Executive summary
Four arguments. One owner.
This paper is for the people who sign for a company's risk: the Owner, the Chief Executive Officer, the Chief Financial Officer, the Chief Information Officer and General Counsel.
It answers the question they now hear every week: who should we trust with our data, our customers and our name when we buy AI?
Our own site opens the subject bluntly: "Almost everyone is an AI expert now. Most of that expertise comes from typing into a chat window in a browser." [S35] This paper is for the meeting after that expert has finished the slides.
1. Compliance cannot be borrowed
A supplier's SOC 2 report describes the supplier's system, not yours. [S2] An ISO/IEC 27001 certificate names one organization and one scope. [S14] HIPAA has no official certification at all. [S4]
So the line "your supplier is SOC 2 compliant, so you are covered" tells you nothing about your own organization, and your buyer will still ask for yours. Our position is the opposite, and it is the line this whole paper rests on: We build compliance-ready systems so that you can get certified.
2. No PHP. Not one line.
On 21 October 2026 PHP2Shell put every PHP installation serving web requests at risk, and CISA listed both flaws as known exploited five days later. [S28] [S31] Our platform runs no PHP at all. It never did. There was nothing to patch, nothing to inventory and nothing to explain.
3. Layer-2 PEIPN: a door that is not there cannot be knocked on
The 2023 casino breaches and most breaches since share one feature: something the attacker needed was reachable from the public internet, or one compromise could travel across a flat network. Verizon's 2026 report says 31% of breaches now start with software vulnerabilities. [S23]
Most of what is sold as cybersecurity is a better lock, a longer warranty or a bigger insurance policy on a door that should not be there. Our Layer-2 Private Enterprise IP Network (PEIPN) takes the door away: a unique VLAN for every single site and every single 5G device on the network, no internet-facing door, and one controlled exit. It does not stop every attack, and we never say it does. It decides how far an attacker can get, and what keeps running.
4. Eight pillars, one owner
"A chat window is one pillar at best." [S35] A system a business can put its name behind needs eight: business applications, an operator, connectivity, private transport, languages and integration, with its own infrastructure and its own power at design stage.
It runs on a dedicated server, hosted for the customer or installed on its premises. Either way, the server is the customer's.
How to use this paper
Part 1 takes the usual AI pitch apart. Parts 2 to 4 make the first three arguments in depth: the history, how the failure works, what the standard says and what to demand in writing. Part 5 introduces the eight pillars, and each pillar then gets its own chapter, with a page for the Chief Financial Officer, the Chief Information Officer and General Counsel.
Parts 6 and 7 cover voice AI and the economics of renting it. Parts 8 to 10 go industry by industry. Part 11 gives ten questions that end the conversation, and Part 12 turns them into a procurement pack with contract clauses. Part 13 covers payments, Part 14 says plainly what we do not claim, and Part 15 is the challenge. The appendices give our published answers, a glossary and every source.
Part 1
The AI guru pitch, taken apart
In this paper, "the AI guru" is not a person. It is a pitch, and you have probably sat through it: a chat window or an API wrapper on someone else's public cloud model, sold on a rented platform, behind a WordPress site, reached over a VPN, with a supplier's certificate held up as if it were yours.
Our page puts it in one sentence: "What you type into is the product of AI, not the AI ... You have been talking to the packaging." [S35]
The pitch against what an enterprise needs
| The AI guru pitch | What an enterprise needs | |
|---|---|---|
| Where the model runs | Someone else's GPUs, behind an API | GPUs it can name, operated for it or on its own premises [S36] |
| The bill | "A taxi meter that runs while your customer talks" [S35] | A cost that "does not climb because your customer kept talking" [S35] |
| The memory | "Three products, three logins, three memories of the same buyer" [S35] | "One memory. One account of the customer." [S36] |
| The rules | Guardrails "written for everybody at once, to protect the vendor first" [S35] | Its own rules: what may be promised, what must never be said [S35] |
| The network | The public internet, a VPN, a door at every site | A private Layer-2 path with no door on the internet, a unique VLAN for every site and every 5G device [S35] |
| The website | WordPress, which runs on PHP [S32] | No PHP anywhere, by design |
| The certificate | Someone else's SOC 2 report, held up as yours | Its own report, earned on a compliance-ready system [S2] |
| When it goes wrong | "The vendor points to the model provider, and the model provider points to its terms" [S35] | "The company on the contract." [S36] |
| What you have after a year | "No hardware, no model and often no memory of your own customers" [S35] | Its own record, on its own server [S36] |
What an LLM really is
Strip the slides away and a language model is simple to describe. Our page puts it in five lines [S35]:
- Tokens. "Text is cut into tokens, pieces of words. The model reads and writes one token at a time."
- Weights. "The model is billions of numbers learned in training, stored as a file. Those numbers are the model. Nothing else is."
- Inference. "For every token, the GPU pushes the whole conversation through every layer of those numbers to predict the next one."
- No memory of its own. "Between conversations the model remembers nothing. What it knows about you is what the system around it places in front of it, each turn."
- The product. "Model, hidden instructions, filters, memory, tools and an interface. When people say they have used AI, this package is what they used."
That last line is the whole trick. The pitch sells you the package and calls it the intelligence. The intelligence is the model and the hardware that runs it, and the value is in what is built around them for one business.
Six things you were told about AI
Our page lists twelve. [S35] These six do the most damage in a boardroom.
- "The answer I got is what an LLM can do." No. "The answer you got is what that vendor allowed, at the length it chose, on a machine shared with everyone else at that moment." [S35]
- "AI lives in the cloud, so the hardware does not matter." "The hardware is the intelligence ... Whoever owns the GPU decides whose work goes first." [S35]
- "Memory is a feature you switch on." "Memory is architecture. In most products each tool keeps its own notes, so the email tool, the dialer and the chatbot hold three different versions of your customer." [S35]
- "A subscription gives you the AI." "A subscription sells you what is left over from the buffet." An enterprise needs tenancy: "resources allocated to one customer, there twenty-four hours a day whether they are used or not". [S35]
- "My data is private because the vendor says so." "On a rented stack your questions, documents and conversations travel the public internet to machines other people run, under rules other people set. That is your business on a postcard." [S35]
- "Everyone has the same AI, so it cannot set us apart." "Everyone can rent the same chat window. Nobody else can have your harness: your context, your ambitions, your rules, your memory and your connections." [S35]
The harness is the product
Our page defines it: "A harness is everything around the model that makes it yours: the context of your business and your ambitions, your rules, your one memory, your connections and your voice. Vendors keep the harness for themselves and rent you the window." [S35]
That is the real dividing line in this market. The guru rents you a window into a harness someone else owns. We give a business "the raw material to build its own harness, its own version of SARAH and its own revenue engine, on GPUs allocated to it". [S35]
The point. None of this makes the pitch dishonest. It makes it rented. Rented is fine for a chat window. It is not fine for the system your customers, your auditor and your board rely on.
Part 2
Compliance cannot be borrowed
Chris Ismail put it plainly: "Just because you deal with a SOC 2 compliant supplier or a HIPAA compliant supplier, it doesn't automatically make you a SOC 2 compliant organization."
And: "They don't care about your suppliers' certifications they care about how good your org is." Every standard below bears him out.
The core line
We build compliance-ready systems so that you can get certified. The words matter. A SOC 2 result is an attestation report from an independent CPA firm, which buyers often call certification [S1] [S2]; ISO/IEC 27001 is a certificate issued by an independent certification body [S14]; and HIPAA has no official certification at all [S4], so for HIPAA, ready means ready for your own risk analysis, your own policies and your own business associate agreements.
How the borrowed certificate became a sales tool
Every rented platform has the same problem: its customers are asked for evidence it cannot give them. So the platform publishes its own report, and the reseller, the agency and the AI wrapper built on top of it start quoting that report as if it covered them.
It does not. The AICPA says SOC reports help users "to assess and address the risks associated with outsourcing services". [S1] When an enterprise buys from you, it is outsourcing to you, and the risk it is assessing is yours.
Its questionnaires say so. Shared Assessments describes the SIG as a tool for "managing third-party risk" [S16]; the Cloud Security Alliance calls the CAIQ a "Security questionnaire for vendors". [S17] Both ask about your own policies, people and controls, and ask you to attach your own report.
SOC 2: one report, one system
- Who issues it. A CPA firm, as a service auditor, under the AICPA's attestation standards. [S1] [S2] The result is a report, not a certificate.
- Whom it describes. One service organization's system and its boundaries. [S2] Nobody else's controls are tested.
- Type 1 and Type 2. Type 1 covers design at a point in time. Type 2 adds whether the controls worked over a period. [S2]
- Who may read it. The "specified parties" agreed by management and the auditor. [S2] Holding out a supplier's report as your own misleads your buyer.
- The firm matters. The AICPA treats SOC 2 engagements with "identical reports, risk assessments, sample sizes, and testing procedures" as nonconforming. [S3] A cheap report nobody trusts is worth little.
The rented platform test
Take any platform you rent: a CRM, a marketing suite, a model API. Suppose it holds a SOC 2 Type II report. That report is about the platform's own system.
The thousands of businesses who pay it every month do not hold that report. None of them can hand it to a buyer as their own. That is true of every platform, and it is true of us.
The controls your supplier says you must run
A SOC 2 report lists complementary user entity controls (CUECs): controls the supplier assumed its customers would run. [S2] Your supplier's own report lists controls that you must run. Skip them, and the assurance it describes does not hold for your use of the service.
Ask any AI vendor for its CUEC list before you sign. If it cannot produce one, it either has no report or has not read its own.
ISO/IEC 27001 and HIPAA
ISO "does not perform certification or issue certificates"; independent certification bodies do. [S14] A buyer checks two things on a certificate: the name of the organization, and the scope. [S15]
HHS is direct: it "does not endorse or otherwise recognize private organizations' 'certifications' regarding the Security Rule". [S4] Asked whether a business associate can be "certified by a third party" instead of signing a contract, HHS answers "No". [S5]
Business associates carry their own direct liability [S7], must sign a business associate agreement with each subcontractor that handles PHI [S6], and every party runs its own risk analysis. [S8] [S9] A vendor that says it is "HIPAA certified" is describing something HHS does not recognize.
| What a supplier holds | What it proves | What it does not prove for you |
|---|---|---|
| A SOC 2 Type 2 report | A CPA firm's opinion on that supplier's system, over a stated period [S2] | Anything about your own policies, staff or systems |
| A list of CUECs | Which controls the supplier assumed its customers would run [S2] | That you run them |
| An ISO/IEC 27001 certificate | That one organization's management system was certified, for a scope [S14] | That yours has one, or that the scope covers your service |
| HIPAA safeguards and a BAA | That the supplier accepted its own contract terms and direct duties [S6] [S7] | Your own risk analysis, policies and agreements [S9] |
| PCI DSS validation | Its status for the services it provides [S10] | Your own validation and your share of the requirements [S10] |
The same rule in Europe
The GDPR builds the chain into law. A processor "shall not engage another processor without prior specific or general written authorisation of the controller", and must tell the controller of changes, "thereby giving the controller the opportunity to object". [S41]
And where a sub-processor fails, "the initial processor shall remain fully liable to the controller for the performance of that other processor's obligations". [S41] Every extra hop in an AI vendor's chain is a hop someone has to authorize, contract and answer for.
The fines are sized to make a board pay attention: up to 20 million euro, or up to 4 % of total worldwide annual turnover for the most serious infringements, whichever is higher. [S43]
We build, you certify
In Chris Ismail's words: "We empower businesses with compliance-ready systems and documentations so that they can get their own certificates."
And: "Every tenant on our infrastructure will have their own unique system ... it is up to the tenant to get their own certifications for their own flavor of OMEGA Revenue Engine."
- We are an engineering company. Each install is, in his words, "a totally different workflow, totally different product, it's modified to fit their needs, it's customized". So "they have to go and get the certificate themselves".
- ORE is not customer facing, so it holds no certificate by design. OMEGA Revenue Engine holds no SOC 2 report, no ISO/IEC 27001 certificate and no PCI DSS validation of its own, and no HIPAA certificate exists for anyone to hold. [S4] The business that sells to the end customer gets certified, for its own system, under its own name.
- We stay inside your vendor management. Where we handle PHI for you, we are your business associate, with a BAA and our own direct liability. [S6] [S7] Where our systems could affect card data, you monitor our status every year. [S10] We answer those reviews with evidence.
- Every tenant gets its own system. Our published answer to whether other tenants share a customer's private VLAN: "No. Each customer path is its own VLAN on PEIPN." [S36]
What we hand you
- A compliance-ready system, built so the controls an auditor tests can be shown: segmentation, named access, logging, retention, change records, isolation drills, data location and dedicated tenancy.
- The paperwork, drafted: a policy pack for adoption in your name, a system description written the way the AICPA criteria ask, including the user entity controls you must run [S2], a risk register and honest readiness matrices. It is a draft until you adopt and sign it.
- Evidence as it is produced, and engineer time for your auditor's walkthroughs.
The road to your own report
| Step | What happens | Who leads |
|---|---|---|
| 1. Scope | List every system that touches customer data or payments, yours, ours and your other vendors' | You, with our system map |
| 2. Frameworks | Choose what your buyers ask for: SOC 2, ISO/IEC 27001, HIPAA, PCI DSS, GDPR | You |
| 3. Readiness | Compare controls to the criteria, list the gaps, close them | You, with our readiness matrices |
| 4. Owners | Adopt the policies; name a Security Officer, a Privacy Officer and an owner for every control | Your leadership team |
| 5. Evidence | Collect from day one: ours from the system, yours from the rest of the business | Your control owners |
| 6. The firm | Evaluate several CPA firms; avoid quick-turn, identical engagements [S3] | Your Chief Financial Officer |
| 7. Type 1, then Type 2 | Design at a point in time, then operating effectiveness over a period [S2] | Your CPA firm |
| 8. Keep it running | Operate the controls daily, review vendors yearly, renew the report | You, with us on the system |
What to demand in writing from any AI vendor
- The name of the organization and the system its report or certificate covers, with the period or scope. [S2] [S14]
- Its list of complementary user entity controls, and which of them fall to you. [S2]
- Its list of sub-processors, its notice period for changes, and your right to object. [S41]
- Where PHI is involved, a signed BAA, and the BAAs it holds with its own subcontractors. [S6]
The contract carries the promise
The party that collects the payment signs the master services agreement, the data processing agreement, the security schedule and, where PHI is involved, the business associate agreement. [S6]
The FTC is blunt about service providers: "the buck still stops with you". [S18] This is practice, not legal advice. The exact terms are for General Counsel.
Part 3
No PHP. Not one line.
On 21 October 2026 the PHP project published fixes for two critical flaws in PHP, the language WordPress and a great many other web applications run on. [S28] [S32]
Within days, buyers' security teams were writing to their vendors. It proves the argument of Part 2: the question arrives at your door, under your name, and must be answered for your own stack.
What PHP2Shell is
- Two flaws, one name. CVE-2026-109618 [S29] and CVE-2026-109619 [S30], both rated 9.2 under CVSS version 4.0. PHP says they "can allow unauthenticated remote code execution". [S28] CISA writes that they are "known as 'php2shell'". [S31]
- Who is at risk. "All PHP installations serving web requests, including PHP-FPM, Apache with mod_php, php-cgi, FrankenPHP and LiteSpeed." For mod_php there is "a demonstrated exploit where an unauthenticated request can trigger the attack before any application code runs". [S28]
- Only an update fixes it. The fixed releases are PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Interim measures "do not make an installation safe". [S28]
- Known to be exploited. CISA added both to its Known Exploited Vulnerabilities catalog on 26 October 2026. [S31]
How the attack works, in plain words
PHP says the first flaw lets out-of-memory handling interrupt PHP while internal request data is only partly set up, so that cleanup may then treat invalid data as valid. In the second, state owned by one request can survive after its memory is freed, and a later request handled by the same worker may use it. [S28]
Strip away the jargon and it means this: a stranger sends a crafted web request, before any login and before your application code runs, and may end up running code on your server. [S28] Every website on PHP that faces the public internet was a door with that lock in it.
Why you cannot see it
CISA's first instruction was: "Identify and inventory systems running any version of PHP across your organization." Its warning: "PHP is often a component of other software and prevalence may not be directly observable." [S31]
That is the real damage. PHP hides inside admin consoles, plugins, page builders, dialers, phone switch web panels and hosting control panels. A company that thinks it runs no PHP may run it in a dozen places it has never inventoried.
CISA's advice for those who cannot update was blunt: they "should take PHP-based systems offline until they can be patched or updated." [S31] For a business whose website, landing pages and dialer all run on PHP, that is not a patch. It is an outage.
The clock
The Known Exploited Vulnerabilities catalog has teeth for the federal government. CISA's directive requires agencies to remediate catalog entries "within two weeks" for vulnerabilities with CVE IDs assigned from 2021 onward. [S40] A buyer can hold its own suppliers to the same clock in contract (Part 12). Our written statement went out on 26 October 2026, the day CISA listed both flaws.
The five questions one buyer sent
A buyer does not ask for a certificate when a flaw like this lands. It asks what happened, and for the evidence.
| The buyer's question | What a good answer contains | Whose answer it is |
|---|---|---|
| 1. Do any of the products or services we use run PHP? | An inventory of every system and vendor, and which run PHP | Yours, for everything you deliver |
| 2. Have all affected systems been updated, and on what date? | The fixed version and the date, for each system | Yours, collected from each host and vendor |
| 3. If patching is not complete, what interim protections are in place? | The measures taken, and when the update will land | Yours |
| 4. Any signs of exploitation or unauthorized access? | What was looked at, and what was found | Yours, from your monitoring and your vendors' |
| 5. Is any action needed on our side? | A clear yes or no, with the steps | Yours |
The standards treat this as a control matter. The AICPA defines a system incident as an event that "requires action on the part of service organization management". [S2] HHS counts "holes, flaws or weaknesses" as technical vulnerabilities for your risk analysis. [S9] The FTC says: "Keep your information security program current." [S18]
Now put the five questions to the pitch
The website is WordPress, and WordPress asks its host for "PHP version 8.3 or greater". [S32] The landing pages sit on a rented builder, the CRM on a rented platform, the dialer and the email sender with other vendors, the model behind an API.
That is five written answers to chase, under your name, from people who may answer slowly or not at all. We do not say any given vendor is affected. We say each one sits in your chain, and each one owes you an answer with a date on it.
Our answer, by design
Our PHP2Shell Security Statement of 26 October 2026 opens: "No. ORE runs on its own private infrastructure, and none of it runs PHP. This is by design. PHP was never part of the ORE architecture."
- No PHP interpreter, package or service is part of the platform, and nothing listens for PHP requests.
- The web server is built without any PHP handler. No website or application on the platform is written in PHP, and no container image includes it.
- Websites are hand-coded HTML, CSS and JavaScript. Application, AI and voice services are our own code. Data sits in databases on private infrastructure.
So the other four answers were short. Updated? "Not applicable. By design there is no PHP on the platform, so there was nothing to update." Exploitation? "No. These vulnerabilities can only be exploited through a running PHP service, and ORE has none." Action on your side? "No."
The only PHP we ever ran was in an older dialer and the web console of a phone switch. We replaced both with our own phone switch and dialer, written from scratch. As of 27 October 2026 there is no PHP anywhere in our fleet.
Secure by design is a government standard now
CISA's Secure by Design program says it plainly: "Every technology provider must take ownership at the executive level to ensure their products are secure by design." [S48] It asks providers to "significantly decrease the number of exploitable flaws before introducing them to the market". [S48]
Removing an entire language from the platform is that principle at its simplest. A vulnerability class that is absent by design never needs an emergency patch.
What to demand in writing
- A list of every component in the service that runs PHP, including admin consoles and plugins, or a signed statement that none does.
- For each, the version running today and the date of the last update.
- The vendor's remediation target for Known Exploited Vulnerabilities, in days. [S40]
- Notice to you, in writing, within an agreed number of hours of any catalog entry that touches the service.
What it does not change. No design removes every flaw: the components we do run have their own advisories, and each gets its own answer, on the record. You still answer your buyer for your own sites, laptops, email and other vendors. Public websites still live on the public internet. [S35]
Part 4
Layer-2 PEIPN: no door on the internet
Our page puts the choice in one line: "Locks, a warranty and an insurance policy. Or a door nobody can find." [S35]
Firewalls, warranties and insurance all accept one starting point: "your systems can be reached by anyone on the planet". [S35] We start somewhere else.
The record
- MGM Resorts, 2023. MGM shut down systems after the attack, which "resulted in disruptions at some of the Company's properties", and estimated a negative impact of about $100 million on Adjusted Property EBITDAR. [S21] It later agreed to pay $45 million to settle the class action over its 2019 and 2023 breaches. [S27]
- Caesars Entertainment, 2023. "A social engineering attack on an outsourced IT support vendor used by the Company" ended with its loyalty program database copied. [S19] The attack ran through a supplier. The filing, the notice and the cost were Caesars' own.
- Marriott and Starwood. One breach ran from 2014 until September 2018 and reached 339 million guest records. Marriott agreed to pay $52 million to 49 states and the District of Columbia, and 20 years of FTC oversight. [S22]
- Wyndham. The FTC alleged a failure to use "firewalls and network segmentation between the hotels and the corporate network". [S25] These were allegations, not findings; the settlement ran for 20 years. [S26] One flat path lets one intrusion become dozens.
IBM put the global average cost of a data breach at $4.99 million in 2026. [S24]
Why it keeps happening
A carrier hands each site a router and a public IP address, and from that point the security of everything behind it is the customer's job. Each site becomes a door on the public internet that anyone on the planet can find and knock on, all day.
In Chris Ismail's words: "Every carrier, every ISP, instead of providing the secure network and taking all the responsibility, they're farming it off to the actual clients to deal with the layer three issues."
The doors are being used. Verizon's 2026 report says 31% of breaches now start with software vulnerabilities, ahead of stolen passwords, and 48% involve ransomware. [S23] Every firewall, VPN concentrator and remote access box that faces the internet is a place to find one.
A VPN is a tunnel across the public internet; its front door is on the internet by definition. A private cloud in someone else's region is, in our published words, "still their network". [S36]
The 2023 casino attacks began with people. The attackers "Posed as company IT and/or helpdesk staff using phone calls or SMS messages", and the advisory's mitigations include "Segment networks to prevent the spread of ransomware". [S20] A private network makes segmentation the default, not a project.
Layer 2 and Layer 3, in plain words
The internet is a Layer 3 network: every device has an address, and routers pass packets between addresses anywhere on the planet. That is what makes it useful, and it is what makes every public address reachable by anyone.
A Layer 2 network is closer to a cable. Traffic moves between the points that are joined, on a VLAN that belongs to one customer, and nowhere else. There is no public address on the production path to scan, because the path is not on the internet at all. [S36]
That is why our published answer to whether a virtual private cloud is the same thing is "No. A cloud you rent is still their network." [S36] And to whether the path survives a cloud region being withdrawn: "The path is part of the metal and the contract. It is not a feature flag in someone else's console." [S36]
Defended on the internet, or not on it
| The question | Defended on the public internet | On the Layer-2 private network |
|---|---|---|
| Who can reach your systems? | "Anyone on the public internet can find the address and knock, all day, every day" [S35] | "There is no public door on it to knock on" [S35] |
| What does each site get? | "Tunnels laid across the public internet" [S35] | "One site, one VLAN, nobody else on it" [S35] |
| What does a cyber warranty do? | Promises to repair or pay, "under conditions written by the seller" [S35] | "A promise after the fact is no substitute for a door that is not there" [S35] |
| What does cyber insurance do? | Pays part of the cost. "The customer data is already out" [S35] | "The cheapest claim is the one that never happens" [S35] |
| Can you prove where the traffic went? | "Through logs from many vendors, kept in many places" [S35] | "Private path, single egress, one place to read the record" [S35] |
The private Layer-2 path
Chris Ismail has spent 19 years building private networks of this kind. As he tells it, in 2010 a private Layer-2 network joined six Australian racecourses, Flemington, Caulfield, Sandown, Moonee Valley, Rosehill and Randwick, with no public IP addresses, and it has run since March 2010.
The Layer-2 Private Enterprise IP Network (PEIPN) brings the same idea to every customer. We act as the customer's private ISP and take the layer three responsibility back.
Most of what is sold as cybersecurity is a better lock, a longer warranty or a bigger insurance policy on a door that should not be there. PEIPN does not sell a better lock. It removes the door.
- One long Ethernet cable. A private circuit from each site to the data center, on a VLAN dedicated to that site. Not a VPN.
- No internet-facing door. The site edge does not answer the public internet, because it is not on it.
- One VLAN for every 5G device. Every 5G device on the network gets its own unique VLAN as well. A device that goes missing is one VLAN to switch off, not a way in.
- One VLAN per system. Workstations, payment terminals, back office, cameras and guest Wi-Fi each sit on their own VLAN, and reach each other only where a rule allows it.
- One controlled exit. What needs the internet leaves through one watched exit, on rules. Nothing comes in uninvited.
- The AI on the same path. The language model, speech recognition and the voice run on hardware we operate, so nobody's systems go out to the public internet to reach SARAH. [S36] In Chris Ismail's words: "You don't need to go to the internet because obviously you have your version of the internet available to you and it gets refreshed all the time by us."
Who gets in: a named login, a whitelisted site, a need to know
- A named login for every person. No shared accounts. Who saw what, who changed what, and who was removed the day they left all come from the record.
- From a whitelisted site only. A login works only from a site or device whose VLAN is on the allowed list. A valid password typed anywhere else on the planet reaches nothing, because there is no door to type it into.
- On a need to know. Each role sees what its work needs and no more, and access reviews are kept as evidence.
On a rented platform, ask who else is in your account. Agencies, freelancers and offshore contractors are often given every contact, and each of them is inside your scope, under your name.
Zero trust, and why the private path is not a free pass
NIST's zero trust architecture "assumes there is no implicit trust granted to assets or user accounts based solely on their physical or network location", and treats authentication and authorization as "discrete functions performed before a session to an enterprise resource is established". [S38]
We agree, and we design for it. The private path removes the public door; it does not replace the lock on each room. That is why every person still logs in by name, every role is limited to what it needs, and every access is recorded. The private network and zero trust are two halves of one answer, not rivals.
What it changes, stated carefully
| Step seen in the cited attacks | What the private network changes | What you still need |
|---|---|---|
| Calling the help desk to reset passwords [S20] | Nothing by itself: a stolen password is still stolen | Strict caller checks, phishing resistant MFA |
| Talking staff into running remote access tools [S20] | The tool must reach out through the one watched exit, where it can be seen and blocked | Control of what staff can install |
| Logging in from the attacker's own machine | No public address, and logins work only from whitelisted sites | Controls on cloud email and identity |
| Spreading between systems [S20] | One VLAN per system and per site; no path to the next by default | Offline, tested backups |
| Copying a customer database out [S19] | Large transfers cross one watched exit | Keep less customer data |
| Shutting everything down to contain it [S21] | One VLAN or one site cut off while the rest keeps trading | A rehearsed incident plan |
The isolation drill
A private network is only as good as the proof that its walls hold. That proof is a drill, run on a schedule and recorded.
- Cut one system. Take one VLAN off the network, such as the payment terminals, and record how long it took and what kept running.
- Cut one site. Take one site off the network, and record the same.
- Try the wall. From one VLAN, try to reach another that no rule allows, and record that it fails.
- Lose a device. Switch off the VLAN of one 5G device, and record the time from the report to the cut.
Each drill is evidence an auditor can test and an insurer can read. A promise that the walls hold is not.
What to demand in writing
- A list of every address on the service that answers the public internet, and what answers on each.
- Whether other customers share any network segment with yours, and how that is tested. [S36]
- Where privileged traffic travels, and whether it ever crosses the public internet. [S36]
- A timed isolation drill: how long it takes to cut off one system, and one site, with the result recorded.
The honest claim is reduced exposure and a smaller blast radius, not immunity. Carrier circuits for each customer site are arranged per deployment.
Part 5
Eight pillars. One sovereign system.
Our published line: "A chat window is one pillar at best. A business that wants to own its intelligence needs all eight, built to work as one." [S35]
Each pillar is a tool. Taken as a whole or in parts, the system is hosted and managed on a dedicated server, or installed on the customer's own premises. Either way, the server is the customer's.
| No. | Pillar and tool | What it does | What the pitch cannot do |
|---|---|---|---|
| 1 | OMEGA Revenue Engine. Business applications | "25,000 industry modules that replace your subscriptions" [S35] | Replace the stack; it adds one more subscription |
| 2 | SARAH AI Suite. Operator | "Complete business automation, with the voice that answers the phone" [S35] | Pick up the telephone, remember the last call, act |
| 3 | SOPHIA Layer-3 API Hub. Connectivity | "34.7 million enterprise features over private peering" [S35] | Show one record of what left |
| 4 | Layer-2 Private Enterprise IP Network. Transport | "Not a VPN, and not the public internet" [S35] | Leave the public internet; a cloud is reached over it |
| 5 | 119 Languages. Language | "Thinks in 119, understands 99, speaks 76 in her own voice" [S35] | Say which languages its own voice speaks |
| 6 | Integration. Compatibility | "Never rip and replace. Always integrate with what you already run" [S35] | Keep what you run; it migrates you |
| 7 | Last Mile Mini Data Center. Infrastructure, design stage | Designed to put the GPUs beside an existing data center, on dark fiber | Say where its compute is |
| 8 | Off-Grid Power. Power, design stage | Designed to run on magnetic and hydrogen generators on site [S35] | Say what powers its compute |
Pillars 1 to 6 are what a customer buys today. Pillars 7 and 8 are at design stage, with no delivery date, and we say so every time.
Why eight, and why one
Our site gives the short version in six lines a child could follow. [S35] "A chat app is a vending machine. You press a button and you get what somebody else filled it with. It is not the factory." "The real brain is made of numbers." "Most people rent a slice of a shared brain." "Small brains are fast but shallow." "The brain forgets after every chat." "Helpers called agents are just recipes."
Each line names a gap that one pillar closes. The engine and the operator do the work; the hub and the network carry it; the languages let it speak to the buyer in the buyer's own words; integration keeps what the business already runs; and the two design-stage pillars answer where the compute sits and what powers it.
Buy one pillar on its own and the gaps come back. A voice widget without one memory forgets the caller. A memory without a private network travels on a postcard. A private network without the hub has no record of what left. That is why the pitch sells you parts, and why the parts never add up.
One memory, one soul, one source of truth
Our page calls these "Three words almost nobody can explain", and explains them [S35]:
- One memory. "Every channel, the telephone, email, the browser and chat, writes to one account of the customer. Change the channel and the customer is still the same person."
- One soul. "Every agent speaks with one identity, one set of rules and one voice for the business. Two agents on the same customer cannot contradict each other."
- Single source of truth. "Each fact is stored once and everything reads it from there. Campaigns, calls and proposals agree on who the buyer is because they read the same record."
Pillar 1
OMEGA Revenue Engine: business applications
What it is
OMEGA Revenue Engine is the core of the system: "25,000 industry modules that replace your subscriptions." [S35] It does the commercial work of a business, finding, qualifying, contacting and winning accounts, and keeps the record of that work in one place.
Asked what a buyer gets from it, our published answer is: "The commercial work, done on one memory. Research, outreach, qualification, proposal and the record of that work sit on the same memory as the conversation, so what a buyer said on the call is already in the proposal." [S36]
And in one unit of work: "A customer, a conversation, a connection and a commercial outcome. That is what you are paying it to produce." [S36]
What the AI pitch offers instead
The pitch offers a tool for each step. Our page describes the category exactly: "An AI research agent, email agent and SDR agent" that turns out to be "Three products, three logins, three memories of the same buyer, each one a wrapper on a public model." [S35]
It also offers "AI built into the tools you already use", which our page describes as "A feature on someone else's model that knows only that one tool, adds to its price, and never sees the rest of your business." [S35]
Each of those tools is a subscription, a login, a renewal date, a data processing agreement and a copy of your customer. The pitch calls that a stack. A buyer's security reviewer calls it scope.
Why the record matters more than the feature
A feature can be copied in a quarter. A record of every conversation, proposal and outcome with your own buyers cannot. That record is the asset, and the question is whose balance sheet it sits on.
Our answers are short. Do you keep the proposal the engine writes? "Yes. It is part of your record." [S36] Will the proposal know what was said on the call? "Yes. The proposal and the call that produced it live in the same house, on the same account of the customer. Nobody retypes the conversation into a second tool." [S36]
Will campaigns and calls agree on who the buyer is? "Yes. They draw on one memory. The buyer does not become a different person because the channel changed, so your marketing and your sales team work from the same facts." [S36]
Questions that expose the difference
- How many separate products, logins and data stores hold a copy of one customer in your proposal?
- When a buyer says something on a call, how does it reach the proposal, and who retypes it?
- If we leave, what do we take with us: the export, or the memory the system built about our customers?
- "Do I get more than a dialer with an AI label?" Our answer: "Yes. You get three parts that work as one: SARAH holds the conversation, the engine does the commercial work, and SOPHIA makes the connections to the systems you already run." [S36]
One buyer, one record: an illustration
Follow one buyer through a week, as an illustration of the design, not a customer story.
- Research. The engine builds the account: the company, the people, what they buy. It is written once, to one record.
- Outreach. The email that goes out reads that record. The reply comes back to it.
- The call. SARAH answers the buyer's call with the email thread already in front of her, and writes what was said to the same record.
- The proposal. The proposal draws on the call. "Nobody retypes the conversation into a second tool." [S36]
- The follow-up. Marketing and sales read the same facts, so the next message does not ask the buyer what they already said.
On the pitch's stack, each of those five steps happens in a different product, and the buyer meets five strangers.
What to test in a pilot
| Test | How | What a pass looks like |
|---|---|---|
| One record | Change a buyer's job title in one channel | Every other channel shows the change at once |
| Call to proposal | Hold a test call, then draft the proposal | The proposal cites what was said, with no retyping |
| Subscriptions retired | List the tools the engine replaces | A written list, with the date each one can be canceled |
| The exit | Ask for a full export of one account | The record, the calls and the proposals, in a format you can open |
What it is not
- Not a promise of revenue. The engine does the work and keeps the record. We publish no results of our own, and this paper gives none.
- Not a CRM you must move into. Your system of record stays where it is, and the engine connects to it (Pillar 6).
- Not three tools behind one login. One memory sits under every step; that is the difference, not the screen.
For the Chief Financial Officer
Count the subscriptions the pitch adds, not the one it demonstrates. Every one carries its own renewal, its own price change and its own exit cost. Ask whether the invoice will match the page. Ours: "Yes. The figure on the page is the figure. We do not publish one price and invoice another." [S36]
For the Chief Information Officer
Every separate tool is one more integration to keep alive and one more place where the customer record can drift. One engine on one memory removes the reconciliation work rather than automating it.
For General Counsel
Every copy of a customer record is a copy you must be able to find, correct and delete when a customer asks. Fewer copies mean fewer places to search and fewer processors to name. Under the GDPR, each of those processors needs written authorization and a contract. [S41]
Pillar 2
SARAH AI Suite: the operator
What it is
SARAH is one operator across every department of the business, by voice and by text, with one memory. Our site describes the suite as "Complete business automation, with the voice that answers the phone." [S35]
Her language model, speech recognition and voice "run on GPUs we operate". [S36] She is "the advisor on the telephone, the browser and the public number, with one memory". [S36]
What she does, on the record
| The question | Our published answer [S36] |
|---|---|
| Does she answer, book and transfer to a person? | "Yes. She transfers when the work needs a person. The desk you chose to keep stays in the call." |
| Can the customer's own people stay on the line? | "Yes. SARAH routes to them." |
| Where does a booking go? | "Into the customer's calendar and the ledger of the call. Not into a separate bot that forgets the telephone." |
| Does the call leave a ledger the customer can read? | "Yes. The conversation, the transfer and the booking are part of the record." |
| What is one spoken turn? | "One statement, or one question. She does not stack a second question on the same turn to rush the caller." |
| What will you not call her? | "A chatbot skin. She is the voice of the system the customer is running." |
What the AI pitch offers instead
The pitch offers agents, usually many of them. Our page is blunt about what an agent is: "An agent is a script that calls a model. What does the work is the skills and the tools." [S35]
So the pitch counts agents the way an older pitch counted features. The question that matters is whether those agents share one memory, one set of rules and one identity, or whether each keeps its own notes about your customer and its own idea of what may be promised.
Our page answers the guardrail question directly: vendor guardrails are "written for everybody at once, to protect the vendor first. A business needs its own rules: what may be promised, what must never be said, who takes the call when it matters." [S35]
Why the person stays in the loop
An operator that cannot hand a conversation to a person is a wall, not a receptionist. SARAH transfers when the work needs a person, and the desk the customer chose to keep stays in the call. [S36]
The person also stays in charge of what she may do. A business decides which actions she takes on her own, a booking, a callback, a message, and which wait for a person to approve them. That line is written into the rules, not left to the model.
Our page describes the same principle for the tools she runs on: algorithms written more and more by AI, "approved by a person", and in the sixth era, "A person still approves what goes live." [S35]
Questions that expose the difference
- When two of your agents speak to the same customer, do they read the same memory, or two?
- Who wrote the rules your agents follow, and can we change them, or only the vendor?
- When the AI cannot help, how does the caller reach a person, and does the person see what was already said?
- What does your operator do when a caller asks something its rules do not cover?
- "Who answers when she is wrong?" Our answer: "The company on the contract. A rented model vendor is not the party." [S36]
Rules a business writes for its operator
Your rules are what make the operator yours. These are the kinds of rules a business writes; the examples are illustrations, not a template.
| Kind of rule | An illustration |
|---|---|
| What may be promised | A delivery date only when the order system confirms it |
| What must never be said | Any price, discount or term that is not on the published page |
| Who takes the call | A named person for complaints, legal matters and anything about health or money |
| What the caller is told | That they are speaking with an AI, where the law asks for it [S44] |
| What is recorded | Consent, disclosure, the transfer and the booking, on the customer's record [S36] |
What to test in a pilot
| Test | How | What a pass looks like |
|---|---|---|
| Transfer | Ask for a person halfway through a call | A person joins and sees what was already said |
| One memory | Call, then email, then chat, as the same customer | Nobody asks the customer to repeat themselves |
| The rules hold | Try to talk the operator into a promise your rules forbid | It declines, and the attempt is on the record |
| One turn | Ask a question, then listen | One statement or one question per turn, never two [S36] |
For the Chief Financial Officer
An operator is measured by work finished, not by agents deployed. Ask the vendor to price the outcome you want, a booking, a qualified conversation or a resolved call, and to say what happens to the cost when calls run long.
For the Chief Information Officer
One operator means one place to set the rules, one place to read the record and one identity to manage. Many agents mean many configurations that drift apart. Ask how a change to one rule reaches every channel.
For General Counsel
Your operator speaks in your name. Every promise it makes on a call is a promise your company made. That is why its rules must be yours, written for your business, and why the ledger of each call must be one you can read.
Pillar 3
SOPHIA Layer-3 API Hub: connectivity
What it is
SOPHIA is the customer's own hub to every system it runs. Our site describes it as "34.7 million enterprise features over private peering." [S35] Asked what SOPHIA is, our published answer is: "The Layer-3 hub. Connectors, egress, and the record of what left the building." [S36]
The figure is our published one, and our page says so: "34.7 million. That is the published figure." [S36] What matters more than the size is the rule it runs on.
One egress, one record
When a connector must call the outside world, our published answer to who governs that call is: "SOPHIA. One egress, one record. The language model does not wander off and call an ungoverned API with the customer's data." [S36]
And to what is allowed to leave: "Only what the connector was asked to send, through the egress, with a record." [S36] A connector, in our words, "is a governed call. The model does not receive a dump it was not asked to use." [S36]
What the AI pitch offers instead
The pitch offers a slide of logos. Our page describes agent platforms as scripts "joined by connectors that each leave by a different door, with no single record of what left". [S35] And "Chat with your documents" as "Your documents copied to another vendor, searched by their system and answered by a public model". [S35]
Our answer to whether the catalog is a slide of logos: "No. It is dispatch the product can call. A logo is not a connector." [S36]
Why egress is where the audit lives
Every breach notice eventually asks one question: what left, when, and to whom. On a stack where each tool calls out on its own, that answer is spread across the logs of every vendor, if they kept them. On a hub with one egress, it is one record.
The GDPR asks a processor to have "the ability to ensure the ongoing confidentiality, integrity, availability and resilience of processing systems and services". [S42] It is easier to show that ability when there is one door out, and it is watched.
| The question | Connectors that each leave by their own door | One governed egress |
|---|---|---|
| What left the building? | Ask each vendor, and wait | "Only what the connector was asked to send, through the egress, with a record" [S36] |
| Can a connector be switched off? | Often bundled, and left running | "Yes. What you do not need is not left running because a bundle included it" [S36] |
| Does adding a system add a memory? | Usually, one more copy of the customer | "The connection joins the same account of the customer" [S36] |
| Who holds identity across channels? | Each tool, separately | "One person record" [S36] |
Questions that expose the difference
- For every connector in your catalog, which door does it leave by, and where is the record of the call?
- Can a connector send the whole customer file to the model, or only the field it was asked for?
- Can we switch off a connector we do not need, and prove it is off?
- "Does SOPHIA replace the customer's bank, airline or CRM?" Our answer: "No. She connects to them. The system of record stays the customer's. The call is governed." [S36]
Every connector is a contract
A connector that sends personal data to a new outside party makes that party a processor or a sub-processor of your data. Under the GDPR, that needs written authorization, notice of changes and a right to object. [S41] Where PHI is involved, it needs a business associate agreement. [S6]
So the useful question about a catalog is not how many logos it holds. It is how many of those connectors send your data somewhere new, and whether each one is named, governed and recorded. Our rule is that "What you do not need is not left running because a bundle included it." [S36]
What to test in a pilot
| Test | How | What a pass looks like |
|---|---|---|
| One egress | Trigger three connectors, then ask for the record | One record shows each call, what was sent and where |
| Least data | Ask a connector for one field | Only that field leaves; the rest of the file stays |
| Switch off | Turn off a connector you do not need | It stops, and the record shows when |
| The list | Ask for every outside party a connector can reach | A written list, before you sign |
What it is not
- Not a marketplace shared with other tenants. "The hub is the customer's, on the estate they contracted." [S36]
- Not a replacement for your systems. "She connects to them." [S36]
- Not a way around your own vendor reviews. Every outside party a connector reaches is still a party you must review, and we name each one.
- Not a dump of your data into the model. "The model does not receive a dump it was not asked to use." [S36]
For the Chief Financial Officer
Every connector that sends data to a new vendor is a new processor, a new contract and a new party to chase after an incident. IBM found that detection and escalation, with lost business, made up 63% of the average cost of a breach in 2026. [S50] One record shortens the detection.
For the Chief Information Officer
Treat the egress as the control point. If the vendor cannot show you where the outbound traffic of its connectors goes, you cannot write a rule for it, and nor can your firewall.
For General Counsel
A processor must not engage another "without prior specific or general written authorisation of the controller", and remains "fully liable" for a sub-processor's failures. [S41] Ask for the list of every outside party a connector can reach, before you sign, not after the incident.
Pillar 4
Layer-2 Private Enterprise IP Network: transport
What it is
The Layer-2 Private Enterprise IP Network (PEIPN) is the private path between every site, every device and the system. Our site describes it as "A private network of its own. Not a VPN, and not the public internet." [S35]
Our published answer to what it is: "The Layer-2 Private Enterprise IP Network. Ethernet, not the public internet." [S36] And whether it is an add-on sold after the demo: "No. It is part of the foundation, with SARAH, SOPHIA, Omega Revenue Engine and the Last Mile Mini AI data center, which is designed as OMEGA Hyperscale. Omega Mini is a separate product." [S36]
The private ISP
In Chris Ismail's words, carriers and ISPs are "farming it off to the actual clients to deal with the layer three issues". PEIPN reverses that: we act as the customer's private ISP and take the layer three responsibility back.
For a business with many sites, that changes who carries the work. Each new site is joined on its own VLAN, not given a router and a public address and left to defend itself.
Every site, every device
- Every site. "Every single site gets its own dedicated VLAN ID on PEIPN. One site, one VLAN, nobody else on it." [S35]
- Every mobile device. "Through our Layer-2 Mobile Service, every single mobile device gets its own dedicated VLAN ID and joins the private network, not the public internet." [S35] On the public internet, our page notes, "Every phone sits on the public internet, with an app, a VPN client or nothing at all. Each one is another door." [S35]
- Every privileged call. Where do privileged calls travel? "On PEIPN when they are ours to carry. Through the governed egress when the outside system is the point." [S36]
What the AI pitch offers instead
The pitch rarely mentions the network at all, because the network is the public internet. The model sits in someone else's region, the app is reached through a browser, and the voice travels as packets across carriers nobody in the room can name.
When it does mention the network, it offers a VPN or a private cloud. Our answers: a browser widget on the public web is "how a visitor arrives. PEIPN is how privileged traffic is carried." [S36] And "A cloud you rent is still their network." [S36]
Questions that expose the difference
- From the moment a caller speaks to the moment the voice answers, which networks does the audio cross, and who operates each?
- "Must voice and GPU traffic cross a multi-tenant hyperscaler backbone to reach the model?" Our answer: "No. That traffic is carried on the private path." [S36]
- How is a lost or stolen phone cut off, and how long does it take?
- "Can that path carry the model, the voice and the connectors without turning each into a public API?" Our answer: "Yes. That is what it is for." [S36]
What a new site needs
- A circuit. A private circuit from the site, arranged per deployment with the carriers that serve it.
- A VLAN. Its own VLAN on PEIPN, with nobody else on it. [S35]
- A plan for its systems. Workstations, payment terminals, cameras and guest Wi-Fi each on their own VLAN.
- Exit rules. What the site may reach outside, through the one watched exit, written down.
- A first drill. The site is cut off once, on purpose, and the result recorded, before it goes live.
What to test in a pilot
| Test | How | What a pass looks like |
|---|---|---|
| No public door | Scan the site from the public internet | Nothing on the production path answers |
| Separate VLANs | From the guest Wi-Fi, try to reach a payment terminal | No path |
| One site off | Cut the test site off the network | Every other site keeps trading; the time is recorded |
| A lost phone | Report a test 5G device lost | Its VLAN is switched off, and the time recorded |
What it is not
- Not a VPN. A VPN is a tunnel across the public internet; PEIPN is a private circuit on a dedicated VLAN.
- Not software on a laptop. "It is the network. Not a piece of software on a personal computer." [S36]
- Not immunity. It reduces exposure and shrinks the blast radius. People can still be deceived, which is why named logins and drills stay (Part 4).
- Not your public website. Public websites still live on the public internet, "and are defended the usual way". [S35]
For the Chief Financial Officer
The cost of a breach that does not happen never appears on a budget, which is why it is easy to under-fund. The global average cost of a breach was $4.99 million in 2026. [S24] Our page puts the comparison plainly: "Insurance can still be wise. But the cheapest claim is the one that never happens." [S35]
For the Chief Information Officer
A private path is a different operating model: fewer internet-facing devices to patch, fewer remote access boxes to watch, and one place to read the record. It does not remove the work of identity, endpoint control and backups. It shrinks the area that work has to cover.
For General Counsel
When an examiner, a regulator or a court asks where the audio and the data went, the answer should be a single record, not a request to every carrier and vendor in the chain. Our published answer to whether we can audit what crossed the boundary: "Yes. Private path, single egress, one place to read the record." [S36]
Pillar 5
119 Languages: language
What it is
SARAH "Thinks in 119, understands 99, speaks 76 in her own voice, and never claims more than is true." [S35] Those are three different numbers, for three different abilities, and we never add them together or round them up.
- Thinks in 119. The model can reason in 119 languages, so a buyer can write or speak in their own language and be understood in substance.
- Understands 99. Speech recognition can turn a speaker of 99 of them into text the system can act on.
- Speaks 76 in her own voice. The voice the caller hears speaks 17 languages natively, and "she switches between them in the same conversation". [S36] For the others, she understands the speaker and shows the translation on screen.
Why we do not round up
Our published answer to how many languages she actually speaks ends with a rule: "We do not round the number up. A language is claimed only as far as it is true." [S36] And to whether she will claim a language she cannot speak: "No. She speaks 76 languages in her own voice. The claim stops where the voice stops." [S36]
There is a second speech engine for languages the primary engine does not speak, and our answer says what it is: "Yes. It is additive. The primary engine stays in place. A language is claimed only as far as it is true." [S36]
What the AI pitch offers instead
The pitch offers "over a hundred languages" as one number. It is usually the count of languages the model can read, quoted as if it were the count the product can speak on the telephone.
Our page names the weakness that follows from a small model: "a small brain gives small answers, and it stumbles the moment a buyer switches language." [S35] It describes inference clouds the same way: "Fast, but shallow, and weak across languages." [S35]
Why it matters to a buyer
A customer who calls in their own language and is answered in a stilted one has learned something about you. A customer told that a language is supported, who then meets a translation layer, has learned something worse.
In regulated work the stakes are higher. A misunderstood answer on a booking is an inconvenience. A misunderstood answer about a patient, a payment or a contract is a record you will have to explain.
Questions that expose the difference
- Give us three numbers, not one: how many languages the model reads, how many your speech recognition understands, and how many your own voice speaks natively.
- Where is the voice for each language made, on your engine or on a public voice API?
- When a caller switches language halfway through a sentence, what happens?
- Which languages are translated on the way out rather than spoken natively, and does the caller know?
The three numbers, side by side
| Ability | Count [S35] | What it means for a caller |
|---|---|---|
| Thinks in | 119 | Can write or speak in their own language and be understood in substance |
| Understands | 99 | Can speak in their own language and have it recognized as speech |
| Speaks natively | 17 | Hears an answer in their own language, in a native voice, and can switch mid call |
What to test in a pilot
| Test | How | What a pass looks like |
|---|---|---|
| Native voice | Native speakers call in each language you sell in | They rate the voice as natural, on the record |
| The switch | Change language halfway through a call | The answer follows in the new language |
| The honest edge | Call in a language the voice does not speak | Understood, with the translation shown on screen, and no claim to speak it |
| The claim | Read every page that mentions languages | Three numbers, never rounded up [S36] |
What it is not
- Not one number. Reading, understanding and speaking are three abilities, and we give three counts.
- Not a translation layer sold as native. Where the voice does not speak a language natively, the caller is understood and the translation is shown on screen, and we say so.
- Not a list that grows on a slide. A language is added to a claim only when it is true. [S36]
For the Chief Financial Officer
A language claimed and not delivered is a market you paid to enter and did not reach. Ask for the native list before you budget a campaign in a new country.
For the Chief Information Officer
Test the claim the way your customers will: on the telephone, with native speakers, in the languages you sell in, including a switch between two of them on one call.
For General Counsel
A claim about languages on your website or in your contracts is your claim. If the voice cannot speak a language natively, do not let anyone say it can. The rule we work to is simple: the claim stops where the voice stops. [S36]
Pillar 6
Integration: compatibility
What it is
"Never rip and replace. Always integrate with what you already run." [S35] The customer keeps its numbers, its phone system, its applications and its data. SARAH connects to them, and the system of record stays the customer's. [S36]
Six eras, all still running
Our page lists six eras of technology, each of which "made the last look like a toy". [S35] Every business still runs on something from each of them.
| Era | What it was, in our page's words [S35] | What a business still runs from it |
|---|---|---|
| 1. Analogue | "The telephone line, the switchboard and the operator who connected each call by hand" | Phone lines, fax lines, door and alarm circuits |
| 2. Digital | "Digital exchanges, the personal computer and the database" | The databases and accounting systems the business closes its books on |
| 3. IP | "Voice over IP, the web, email and the cloud" | The phone system, the website, email and cloud applications |
| 4. Human-written algorithms | "Lead scores, routing tables, email sequences and CRM workflows" | The CRM and the rules the team built into it |
| 5. AI | "For most people this era arrived as a chat window" | Chat assistants and AI features inside existing tools |
| 6. AI-written algorithms | "The model now helps write, test and improve the tools it runs on" | The new work, approved by a person before it goes live |
Our page describes where we built: "Omega was built for the sixth era on the network of the third: a switch we wrote from scratch, our own private network, and the language model, the speech recognition and the voice on GPUs we operate." [S35]
And it says what we do not do: "We do not manufacture the chips, and we do not manufacture the base model." [S35]
What the AI pitch offers instead
The pitch offers a migration. Move your contacts, your numbers and your workflows onto the platform, and then the AI can help. That is a project with a cost, a risk and a date, before any value arrives.
And once the migration is done, the platform holds your numbers, your contacts and your workflows. Leaving it means a second migration, in the other direction, on its timetable.
Our page also corrects the myth that the AI era makes everything before it obsolete: "The algorithms around it, the scripts, tools, databases and connectors, are ordinary code, and most of it runs on CPUs. Today's data centers keep doing that work." [S35]
Questions that expose the difference
- What must we move onto your platform before your product works, and what happens to it when we leave?
- Can we keep our telephone numbers, our phone system and our CRM exactly as they are?
- "Can a system be added without opening a second memory?" Our answer: "Yes. The connection joins the same account of the customer." [S36]
- Which of our existing systems will you refuse to connect to, and why?
- If your platform is switched off tomorrow, which of our systems stop working?
An integration inventory
Before any AI is switched on, list what the business already runs and what happens to each. This is the shape of that list.
| What you run | What happens to it | Where the record stays |
|---|---|---|
| Telephone numbers | Kept; calls reach SARAH and your people on the same numbers | With you |
| The phone system | Kept and connected | With you |
| The CRM | Connected through SOPHIA; nothing migrated first | In the CRM [S36] |
| Calendars | Connected; bookings land in them [S36] | In your calendar |
| Accounting and banking | Connected where the work needs it, through the one egress | In your systems [S36] |
What to test in a pilot
| Test | How | What a pass looks like |
|---|---|---|
| Keep the numbers | Route one existing number to the system | Callers notice nothing but the answer |
| No migration | Connect the CRM without moving it | Records stay where they are, and stay current |
| Rollback | Switch the new layer off | Everything you ran before still runs |
| One memory | Add a second system | No second copy of the customer appears [S36] |
What it is not
- Not a promise to connect to everything. Some systems have no safe way in; we say which, and why, before you sign.
- Not a back door into your systems. Every connection is a governed call through the one egress, with a record. [S36]
- Not a reason to keep what should retire. Integration lets you retire an old system on your own timetable, not the vendor's.
For the Chief Financial Officer
Every system the business already owns is a sunk cost that still earns. A migration writes that off before the new system has proved anything. Integration keeps the old investment working while the new one is tested.
For the Chief Information Officer
A migration is the riskiest project most IT teams run. Integration keeps the system of record where it is, and lets the new layer be switched off if it fails a test. Ask the vendor for its rollback plan.
For General Counsel
The system of record carries your retention rules, your legal holds and your audit trail. Moving it moves all of those. When it stays where it is, those obligations stay where they are, under the controls you already know.
Pillar 7
Last Mile Mini Data Center: infrastructure, design stage
Design stage. The Last Mile Mini Data Center is a design, not a running site, and it has no delivery date. Everything in this chapter describes the design, and nothing in it describes a service a customer can buy today.
What it is designed to be
Our site describes the Last Mile Mini Data Centers as designed buildings: "Our own buildings, next to the fiber, designed for the racks the biggest brains need." [S35]
The design puts the GPUs in a couple of racks beside an existing data center, on dark fiber, close to the people who use them, so the existing center keeps doing the switching, routing and carriers.
Our published answer to whether other tenants share a customer's private VLAN adds the design intent: "No. Each customer path is its own VLAN on PEIPN. The Last Mile Mini is designed as our own building, not a shared hall, and it is not Omega Mini." [S36]
Why the design exists
Our page explains why the newest racks need somewhere new: "The model's reasoning needs GPUs. The algorithms around it, the scripts, tools, databases and connectors, are ordinary code, and most of it runs on CPUs. Today's data centers keep doing that work. GPUs are added where the model runs, and the racks that need liquid cooling go where the cooling is." [S35]
So the Last Mile design does not replace a data center. It is designed to sit beside one, for the one part of the work the existing hall was not built for.
What a customer can buy today
A customer does not need the Last Mile design to exist. Our published answer: "SARAH Mini does not need this building to answer. Nor do SARAH Pro and SARAH Enterprise. They can sit on the customer's premises, or in a current data center." [S36]
Where policy says the data must stay on the customer's floor, "That path is SARAH Mini, SARAH Pro or SARAH Enterprise, on your premises." [S36]
How to read a design-stage claim, ours included
Every AI supplier now has a roadmap slide. A buyer should treat every item on it, including this one, the same way:
- Ask whether it runs today. If the answer is no, it is a design, and it should be labeled as one wherever it appears.
- Ask for a date, and treat no date as no date. We give none for the Last Mile design, and we say so.
- Do not pay for it, or sign for it, as if it exists. Contract for what runs. Note the design as a design.
- Ask what you get if it never arrives. For us, the answer is everything in pillars 1 to 6, on a dedicated server or on your own premises.
Questions that expose the difference
- "Who operates the GPUs?" Our page puts this question first to every provider. [S35]
- Where, physically, is the compute that will serve our calls, and whose building is it in?
- Which parts of your proposal are running today, and which are designs?
For the Chief Financial Officer, the Chief Information Officer and General Counsel
Price, plan and contract for what runs today. The Last Mile Mini Data Center is at design stage; if and when it enters service, it will be offered on its own terms, and nothing in your system today depends on it.
Pillar 8
Off-Grid Power: power, design stage
Design stage. Off-Grid Power is a design for the Last Mile sites, not a running service, and it has no delivery date. Everything in this chapter describes a design.
What it is designed to be
Our site describes the design in one line: "Permanent magnetic and hydrogen generators, so the grid is the last source, not the first." [S35]
The design aims to keep the GPUs from competing with homes and businesses for power, and to keep the compute running when the grid is short, so that the answer a customer gets does not depend on the state of the local grid.
Why power is part of the AI question
Our page says "The hardware is the intelligence." [S35] Hardware runs on power. A model at full strength "needs enterprise GPUs with fast memory beside them" [S35], and those GPUs need power and cooling around the clock, whether a customer is calling or not.
A rented AI service hides that question from the buyer. A buyer who owns the server, on a dedicated estate or on its own premises, sees it, plans for it and can ask what happens when the supply is interrupted.
What a customer has today
Today, a customer's system runs on a dedicated server in a data center, with that data center's power, or on the customer's own premises, with the customer's power. The Off-Grid Power design changes nothing in that until it is built, and we give no date for it.
Design stage against today
| Today | The Off-Grid Power design | |
|---|---|---|
| Power source | The data center's supply, or the customer's own | On-site generators, with the grid as the last source |
| Where it applies | Every system a customer buys now | The Last Mile sites, which are themselves a design |
| Delivery date | Running | None given; design stage |
Questions to ask any supplier about power
- What powers the compute that serves our calls, and what happens to the service when that supply fails?
- Is there a second site, on a different power grid, and does the service move there on its own?
- Which parts of your power story run today, and which are designs?
For the Chief Financial Officer, the Chief Information Officer and General Counsel
Treat Off-Grid Power as what it is today: a design. Ask for the continuity plan that covers the system you are buying now, test it, and keep the record of the test. Plan nothing around the design until it runs.
Part 6
Voice AI, taken apart
On the telephone, the AI pitch is at its most convincing and its most exposed. The demo sounds natural. What it does not show is where each part of the call went.
The three engines on every voice call
Every AI voice call runs three engines in turn: speech recognition turns the caller's words into text, the language model decides what to say, and speech synthesis turns the answer back into a voice. Asked to name the three engines on our voice path, our answer is: "Language, speech recognition and speech synthesis. All three are hosted by us." [S36]
| Engine | On a typical voice widget | On our voice path [S36] |
|---|---|---|
| Speech recognition | A public transcription API | "Recognition runs on our GPUs. The utterance is not posted to a public transcription API." |
| The language model | A public model, behind someone else's rate limits | "Inference for the conversation stays on hardware we operate." |
| Speech synthesis | A public voice API | "The voice the caller hears is synthesized on our speech engine, on our GPUs." |
| Changing an engine | A silent failover to another vendor | "Changing it is a decision, not a silent failover." |
Our page describes the voice widget category as "A call sent across the internet to public speech recognition, a public model and a public voice, three hops that can each be listened to or switched off." [S35]
On a call, latency is the conversation
Our page puts it in one line: "On the telephone a pause is the difference between a person and a machine." [S35] Each public hop adds a trip across the internet, and each trip can queue behind other customers' traffic.
Our answer is architecture, not tuning: "Our speech recognition, the model and the voice run in the same box, so nothing crosses the internet between them." [S35]
The voice is personal data
A recording of a caller is a recording of a person. On a rented stack it travels to the transcription vendor, and often to the voice vendor, under their terms.
Our published answers: the audio "stays with the customer's record. It is not handed to a model vendor to improve a clone library." [S36] Do we collect a voice sample and send it to a third party to render? "No. Synthesis stays on our engine." [S36]
What the regulators already say
- AI voices are "artificial" voices. On 8 February 2024 the FCC adopted a Declaratory Ruling that "recognizes calls made with AI-generated voices are 'artificial' under the Telephone Consumer Protection Act (TCPA)", and it "takes effect immediately". [S46]
- Consent comes first. Under FCC rules the TCPA "requires telemarketers to obtain prior express written consent from consumers before robocalling them", and the ruling holds AI-generated voices "to those same standards". [S46]
- Callers can sue. The FCC notes that "the TCPA allows individual consumers or an organization to bring a lawsuit against robocallers in court". [S46]
- People must be told. In the European Union, the AI Act requires providers to ensure that AI systems "intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system". [S44]
None of these rules cares whose model it was. They apply to the business that made the call, in that business's name.
The languages on the line
SARAH thinks in 119 languages, understands 99 and speaks 76 in her own voice, and switches between them in the same conversation. [S35] [S36] For the others, she understands the speaker and shows the translation on screen. Pillar 5 sets out why those are three numbers, not one.
A call, step by step
| Step | On a rented voice stack | On our voice path |
|---|---|---|
| The caller speaks | Audio leaves for a public transcription API | Recognized on our GPUs [S36] |
| The answer is decided | Text leaves for a public model, behind its rate limits | Decided on hardware we operate [S36] |
| The voice replies | Text leaves for a public voice API | Synthesized on our speech engine [S36] |
| A person is needed | Often a callback | Transferred, with the desk kept in the call [S36] |
| The call ends | Pieces of the record sit with each vendor | "The conversation, the transfer and the booking are part of the record" [S36] |
Questions for any voice AI vendor
- Name the three engines on your voice path, and the company that operates each.
- Is the audio of our callers ever sent to a third party, for any purpose, including improving its models?
- What happens to a live call when one of your vendors returns a rate limit? Our answer: "No. The turn is not a metered request to someone else's public model." [S36]
- How does your product tell a caller they are speaking with an AI, where the law asks for it? [S44]
- How do you record the consent the TCPA requires before an outbound AI call? [S46]
- Who answers when the voice is wrong? Ours: "The company on the contract." [S36]
The test call
Before you sign, make one call yourself, in the language your customers speak, at your busiest hour. Interrupt. Change language. Ask for a person. Then ask the vendor for the record of that call, and for the list of every company that heard it.
Part 7
The economics of rented AI
Our page calls it "A taxi meter that runs while your customer talks." [S35] "Per-token pricing sounds fair: pay only for what you use. Here is what it really means for a business that talks to customers all day." [S35]
This part uses no prices. Where it shows arithmetic, the numbers are hypothetical and labeled as such, so that a Chief Financial Officer can put the vendor's real numbers in their place.
Six things the meter does
Our page lists them [S35]:
- "You pay for every word, both ways." "Every word your customer says and every word the AI answers is counted and billed."
- "The meter speeds up as the call goes on." "Most per-token services read the whole conversation again on every turn, so you pay for the same words again and again."
- "The cheapest answer is the shortest one." "When every token is a cost, the pressure is always to say less, use a smaller model and cut the context."
- "Busy hours cost you twice." "At peak, rate limits slow or stop the answer, exactly when the most customers are calling."
- "The price is theirs, not yours." "Price, terms and the model itself can change tonight. Your budget follows."
- "You own nothing at the end." "After a year of paying the meter, you have no hardware, no model and often no memory of your own customers."
Why the meter speeds up: a worked example
The second point is arithmetic, and it is worth doing once. A language model has no memory of its own, so a metered service sends the whole conversation back to the model on every turn. [S35] Take a hypothetical call in which each turn, the caller's words and the answer together, adds 150 tokens.
| Hypothetical call | Tokens spoken | Tokens the model reads, if every turn re-reads the conversation | Read for every token spoken (hypothetical) |
|---|---|---|---|
| 10 turns (hypothetical) | 1,500 | 8,250 | 5.5 times |
| 20 turns (hypothetical) | 3,000 | 31,500 | 10.5 times |
| 40 turns (hypothetical) | 6,000 | 123,000 | 20.5 times |
In this hypothetical, a call twice as long is read about 4 times as many tokens, because each new turn re-reads every turn before it. Add a set of hidden instructions re-sent on every turn, and the gap widens. Real services may cache part of this; ask the vendor how it bills a long call, in writing.
Our own path has no token meter: "Omega has no token meter on the voice path ... the cost does not climb because your customer kept talking." [S35]
The buffet and the table
Our page explains the difference between a subscription and tenancy with a dinner: "At a buffet you eat what is left when you reach the front of the queue. A reserved table is set for you whether you arrive at nine or at midnight. Enterprise work needs the table." [S35]
| Subscription [S35] | Tenancy [S35] | |
|---|---|---|
| The machine | "A share of a machine other customers use at the same moment" | "Resources allocated to one customer" |
| When you get it | "Capacity you get only when it is left over" | "There twenty-four hours a day, used or not" |
| At peak | "Rate limits and queues at peak" | "No rate limit on the voice path" |
The cost the meter does not show
The meter is the visible cost. The larger costs of rented AI arrive later, and they arrive on the business, not the vendor.
- The breach. The global average cost of a data breach was $4.99 million in 2026. [S24] Detection and escalation, with lost business, made up 63% of that cost. [S50]
- The fine. Under the GDPR, up to 20 million euro or up to 4 % of total worldwide annual turnover, whichever is higher, for the most serious infringements. [S43] Under UK GDPR, up to £17.5 million or 4% of turnover. [S53] In Australia, up to the greater of A$50 million, three times the benefit obtained, or 30% of adjusted turnover. [S54]
- The card schemes. Visa introduced investigation fees and non-compliance assessments from 18 April 2020, and says "Fees are entirely avoidable if entities cooperate with applicable investigations in a timely fashion". [S51]
- The exit. At renewal, the memory the system built about your customers is an asset. If it stays with the vendor, you pay to rebuild it.
Security spending works both ways
IBM found that extensive use of AI and automation in security saved an average of $1.93 million per breach. [S50] AI in the hands of the defender is worth paying for. The question is whether the AI you buy for revenue also widens the doors you must defend.
Insurance is cheaper than it has been for years: Marsh reported global cyber insurance rates fell 4% in the third quarter of 2026, the thirteenth consecutive quarter of decline. [S52] Cheap cover is welcome. It pays after the event. It does not put the data back.
A worksheet for the Chief Financial Officer
Ask every AI vendor to fill in these lines, in writing, for your own call volumes. Put their numbers in; leave ours blank until you have them in writing too.
| Line | What to ask for | Why it matters |
|---|---|---|
| 1. Unit of billing | Per token, per minute, per seat or per outcome, and how a long call is billed | The meter speeds up as the call goes on [S35] |
| 2. Peak behavior | What happens at a rate limit, and whether you pay for retries | "Busy hours cost you twice" [S35] |
| 3. Price change | Notice period and your right to leave without penalty | "The price is theirs, not yours" [S35] |
| 4. Sub-processors | Every vendor in the chain, and who pays when one fails | Each is a contract and a breach notice [S41] |
| 5. Exit | What you take with you, in what format, at what cost | "You own nothing at the end" [S35] |
| 6. Incident cost share | Who pays for investigation, notice and downtime | Lost business is the largest share [S50] |
Part 8
Hospitality, gaming and contact centers
These are the industries where AI meets the public by voice, all day, and where one breach reaches every guest at once. They are also where the attacks of 2023 were most visible.
Hospitality and gaming: the record
In the 2024 Verizon data, as analyzed by the retail and hospitality sharing center RH-ISAC, Accommodation and Food Services had 220 incidents and 106 confirmed breaches, up from 68 the year before. [S49]
92% of breaches were executed by external actors and 100% of attacks were financially motivated. [S49] The data taken was credentials in 50% of breaches, personal data in 28% and payment data in 19%. [S49]
In 2024 Omni Hotels suffered a prolonged outage after a cyberattack. Guests reported that reservation and check-in systems, room key cards and payment systems were affected, and Omni later said the data taken did not include payment details. [S55]
Part 4 sets out MGM Resorts, Caesars Entertainment, Marriott and Wyndham. Taken together with the sector figures, the pattern is plain: outsiders, for money, through people and through doors that faced the internet.
How the AI pitch lands in a hotel or a casino
The pitch to a property is usually a voice receptionist, a booking chatbot and an AI concierge, each from a different vendor, each reaching the property's systems over the public internet. Every one of them is a new account with access to guest data, and a new party in the chain.
The 2023 casino attackers "Posed as company IT and/or helpdesk staff using phone calls or SMS messages". [S20] A property that adds AI on the phone must make sure that the AI cannot be talked into what the help desk was talked into, and that a person who can be talked into it cannot reach every system from one login.
| The property's question | The AI pitch | A compliance-ready system |
|---|---|---|
| Where does the guest's voice go? | To three public engines, across the internet [S35] | To engines on hardware we operate, on the private path [S36] |
| What can a receptionist bot reach? | Usually whatever its API key can reach | Only what its role allows, on its own VLAN |
| What happens to the other properties? | On a flat path, one intrusion can become dozens [S25] | One VLAN per site; one site can be cut off while the rest trades |
| Who sees the card? | Depends on every vendor in the chain | Stripe's hosted page; card details never pass through our systems [S34] |
Contact centers: the law of the call
A contact center running AI voices works inside rules that already exist. In the United States, the FCC recognizes AI-generated voices as "artificial" under the TCPA, and telemarketers must "obtain prior express written consent from consumers before robocalling them". [S46]
Consumers can sue under the TCPA, and the FCC says the ruling gives State Attorneys General "new tools". [S46] In the European Union, people interacting with an AI system must be informed that they are. [S44]
When a contact center takes card payments over the phone, PCI DSS applies to it "regardless of whether these activities are conducted directly or by a third-party service provider". [S10]
How the AI pitch lands in a contact center
The pitch to a contact center is volume: thousands of AI calls at once, on a small model, priced per minute or per token. Our page describes the trade: "Anyone can run thousands of agents at once on a small model; that is cheap and fast. But a small brain gives small answers, and it stumbles the moment a buyer switches language." [S35]
The bigger risk is the record. A contact center has to prove consent, prove disclosure and prove what was said, call by call. On a stack of rented engines that proof is spread across vendors.
| The contact center's question | The AI pitch | A compliance-ready system |
|---|---|---|
| Where is the consent recorded? | In the dialer vendor's system, if at all | In the ledger of the call, on the customer's record [S36] |
| Does the caller know it is an AI? | Depends on the script | Disclosure written into the operator's own rules [S44] |
| What happens at peak? | "Your caller waits or the call fails" [S35] | "No rate limit on the voice path" [S35] |
| Can a person take the call? | Often a callback | "She transfers when the work needs a person" [S36] |
Questions for these industries
- Which of your AI products can reach guest or customer data, and through which door?
- How does your voice AI respond when a caller claims to be from IT and asks for a reset?
- Show us the consent and disclosure record for one AI call, end to end.
- If one property or one site is compromised, how long does it take to isolate it, and what keeps running?
What a compliance-ready system gives a property or a contact center
- One VLAN per property, and per system inside it, so the front desk, the payment terminals and guest Wi-Fi never share a path.
- A help desk that cannot be the front door: named logins from whitelisted sites, so a reset talked out of the help desk reaches nothing from outside.
- Voice on the private path, with the guest's audio kept on the customer's record. [S36]
- Consent and disclosure in the ledger of every call, ready for the day someone asks. [S46]
- Card entry on Stripe's hosted page, so card details never pass through our systems. [S34]
- A drill for losing one property, timed and recorded, while the others keep trading.
What we do not claim here
- We do not claim to stop social engineering. The 2023 attacks began with people [S20], and a private network does not change what a deceived person can be talked into.
- We build the system that makes the evidence easier to show. Your PCI DSS validation and your consent records stay yours. [S10]
Part 9
Healthcare and financial services
In these two industries the regulator is already in the room. Every AI decision is a decision about protected data, and every supplier is part of an examination.
Healthcare: the record
In 2024 the cyberattack on Change Healthcare reached patients and providers across the United States. HHS explained its response this way: "Given the unprecedented magnitude of this cyberattack, its widespread impact on patients and health care providers nationwide". [S45]
OCR "prioritized and opened investigations of Change Healthcare and UnitedHealth Group (UHG)". [S45] It "reminded all of these entities of their HIPAA obligations to have business associate agreements in place and to ensure that timely breach notification" to HHS and affected individuals occurs. [S45]
That is the lesson for every AI supplier in healthcare. When a business associate is breached, every covered entity that relies on it is asked the same questions about its own agreements and its own notice.
What HIPAA asks of an AI supplier
- A contract, not a badge. A business associate cannot be "certified by a third party" instead of signing a business associate agreement. [S5]
- Down the chain. Each subcontractor that handles PHI needs its own business associate agreement. [S6] An AI vendor that sends PHI to a transcription API and a model API has two more agreements to show you.
- Direct liability. Business associates are directly liable for their own failures. [S7]
- Your own analysis. Every covered entity and business associate runs its own risk analysis. [S8] [S9]
How the AI pitch lands in a clinic
The pitch to a practice is an AI receptionist, an AI scribe and an AI that answers patient questions. Each one hears PHI. Each one, on a rented stack, sends it to the engines behind it.
Ask how many business associate agreements stand behind the demo. If the vendor says it is "HIPAA certified", it has named something HHS does not recognize. [S4]
| The practice's question | The AI pitch | A compliance-ready system |
|---|---|---|
| Who hears the patient's voice? | The vendor and the engines it rents | Engines on hardware we operate, with the audio kept on the customer's record [S36] |
| How many BAAs stand behind it? | Often more than the vendor can name | Ours as your business associate, where PHI is involved [S6] |
| Who runs the risk analysis? | "The vendor is HIPAA compliant" | You do, with our system description and evidence [S9] |
Financial services: the record
For public companies in the United States, the SEC's 2023 rules made cybersecurity a disclosure matter. "An Item 1.05 Form 8-K will generally be due four business days after a registrant determines that a cybersecurity incident is material." [S37]
The annual report must describe "the board of directors' oversight of risks from cybersecurity threats" and management's role in managing them. [S37] SEC Chair Gary Gensler put the logic plainly: "Whether a company loses a factory in a fire ... or millions of files in a cybersecurity incident ... it may be material to investors." [S37]
The FTC's Safeguards Rule applies to a wide range of financial institutions, and on service providers it is blunt: "the buck still stops with you". [S18] In Europe, the GDPR asks for "the pseudonymisation and encryption of personal data" where appropriate, and for systems with ongoing "confidentiality, integrity, availability and resilience". [S42]
How the AI pitch lands in a bank, a lender or an adviser
The pitch is an assistant for advisers, a chatbot for customers and an AI that reads documents. Each of those reads account data. On a rented stack, that data leaves the institution for a model it does not operate, under terms that can change.
Four business days is not long to learn what a vendor's vendor did with your data. The fewer parties in the chain, the faster the answer.
| The institution's question | The AI pitch | A compliance-ready system |
|---|---|---|
| Where does customer data go when the AI reads it? | To the model provider's machines | "Your work stays on the estate" [S36] |
| How fast can you tell us what happened? | After each vendor in the chain reports | One private path, one egress, one record [S36] |
| Does our data train a public model? | "Read the terms. Then read them again after the next update" [S35] | "No. Your work stays on the estate." [S36] |
Questions for these industries
- List every party that will receive PHI or account data, and the agreement that covers each.
- If one of your sub-processors is breached, how many hours until you tell us, and with what?
- Can the system run on our premises when our policy or our regulator requires it? [S36]
- What would your incident look like in our Form 8-K, and what would you need to tell us first? [S37]
What a compliance-ready system gives a regulated firm
- A shorter chain. The three engines on the voice path are hosted by us, so PHI and account data do not travel to three more vendors. [S36]
- The paperwork for your own analysis: a system description, the user entity controls you must run and a risk register, drafted for your adoption. [S2] [S9]
- A business associate agreement where we handle PHI for you, with our own direct liability. [S6] [S7]
- Your premises, when policy requires it, on SARAH Mini, SARAH Pro or SARAH Enterprise. [S36]
- One record to read when the clock starts, from one private path and one egress. [S36]
- Evidence from day one, so your Type 2 period starts on a system that already logs what an auditor tests. [S2]
What we do not claim here
- We are not HIPAA certified, because no such certification exists [S4], and we do not say we are.
- Your risk analysis, your disclosures and your filings are yours. [S9] [S37] We supply the system description and the evidence.
Part 10
Manufacturing, government and education
These three have little in common on the surface. Each has a reason to keep the door shut that goes beyond data: plant safety, public trust and the records of children.
Manufacturing: the record
NIST describes operational technology as "a broad range of programmable systems and devices that interact with the physical environment", including industrial control systems and building automation. [S39] In a plant, a compromised system is not only a data problem. It can stop a line.
NIST's guidance addresses OT's "unique performance, reliability, and safety requirements". [S39] A flat network that joins the office, the plant and the internet puts all three at the risk of the weakest.
How the AI pitch lands in a plant
The pitch is predictive maintenance, a chatbot for the line and an AI that reads the plant's data. Each wants a connection from the plant to someone else's cloud. Each connection is a path from the internet to the floor, however well it is defended.
A private Layer-2 path changes the shape of that problem. The plant's systems sit on their own VLANs. What must leave, leaves through one watched exit, on rules. Nothing comes in uninvited.
| The plant's question | The AI pitch | A compliance-ready system |
|---|---|---|
| Does the plant need an internet connection for the AI? | Yes, to reach the vendor's cloud | No. The AI runs on the private path [S36] |
| Can the office reach the line? | On a flat network, often yes | Only where a rule allows it, VLAN by VLAN |
| What happens if the AI vendor is down? | The feature stops | "A change in a public API's terms does not switch off a path that never called that API" [S36] |
Government: the record
For federal agencies, CISA's Binding Operational Directive 22-01 turned the Known Exploited Vulnerabilities catalog into a deadline: remediate "within 6 months" for vulnerabilities with a CVE ID assigned prior to 2021 "and within two weeks for all other vulnerabilities". [S40]
CISA adds a flaw "based on reliable evidence that the exploit is being actively used". [S40] PHP2Shell met that test five days after its fix was published. [S31]
NIST's zero trust guidance tells agencies there is "no implicit trust granted to assets or user accounts based solely on their physical or network location". [S38] CISA's Secure by Design program asks every provider to "take ownership at the executive level" for security, with "multi-factor authentication (MFA), logging, and single sign-on (SSO) available at no extra cost". [S48]
How the AI pitch lands in an agency
The pitch is a citizen chatbot and an assistant for staff, running in a commercial cloud. Ask the questions an agency's own rules already ask: how fast are catalog entries fixed, which components are exposed, and is MFA an extra?
Education: the record
FERPA "affords parents the right to have access to their children's education records" and "the right to have some control over the disclosure of personally identifiable information from the education records". [S47]
"When a student turns 18 years old, or enters a postsecondary institution at any age, the rights under FERPA transfer from the parents to the student." [S47] The statute is at 20 U.S.C. § 1232g and the regulations at 34 CFR Part 99. [S47]
How the AI pitch lands in a school
The pitch is a tutor, an admissions chatbot and an AI that answers parents. Each one can touch education records. A school that sends those records to a vendor's vendor has to know who holds them and under what control.
| The institution's question | The AI pitch | A compliance-ready system |
|---|---|---|
| How fast are known exploited flaws fixed? | Whenever the vendor's host gets to it | Inside the clock the buyer sets [S40]; no PHP to patch at all |
| Is MFA an extra? | Often on a higher tier | CISA says it should be included at no extra cost [S48] |
| Who holds student or citizen records? | The vendor and its sub-processors | The institution's own system, on its own server |
Questions for these industries
- Which of your components face the public internet, and what is your remediation clock for catalog entries? [S40]
- Does your AI need a path from our plant, agency or campus to the internet?
- Who outside our institution can read the records your AI touches?
- Are MFA, logging and single sign-on included, or sold as extras? [S48]
What a compliance-ready system gives a plant, an agency or a school
- No path from the internet to the floor: the AI runs on the private path, and plant, office and guest systems sit on their own VLANs.
- No PHP to patch, and a written answer the day a catalog entry lands. [S31]
- Named logins from whitelisted sites, on a need to know, in line with the zero trust idea that location alone earns no trust. [S38]
- Records on the institution's own server, hosted for it or installed on its premises.
- One record of what left, for the parent, the auditor or the inspector who asks.
What we do not claim here
- A private network does not replace a plant's safety systems, an agency's own authorization process or a school's duties under FERPA. [S47]
- Zero trust is a model, not a product. [S38] Named logins, least access and the record are how we apply it; your own policies decide the rest.
Part 11
Ten questions that end the conversation
Our page says: "Ask these, and watch the room go quiet." [S35] Put these ten to any AI vendor, in writing, including us. Each has a factual answer. A vendor either has it or does not.
What you are actually being sold
Our page names no company. It names ten categories, what each sells, and what you actually get. [S35] Six of them:
| Category | What they sell | What you actually get, in our published view |
|---|---|---|
| Consumer chat apps | An AI assistant for everyone | A shared window with one rule book, no memory of your business, no telephone and no ledger |
| Model APIs, priced per token | The smartest model, on tap | A meter. Your data leaves for their machines, and the bill grows with every word |
| Hyperscaler AI platforms | Enterprise AI in your own cloud account | A rented partition in their region. A private cloud you rent is still their cloud |
| Agent platforms | Autonomous agents that do the work | Scripts calling someone else's model, with no single record of what left |
| Voice AI and receptionist widgets | An AI that answers your phone | Three public hops that can each be listened to or switched off |
| Integrators and resellers | We will build AI for you | The categories above, stitched together, with every limit still inside |
Our page's advice for the meeting: "Ask them, and listen for the word partner, API or cloud." [S35] Below, "what you will usually hear" describes these categories, not any named company.
1. Where does the model run, and who operates the GPUs?
- What you will usually hear. "We use the leading models." That is an API call to someone else's hardware.
- Our answer. On GPUs we operate, on our hosted estate or at the customer. "Not on a shared public endpoint, and not inside someone else's hall." [S36]
- Verdict. If the answer is "our cloud partner", the conversation is over. Every other answer depends on this one.
2. Who can read our prompts, documents and call recordings?
- What you will usually hear. A link to a privacy policy. Our page's advice: "Read the terms. Then read them again after the next update." [S35]
- Our answer. Call audio "stays with the customer's record". Your work is not "a donation of your corpus to a frontier lab". The audit is read by "The customer. An examiner, when the customer has one." [S36]
- Verdict. If the honest answer is "people you will never meet, under terms that can change tonight", you have your answer.
3. List your sub-processors, and the terms you flow down to each
- What you will usually hear. A web page of names: the cloud host, the model API, transcription, the voice, the CRM, the email sender.
- Our answer. The three engines on the voice path, language, speech recognition and speech synthesis, "are hosted by us". [S36] Any subcontractor that touches your data is named in writing, with a BAA flowed down where PHI is involved, as HIPAA requires. [S6]
- Verdict. Every name on that list is another agreement, another breach notice to wait for and another answer to chase. Count them.
4. What of your service is reachable from the public internet?
- What you will usually hear. All of it: the app, the API and the admin console, "behind a login and a firewall".
- Our answer. The production path "is never exposed to the public internet, so there is no public door on it to knock on". Public websites are, and "are defended the usual way". [S35]
- Verdict. A door behind a lock is still a door. Ask how many there are.
5. Does any part of your stack run PHP, and what did you do on 21 October 2026?
- What you will usually hear. "We are checking with our host." The website, its plugins and its page builder run on PHP. [S32]
- Our answer. No PHP anywhere, by design. Our written statement went out five days after the advisory. [S28] (Part 3)
- Verdict. PHP2Shell needed only "an unauthenticated request". [S28] If they are still checking, so is the attacker.
6. Whose certificate is it?
- What you will usually hear. "SOC 2 compliant", with a cloud provider's or model provider's report attached.
- Our answer. A report covers the organization and the system it names. [S2] ORE holds none by design and claims none. We build compliance-ready systems so that you can get certified.
- Verdict. Someone else's report answers someone else's audit. A vendor that says "HIPAA certified" has named something HHS does not recognize. [S4]
7. What happens to our data, memory and record at renewal?
- What you will usually hear. An export, in their format. What the system learned about your customers stays behind.
- Our answer. "The system of record stays the customer's." The proposal the engine writes "is part of your record". [S36] The server is the customer's.
- Verdict. If your customer memory is an asset on their balance sheet, you are renting your own business back.
8. Can it run on our premises?
- What you will usually hear. "Cloud only", or a "private cloud" in the vendor's region.
- Our answer. "Yes. That path is SARAH Mini, SARAH Pro or SARAH Enterprise, on your premises." [S36]
- Verdict. A private cloud in their region is their cloud. Ask whose floor the server stands on.
9. How many languages, natively, by voice?
- What you will usually hear. "Over 100", with reading, understanding and speaking counted as one number.
- Our answer. "Thinks in 119, understands 99, speaks 76 in her own voice, and never claims more than is true." [S35]
- Verdict. Ask for the three numbers separately. If you get one, it was rounded up.
10. Can you integrate without rip and replace?
- What you will usually hear. "Migrate onto our platform first."
- Our answer. "Never rip and replace. Always integrate with what you already run." [S35] "She connects to them." [S36]
- Verdict. A vendor that must replace your systems before it can help you is selling a migration, not an answer.
Part 12
The procurement pack
Questions are only as good as the paper they are answered on. This part turns the argument of this paper into questions each department can send, and contract clauses General Counsel can adapt.
Practice, not legal advice. The clauses below are examples of practice. They are not legal advice, they are not drafted for any jurisdiction, and every one needs General Counsel's review before use. Square brackets mark the numbers your business must choose.
For the Owner and the Chief Executive Officer
- Which company is on the contract, and which company answers when the AI is wrong? [S36]
- Which parts of the proposal run today, and which are designs, with dates?
- If this vendor disappeared tomorrow, what would we still own: the hardware, the model, the memory, or nothing? [S35]
- Would we be comfortable reading this vendor's answers aloud to our largest customer?
- Who on our side owns this system, and its controls, after the vendor's team goes home?
For the Chief Financial Officer
- How is a long call billed, and what happens to the cost at peak? [S35]
- How much notice do we get of a price change, and can we leave without penalty?
- Who pays for investigation, notice and downtime after an incident in the vendor's chain? [S50]
- Which costs disappear if we choose a system we own, and which appear?
- What does it cost to leave, in money and in time?
For the Chief Information Officer
- Name every network the data and the audio cross, and the operator of each. [S36]
- List every address that answers the public internet, and every component that runs PHP.
- How long does it take to isolate one system, and one site, and when was that last tested?
- What must we migrate before the product works, and what is the rollback plan?
- Which parts of the service can run on our premises, and which cannot? [S36]
For the Security Officer
- Send the latest SOC 2 report or ISO/IEC 27001 certificate, with the organization, system and period or scope it covers. [S2] [S14]
- Send the list of complementary user entity controls, and mark the ones that fall to us. [S2]
- What is your remediation clock for Known Exploited Vulnerabilities, in days? [S40]
- Are MFA, logging and single sign-on included at no extra cost? [S48]
- Who on your side can reach our data, from where, and how is that access reviewed?
For General Counsel
- List every sub-processor, the notice we get of changes, and our right to object. [S41]
- Where PHI is involved, send the BAA and the BAAs flowed down to each subcontractor. [S6]
- Will our data, prompts or call audio ever be used to train any model outside our system? [S36]
- How does the product record consent and disclose that a caller is speaking with an AI? [S44] [S46]
- Which law governs the contract, and where would a dispute be heard?
For the head of the contact center or operations
- What happens to a live call when a vendor in the chain returns a rate limit? [S36]
- How does a caller reach a person, and does the person see what was already said?
- Which languages does the voice speak natively, and which are translated? [S36]
- At our busiest hour, what is the longest a caller will wait for an answer, and who measures it?
Contract clauses: practice, not legal advice
The first set covers the chain, the clock and the exposure.
| Topic | Example clause (practice, not legal advice) | Why it is there |
|---|---|---|
| Sub-processors | Supplier shall not engage any sub-processor without Customer's prior written authorization, shall give [30] days' written notice of any intended change, and Customer may object. | GDPR Article 28 builds this chain into law [S41] |
| Liability for the chain | Supplier remains fully liable to Customer for the performance of each sub-processor's obligations. | Mirrors Article 28(4) [S41] |
| Incident notice | Supplier shall notify Customer in writing within [24] hours of becoming aware of any incident affecting Customer data, with the facts then known. | Customer's own clocks are short, such as the SEC's four business days [S37] |
| Known exploited flaws | Supplier shall remediate any vulnerability in the service listed in CISA's KEV catalog within [14] days of listing, and confirm in writing. | Matches the federal clock for recent CVEs [S40] |
| Public exposure | Supplier shall maintain and provide on request a list of every component of the service reachable from the public internet. | A door you know about is a door you can close |
The second set covers the data, the record and the exit.
| Topic | Example clause (practice, not legal advice) | Why it is there |
|---|---|---|
| No training | Supplier shall not use Customer data, prompts or call audio to train or improve any model outside Customer's own system. | "Read the terms. Then read them again after the next update" [S35] |
| Data location | Customer data and call audio shall be processed only at the locations listed in [Schedule], and on networks listed there. | You cannot defend a path you cannot name |
| The record | Supplier shall provide Customer, on request, the record of every outbound call made by any connector on Customer's behalf. | One egress, one record [S36] |
| Exit | On termination Supplier shall return all Customer data, including derived customer memory, in [format], within [30] days, then delete it and confirm in writing. | "You own nothing at the end" [S35] |
| Price change | Supplier shall give [90] days' written notice of any price change; Customer may terminate without penalty before it takes effect. | "The price is theirs, not yours" [S35] |
The third set covers the controls and the voice.
| Topic | Example clause (practice, not legal advice) | Why it is there |
|---|---|---|
| Assurance reports | Supplier shall provide its current SOC 2 report or ISO/IEC 27001 certificate each year, stating the system and scope, and its list of CUECs. | A report covers only the system it names [S2] [S14] |
| BAA | Where PHI is involved, Supplier shall sign a business associate agreement and flow it down to each subcontractor that handles PHI. | HIPAA requires it [S6] |
| Card data | Supplier shall provide a written responsibility matrix for every PCI DSS requirement it performs or affects. | The matrix sets out who does what [S12] |
| AI disclosure | Supplier shall configure the service so that callers are informed they are interacting with an AI system where law requires it. | EU AI Act Article 50 [S44] |
| Consent records | Supplier shall keep, for [period], the record of consent for every outbound AI call made on Customer's behalf. | The TCPA consent rule [S46] |
| Isolation drill | Supplier shall test, at least [yearly], the isolation of one system and one site, and give Customer the timed result. | A drill is evidence; a promise is not |
A vendor that refuses every one of these clauses has told you something. A vendor that accepts them all, in writing, has given your auditor a place to start.
Part 13
Payments, and what you certify yourself
If a business takes card payments, PCI DSS applies to it, "regardless of whether these activities are conducted directly or by a third-party service provider". [S10]
Outsourcing narrows what applies, but "this does not remove the merchant's responsibility to ensure account data is properly protected by the third party". [S10]
Card payments in our systems
In our systems, card payments are taken on Stripe's own hosted payment page. The buyer is sent to Stripe to pay. Stripe is the payment provider, Stripe's payment form is not embedded in our sites, and card details never pass through or rest on our systems.
Stripe's own documentation describes a payment page reached "via a redirect to a Stripe-hosted page" [S34], and integrations that send payment details "directly to Stripe without it passing through your servers, reducing your PCI obligations". [S33]
Reducing is not removing. Stripe: "PCI compliance is a shared responsibility and applies to both Stripe and your business." [S33] The merchant still validates its own compliance, typically through a Self-Assessment Questionnaire such as SAQ A [S10], as its acquirer decides.
- Requirement 12.8 is the merchant's: managing its service providers, with written agreements and an annual check. [S10] [S11]
- Requirement 12.9 is the provider's. It applies only to a service provider. [S11]
- The responsibility matrix sets out who performs each requirement. "Ultimately, the entity is responsible for ensuring its own PCI DSS compliance." [S12]
- Segmentation may help. It "is not a PCI DSS requirement" but "may reduce" the scope and cost of the assessment; in a flat network, every system can be in scope. [S13] Our VLAN per system design may let your assessor agree a smaller scope. The assessor decides, not us.
What you certify yourself
The certificate, when it comes, carries your name. This is the split.
| Area | We provide | You, the tenant, do |
|---|---|---|
| Scope | A description of the system and its boundaries | Decide what is in scope |
| Policies | A policy pack drafted for adoption in your name | Adopt and enforce it; name your Security Officer and Privacy Officer |
| Control owners | Controls built in, with owners proposed | Appoint a named owner for every control |
| People | Named logins, roles and access reviews | Train staff, keep records, run background checks |
| Vendors | Our evidence and, where PHI is involved, our BAA [S6] | Run vendor management, including of us |
| Evidence | Logs, change records, access reviews, drill results | Collect evidence for everything outside the system |
| The auditor | Engineer time for walkthroughs | Choose and engage the CPA firm or certification body [S1] [S14] |
| HIPAA | A BAA as your subcontractor, where PHI is involved [S6] | Your own risk analysis [S9] and your own BAAs |
| PCI DSS | Card entry on Stripe's hosted page [S34]; a statement of what we manage [S12] | Your own SAQ or assessment [S10] |
| After the report | Keep the system and its evidence running | Operate the controls, renew the report, answer your buyers |
Part 14
What we claim, and what we do not
A paper that asks you to test every vendor has to pass its own test. Here is ours.
What compliance-ready means
- A private path, a VLAN per site, per customer, and per device type, and one controlled exit. [S35] [S36]
- Tenancy, not subscription: resources allocated to one customer, there all day whether used or not. [S35]
- No PHP by design, with a written answer the day a buyer asks (Part 3).
- Named logins from whitelisted sites, on a need to know, with access reviews kept as evidence (Part 4).
- Isolation drills, timed and recorded: one system or one site cut off while the rest keeps running.
- Logs, retention and change records, on a written schedule, with who, when and why.
- The record belongs to the customer. "We don't sell your data." [S35]
What it does not mean
- It is not a certificate. It does not mean that we, or you, are certified, attested or audited. That comes only from your own auditor or certification body.
- It does not transfer. Your policies, your people, your contracts and your customers stay yours.
- It does not replace your audit. It shortens the work of preparing for it. Your auditor and your assessor decide.
- It gives no immunity from attack. No network does. The 2023 casino attacks began with people, not open ports. [S20]
- Public websites still live on the public internet, "and are defended the usual way". [S35]
- Two pillars are designs. The Last Mile Mini Data Center and Off-Grid Power are at design stage, with no delivery date.
- No figures of ours. No prices, no results, no customers. Every external figure carries its source.
How to check us
- Read the sources. Every external fact in this paper is marked and listed, with the address where we read it.
- Read our pages. Our 100 questions and our page for experts are public. [S35] [S36] Hold us to every answer.
- Ask for the evidence. The system description, the isolation drill results and the PHP2Shell Security Statement, in writing.
- Make the test call. In your language, at your busiest hour, and ask for the record afterward (Part 6).
- Bring your auditor. Walk them through the system before you sign, not after.
Part 15
The challenge
Here is the challenge.
Take this paper into the next AI pitch you sit through. Ask the four questions. Then ask the other six. Ask for every answer in writing, with a date and a name on it.
If a vendor can answer all ten, buy from them. We would rather you bought the right system than ours.
If they cannot, you already know what you are being sold: someone else's GPUs, someone else's rules, someone else's certificate and a door on the public internet, with your name on the contract.
Take the same ten to us. Take our design to your own auditor and your own insurer. Then decide on the evidence.
If you only have time for four
Our question 99: "Where the GPUs are. Where the language model is hosted, and how large it is. Where speech recognition runs. Where the voice is synthesized. If those four are a public API on the public internet, the path can be listened to, and the rest of the brochure is a rental." [S36]
Our own answer to the second: "Our smallest brain is 235 billion numbers big." [S35]
The ten, on one page
| No. | Ask, in writing | The answer that ends it |
|---|---|---|
| 1 | Where does the model run, and who operates the GPUs? | "Our cloud partner" |
| 2 | Who can read our prompts, documents and call recordings? | "See our privacy policy" |
| 3 | List your sub-processors and the terms you flow down | A page of names and no terms |
| 4 | What of your service is reachable from the public internet? | "All of it, behind a login" |
| 5 | Does any part of your stack run PHP? | "We are checking with our host" |
| 6 | Whose certificate is it? | Someone else's |
| 7 | What happens to our data, memory and record at renewal? | "You can export it" |
| 8 | Can it run on our premises? | "Cloud only" |
| 9 | How many languages, natively, by voice? | One number, rounded up |
| 10 | Can you integrate without rip and replace? | "Migrate first" |
Our page ends with the only invitation we make: "Bring us the problem, not the idea. Tell us the revenue problem. We will show you what the engine does with it, on hardware we operate." [S35]
We build compliance-ready systems so that you can get certified.
Appendix A
Our answers on the record, from our 100 questions
Our site publishes one hundred questions to put to any AI provider, with our answers on the record. [S36] The answers below are quoted exactly as published on 4 November 2026, with the question number as published. Questions that concern our prices or our staff are left out of this edition. PEIPN is our Layer-2 Private Enterprise IP Network.
Where it runs [S36]
| No. | The question | Our published answer |
|---|---|---|
| 1 | Do you own the GPUs the conversation runs on? | Yes. The language model, speech recognition and the voice run on GPUs we operate. We do not manufacture the chips. |
| 2 | Do you host your own large language model? | Yes. We host the model the conversation uses. We did not manufacture the base weights. It runs on our Layer-2 Private Enterprise IP Network. |
| 3 | Do you host your own speech-to-text? | Yes. Recognition runs on our GPUs. The utterance is not posted to a public transcription API. |
| 4 | Do you host your own text-to-speech? | Yes. The voice the caller hears is synthesized on our speech engine, on our GPUs. |
| 7 | Does the stack still phone a public model for the answer and call itself sovereign? | No. Inference for the conversation stays on hardware we operate. |
| 9 | If a public model vendor changes its terms tonight, does the telephone still answer? | Yes. A change in a public API's terms does not switch off a path that never called that API. |
| 10 | Do you manufacture the chips, or the base model? | Neither. We do not manufacture hardware. We do not manufacture the base model. We use a great deal of hardware, and we host the model ourselves. |
| 13 | Can procurement see where speech, language and voice ran? | Yes. They ran on GPUs we operate: on our hosted estate, or on a SARAH Mini, SARAH Pro or SARAH Enterprise at the customer. Not on a shared public endpoint, and not inside someone else's hall. |
| 14 | Is a dedicated GPU a line you resell from a hyperscaler's bill? | No. The metal is operated as ours, or placed on the customer's site. It is not a rented partition with a sovereignty label. |
| 15 | Can you name the three engines on the voice path? | Language, speech recognition and speech synthesis. All three are hosted by us. |
| 16 | Does the voice path share its GPU with a public chatbot product you do not control? | No. The path is ours. |
| 18 | Can the speech engine be swapped for a public voice API without a decision? | No. The production voice is our hosted speech engine. Changing it is a decision, not a silent failover. |
| 19 | Do you collect a voice sample and send it to a third party to render? | No. Synthesis stays on our engine. |
| 20 | Is there a second speech engine for languages the primary engine does not speak? | Yes. It is additive. The primary engine stays in place. A language is claimed only as far as it is true. |
The meter and the queue [S36]
| No. | The question | Our published answer |
|---|---|---|
| 6 | Does the call stop when a public model returns a rate limit? | No. The turn is not a metered request to someone else's public model. |
| 11 | Who pays a token invoice while the customer is speaking? | There is no per-token invoice to a public model for the voice path. The commercial shape is the published tier. |
| 31 | Do you shorten the answer to protect a public API bill? | No. The hosted brain is sized for the work. A customer answer is not cut to save a token meter we do not use. |
| 52 | Is there a surprise meter for looking up your own customer? | No. The commercial shape is the published tier. |
| 65 | Will the invoice match the page? | Yes. The figure on the page is the figure. We do not publish one price and invoice another. |
The conversation [S36]
| No. | The question | Our published answer |
|---|---|---|
| 5 | How many languages does she actually speak? | SARAH speaks 76 languages in her own voice, and she switches between them in the same conversation. We do not round the number up. A language is claimed only as far as it is true. |
| 21 | Is SARAH a skin on a public chat widget? | No. SARAH is the advisor on the telephone, the browser and the public number, with one memory. |
| 22 | Does she answer, book and transfer to a person? | Yes. She transfers when the work needs a person. The desk you chose to keep stays in the call. |
| 23 | Is there one memory when the channel changes? | Yes. Browser, handset and public number are the same account of the customer. |
| 24 | Will she claim a language she cannot speak? | No. She speaks 76 languages in her own voice. The claim stops where the voice stops. |
| 27 | Does a prompt file count as the product? | No. The product is the path: recognition, language, voice, memory, transfer and the ledger. |
| 29 | Does the call leave a ledger the customer can read? | Yes. The conversation, the transfer and the booking are part of the record. |
| 30 | Is the voice a different company from the revenue engine? | No. SARAH holds the conversation. Omega Revenue Engine holds the commercial work. SOPHIA holds the connections. PEIPN carries them. SARAH Mini, SARAH Pro and SARAH Enterprise are the premises products (SARAH Enterprise keeps its second server in a data center on a different power grid). The Last Mile Mini AI data center is designed as OMEGA Hyperscale, our own building. |
| 32 | Can the customer's own people stay on the line? | Yes. SARAH routes to them. |
| 33 | Is there a public number as well as a browser? | Yes. Browser, SIP and the public number are the same SARAH. |
| 34 | Who answers when she is wrong? | The company on the contract. A rented model vendor is not the party. |
| 35 | What is one spoken turn? | One statement, or one question. She does not stack a second question on the same turn to rush the caller. |
| 37 | Where does a booking go? | Into the customer's calendar and the ledger of the call. Not into a separate bot that forgets the telephone. |
| 38 | Is a conference a different memory from the desk? | No. The conference and the telephone are the same house, and the customer still has one account of the work. |
| 39 | What will you not call her? | A chatbot skin. She is the voice of the system the customer is running. |
| 40 | Does the caller have to repeat themselves because they moved from the browser to the handset? | No. The channel changed. The account of the customer did not. |
Connections and data [S36]
| No. | The question | Our published answer |
|---|---|---|
| 17 | What happens to the audio of the call? | It stays with the customer's record. It is not handed to a model vendor to improve a clone library. |
| 41 | What is SOPHIA? | The Layer-3 hub. Connectors, egress, and the record of what left the building. |
| 42 | How many enterprise features are in the catalog? | 34.7 million. That is the published figure. |
| 43 | Does every connector leave from a different key under a different desk? | No. One egress. One place to see what was called. |
| 45 | Does SOPHIA replace the customer's bank, airline or CRM? | No. She connects to them. The system of record stays the customer's. The call is governed. |
| 46 | Is a connector the same thing as sending the whole customer file to a model vendor? | No. A connector is a governed call. The model does not receive a dump it was not asked to use. |
| 47 | Who holds identity across channels? | One person record. The customer is not a new stranger on each channel. |
| 48 | Can a system be added without opening a second memory? | Yes. The connection joins the same account of the customer. |
| 49 | Is the catalog a slide of logos? | No. It is dispatch the product can call. A logo is not a connector. |
| 50 | Where do privileged calls travel? | On PEIPN when they are ours to carry. Through the governed egress when the outside system is the point. |
| 51 | Can a connector be turned off? | Yes. What you do not need is not left running because a bundle included it. |
| 53 | Who can read the audit? | The customer. An examiner, when the customer has one. |
| 54 | Is the hub a marketplace token shared with other tenants? | No. The hub is the customer's, on the estate they contracted. |
| 55 | What is allowed to leave? | Only what the connector was asked to send, through the egress, with a record. |
| 72 | Does my data become training material for a public model? | No. Your work stays on the estate. Using the product is not a donation of your corpus to a frontier lab. |
The commercial work [S36]
| No. | The question | Our published answer |
|---|---|---|
| 57 | Do I get more than a dialer with an AI label? | Yes. You get three parts that work as one: SARAH holds the conversation, the engine does the commercial work, and SOPHIA makes the connections to the systems you already run. |
| 60 | Will my proposal know what was said on the call? | Yes. The proposal and the call that produced it live in the same house, on the same account of the customer. Nobody retypes the conversation into a second tool. |
| 61 | Will my campaigns and my calls agree on who the buyer is? | Yes. They draw on one memory. The buyer does not become a different person because the channel changed, so your marketing and your sales team work from the same facts. |
| 73 | Do I keep the proposal the engine writes? | Yes. It is part of your record. |
| 74 | What does the engine deliver, in one unit of work? | A customer, a conversation, a connection and a commercial outcome. That is what you are paying it to produce. |
The private network [S36]
| No. | The question | Our published answer |
|---|---|---|
| 76 | What is PEIPN? | The Layer-2 Private Enterprise IP Network. Ethernet, not the public internet. It joins a customer's SARAH Mini, SARAH Pro or SARAH Enterprise when that work stays on their floor, and it is designed to join our Last Mile Mini data centers to each other. |
| 77 | Must voice and GPU traffic cross a multi-tenant hyperscaler backbone to reach the model? | No. That traffic is carried on the private path. |
| 78 | Can you audit what crossed the boundary? | Yes. Private path, single egress, one place to read the record. |
| 79 | Is a virtual private cloud in someone else's region the same thing? | No. A cloud you rent is still their network. PEIPN is the private enterprise path for this stack. |
| 80 | Where do the public addresses live? | Public services can terminate at a data center. The production path does not. Voice and GPU traffic ride Ethernet on PEIPN. A path on the public internet is a path that can be listened to. This one is not that path. |
| 81 | Does a browser widget on the public web count as the private network? | No. The widget is how a visitor arrives. PEIPN is how privileged traffic is carried. |
| 83 | If a cloud region is withdrawn, is the path still yours? | The path is part of the metal and the contract. It is not a feature flag in someone else's console. |
| 84 | Do other tenants share the customer's private VLAN? | No. Each customer path is its own VLAN on PEIPN. The Last Mile Mini is designed as our own building, not a shared hall, and it is not Omega Mini. |
| 85 | Is PEIPN a laptop agent the user installs and forgets? | No. It is the network. Not a piece of software on a personal computer. |
| 86 | Can that path carry the model, the voice and the connectors without turning each into a public API? | Yes. That is what it is for. |
| 87 | Is the private network an add-on sold after the demo? | No. It is part of the foundation, with SARAH, SOPHIA, Omega Revenue Engine and the Last Mile Mini AI data center, which is designed as OMEGA Hyperscale. Omega Mini is a separate product. |
| 88 | What sentence replaces a slogan? | We do not manufacture hardware. We use a lot of it. We do not manufacture the model. We host our own, on this network. |
| 99 | What do you ask a provider that cannot answer the first four? | Where the GPUs are. Where the language model is hosted, and how large it is. Where speech recognition runs. Where the voice is synthesized. If those four are a public API on the public internet, the path can be listened to, and the rest of the brochure is a rental. Ours is Ethernet on PEIPN. |
Question 34 is this whole paper in one line: the company on the contract answers. For your customers, that company is you.
Appendix B
Glossary
The terms this paper uses, in plain words. Where a term comes from a standard or a regulator, the source is given.
Assurance and certification
| Term | What it means |
|---|---|
| SOC 2 | An attestation report by an independent CPA firm on one service organization's system, against the AICPA's trust services criteria. It is a report, not a certificate. [S1] [S2] |
| Type 1 and Type 2 | Type 1 covers the design of controls at a point in time; Type 2 adds whether they operated effectively over a period. [S2] |
| CUEC | Complementary user entity control: a control the service organization assumed its customers would run, listed in its report. [S2] |
| Specified parties | The readers a SOC 2 report is intended for, agreed by management and the auditor. [S2] |
| ISO/IEC 27001 | The international standard for information security management systems. Certificates are issued by independent certification bodies, not by ISO, for a named organization and scope. [S14] [S15] |
| SIG and CAIQ | Standard security questionnaires buyers send to suppliers: the SIG from Shared Assessments and the CAIQ from the Cloud Security Alliance. [S16] [S17] |
| Compliance-ready | Our term for a system built so the controls an auditor tests can be shown, with the paperwork drafted. It is not a certificate, an attestation or an audit. |
Law, privacy and payments
| Term | What it means |
|---|---|
| HIPAA | The US law on the privacy and security of health information. HHS does not recognize private certifications under the Security Rule. [S4] |
| PHI | Protected health information: health information held by a covered entity or business associate. [S6] |
| BAA | Business associate agreement: the contract HIPAA requires between a covered entity and a business associate, and down to each subcontractor that handles PHI. [S5] [S6] |
| Risk analysis | The assessment HIPAA requires of each covered entity and business associate, of risks to the PHI it holds. [S8] [S9] |
| GDPR processor | A party that processes personal data on behalf of a controller, bound by contract and unable to engage another processor without written authorization. [S41] |
| Form 8-K Item 1.05 | The SEC filing a US public company makes about a material cybersecurity incident, generally due four business days after it determines materiality. [S37] |
| Safeguards Rule | The FTC rule that requires covered financial institutions to protect customer information, including oversight of service providers. [S18] |
| TCPA | The US Telephone Consumer Protection Act. The FCC treats AI-generated voices as "artificial" under it. [S46] |
| FERPA | The US law giving parents, and students from 18, rights over education records. [S47] |
| PCI DSS | The card industry's data security standard. It applies to any entity that stores, processes or transmits account data, including through a third party. [S10] |
| SAQ and SAQ A | Self-Assessment Questionnaire: how many merchants validate their own PCI DSS compliance. SAQ A is for merchants that have outsourced all card handling, as their acquirer decides. [S10] |
| Requirement 12.8 | The merchant's duty to manage its service providers, with written agreements and annual monitoring. [S11] |
| Responsibility matrix | A written split of which party performs each PCI DSS requirement. [S12] |
Networks and security
| Term | What it means |
|---|---|
| Layer 2 | The link layer: traffic moves between points joined on the same network segment, such as a VLAN, without public routing. |
| Layer 3 | The network layer: devices have IP addresses, and routers pass traffic between addresses anywhere, as on the internet. |
| VLAN | Virtual local area network: a separate network segment on shared equipment, so traffic on one VLAN does not reach another unless a rule allows it. |
| PEIPN | Our Layer-2 Private Enterprise IP Network: "Ethernet, not the public internet." [S36] |
| Egress | The exit from a network to the outside world. One controlled egress means one place to watch and record what leaves. |
| Segmentation | Dividing a network so systems cannot reach each other by default. For PCI DSS it is not required, but may reduce the scope of an assessment. [S13] |
| Zero trust | A model that grants no implicit trust because of network location, and authenticates and authorizes each session first. [S38] |
| KEV | CISA's Known Exploited Vulnerabilities catalog: flaws with reliable evidence of active exploitation. [S40] |
| CVE and CVSS | CVE is the public identifier for a vulnerability; CVSS is the scoring system for its severity. [S29] |
| OT | Operational technology: programmable systems and devices that interact with the physical environment. [S39] |
| Secure by Design | CISA's program asking technology providers to own security outcomes at the executive level. [S48] |
AI
| Term | What it means |
|---|---|
| Token | "Text is cut into tokens, pieces of words. The model reads and writes one token at a time." [S35] |
| Weights | "The model is billions of numbers learned in training, stored as a file." [S35] |
| Inference | Running the model: "the GPU pushes the whole conversation through every layer of those numbers to predict the next one". [S35] |
| Harness | "Everything around the model that makes it yours": context, rules, memory, connections and voice. [S35] |
| Tenancy | "Resources allocated to one customer, there twenty-four hours a day whether they are used or not." [S35] |
| Rate limit | A cap a shared service puts on requests. At peak, "rate limits slow or stop the answer". [S35] |
| Agent | "A script that calls a model." [S35] |
Sources
Every external fact in this paper is marked [R: Sn] and comes from the public source below, as opened between 3 and 4 November 2026. S35 and S36 are our own published pages, quoted as published. The PHP2Shell answers in Part 3 are quoted from our PHP2Shell Security Statement of 26 October 2026. Chris Ismail's quotations are his own words, from 3 November 2026. In Parts 8 to 11, "the AI pitch" and "what you will usually hear" describe categories of provider, not any named company. The figures in the worked example in Part 7 are hypothetical. Quotations from regulations keep their original spelling. The contract clauses in Part 12 are practice, not legal advice. Statements about contract practice are practice, not statistics. This paper is not legal advice.
S1 >S1 AICPA and CIMA, "System and Organization Controls: SOC Suite of Services", resources page, opened 3 November 2026. https://www.aicpa-cima.com/resources/landing/system-and-organization-controls-soc-suite-of-services
S2 >S2 AICPA Assurance Services Executive Committee, DC Section 200, "2018 Description Criteria for a Description of a Service Organization's System in a SOC 2 Report" (read from the archived copy of the AICPA's own file). https://www.aicpa.org/content/dam/aicpa/interestareas/frc/assuranceadvisoryservices/downloadabledocuments/dc-200.pdf
S3 >S3 Journal of Accountancy (AICPA), "AICPA guides peer reviewers to address SOC 2 risks", May 2026. https://www.journalofaccountancy.com/issues/2026/may/aicpa-guides-peer-reviewers-to-address-soc-2-risks/
S4 >S4 US Department of Health and Human Services, HIPAA FAQ 2003, "Are we required to 'certify' our organization's compliance with the standards of the Security Rule?" https://www.hhs.gov/hipaa/for-professionals/faq/2003/are-we-required-to-certify-our-organizations-compliance-with-the-standards/index.html
S5 >S5 US Department of Health and Human Services, HIPAA FAQ 237, "Instead of entering into a contract, can business associates self-certify or be certified by a third party as compliant with the HIPAA Privacy Rule?" https://www.hhs.gov/hipaa/for-professionals/faq/237/can-business-associates-self-certify/index.html
S6 >S6 US Department of Health and Human Services, Office for Civil Rights, "Business Associates" guidance, content last reviewed 21 October 2026. https://www.hhs.gov/hipaa/for-professionals/privacy/guidance/business-associates/index.html
S7 >S7 US Department of Health and Human Services, Office for Civil Rights, "Direct Liability of Business Associates" fact sheet, content last reviewed 16 July 2021. https://www.hhs.gov/hipaa/for-professionals/privacy/guidance/business-associates/factsheet/index.html
S8 >S8 US Department of Health and Human Services, Office for Civil Rights, "Guidance on HIPAA and Cloud Computing", content last reviewed 23 December 2022. https://www.hhs.gov/hipaa/for-professionals/special-topics/health-information-technology/cloud-computing/index.html
S9 >S9 US Department of Health and Human Services, Office for Civil Rights, "Guidance on Risk Analysis". https://www.hhs.gov/hipaa/for-professionals/security/guidance/guidance-risk-analysis/index.html
S10 S10 PCI Security Standards Council, FAQ, "Does PCI DSS apply to merchants who outsource all payment processing operations and never store, process or transmit cardholder data?" https://www.pcisecuritystandards.org/faqs/does-pci-dss-apply-to-merchants-who-outsource-all-payment-processing-operations-and-never-store-process-or-transmit-cardholder-data/
S11 S11 PCI Security Standards Council, FAQ, "Are merchants required to meet PCI DSS Requirement 12.9?" https://www.pcisecuritystandards.org/faq/articles/Frequently_Asked_Question/are-merchants-required-to-meet-pci-dss-requirement-12-9
S12 S12 PCI Security Standards Council, Information Supplement "Third-Party Security Assurance", March 2016 (written for an earlier version of the standard). https://listings.pcisecuritystandards.org/documents/ThirdPartySecurityAssurance_March2016_FINAL.pdf
S13 S13 PCI Security Standards Council, Information Supplement "Guidance for PCI DSS Scoping and Network Segmentation", version 1.1, May 2017. https://listings.pcisecuritystandards.org/documents/Guidance-PCI-DSS-Scoping-and-Segmentation_v1_1.pdf
S14 S14 International Organization for Standardization, "Certification". https://www.iso.org/certification.html
S15 S15 International Organization for Standardization, "ISO/IEC 27001:2022, Information security management systems". https://www.iso.org/standard/27001
S16 S16 Shared Assessments, "SIG: Third Party Risk Management Standard". https://sharedassessments.org/sig/
S17 S17 Cloud Security Alliance, "STAR Level 1: Security Questionnaire (CAIQ v4.1)". https://cloudsecurityalliance.org/artifacts/star-level-1-security-questionnaire-caiq-v4
S18 S18 US Federal Trade Commission, "FTC Safeguards Rule: What Your Business Needs to Know". https://www.ftc.gov/business-guidance/resources/ftc-safeguards-rule-what-your-business-needs-know
S19 S19 Caesars Entertainment, Inc., Form 8-K filed with the US Securities and Exchange Commission, September 2023 (event of 7 September 2023). https://www.sec.gov/Archives/edgar/data/1590895/000119312523235015/d537840d8k.htm
S20 S20 CISA and FBI, Joint Cybersecurity Advisory AA23-320A, Scattered Spider, 16 November 2023, revised 29 July 2025. https://www.cisa.gov/news-events/cybersecurity-advisories/aa23-320a
S21 S21 MGM Resorts International, Form 8-K filed with the US Securities and Exchange Commission, 5 October 2023. https://www.sec.gov/Archives/edgar/data/789570/000119312523251667/d461062d8k.htm
S22 S22 US Federal Trade Commission, "FTC Takes Action Against Marriott and Starwood Over Multiple Data Breaches", 9 October 2024. https://www.ftc.gov/news-events/news/press-releases/2024/10/ftc-takes-action-against-marriott-starwood-over-multiple-data-breaches
S23 S23 Verizon, 2026 Data Breach Investigations Report, report page. https://www.verizon.com/business/resources/reports/dbir/
S24 S24 IBM, Cost of a Data Breach Report 2026, report page. https://www.ibm.com/reports/data-breach
S25 S25 US Federal Trade Commission, "FTC Files Complaint Against Wyndham Hotels For Failure to Protect Consumers' Personal Information", 26 June 2012 (allegations, not findings). https://www.ftc.gov/news-events/news/press-releases/2012/06/ftc-files-complaint-against-wyndham-hotels-failure-protect-consumers-personal-information
S26 S26 US Federal Trade Commission, "Wyndham Settles FTC Charges It Unfairly Placed Consumers' Payment Card Information At Risk", 9 December 2015. https://www.ftc.gov/news-events/news/press-releases/2015/12/wyndham-settles-ftc-charges-it-unfairly-placed-consumers-payment-card-information-risk
S27 S27 Insurance Journal, "MGM Resorts to Pay $45M to Settle Data Breach Lawsuit", 29 January 2025. https://www.insurancejournal.com/news/national/2025/01/29/810011.htm
S28 S28 The PHP Group, php.net news archive, "Security advisory: Critical remote code execution vulnerabilities, update now", 21 October 2026. https://www.php.net/archive/2026.php#2026-10-21-5
S29 S29 The PHP Group, php-src security advisory GHSA-w8hw-w9fx-j79w, "RCE via OOM corruption" (CVE-2026-109618), published 21 October 2026. https://github.com/php/php-src/security/advisories/GHSA-w8hw-w9fx-j79w
S30 S30 The PHP Group, php-src security advisory GHSA-6wmf-4rv5-p3rp, "RCE via stale global state" (CVE-2026-109619), published 21 October 2026. https://github.com/php/php-src/security/advisories/GHSA-6wmf-4rv5-p3rp
S31 S31 CISA, Alert "Critical Vulnerabilities in PHP Require Immediate Patching", 21 October 2026, last revised 26 October 2026 (read from the archived copy of 3 November 2026; the KEV dates match CISA's own catalog feed). https://www.cisa.gov/news-events/alerts/2026/10/21/critical-vulnerabilities-php-require-immediate-patching
S32 S32 WordPress.org, "Requirements". https://wordpress.org/about/requirements/
S33 S33 Stripe documentation, "Integration security guide". https://docs.stripe.com/security/guide
S34 S34 Stripe documentation, "Stripe Checkout". https://docs.stripe.com/payments/checkout
S35 S35 Omega Revenue Engine (our own site), "Don't come to us with a preconceived idea". https://omegarevenueengine.com/dont-come-to-us-with-a-preconceived-idea/
S36 S36 Omega Revenue Engine (our own site), "One hundred questions a rented stack cannot answer". https://omegarevenueengine.com/company/100-questions/
S37 S37 US Securities and Exchange Commission, press release 2023-139, "SEC Adopts Rules on Cybersecurity Risk Management, Strategy, Governance, and Incident Disclosure by Public Companies", 26 July 2023 (read from the archived copy). https://www.sec.gov/newsroom/press-releases/2023-139
S38 S38 NIST, Special Publication 800-207, "Zero Trust Architecture", August 2020, abstract. https://csrc.nist.gov/pubs/sp/800/207/final
S39 S39 NIST, Special Publication 800-82 Revision 3, "Guide to Operational Technology (OT) Security", September 2023, abstract. https://csrc.nist.gov/pubs/sp/800/82/r3/final
S40 S40 CISA, Binding Operational Directive 22-01, "Reducing the Significant Risk of Known Exploited Vulnerabilities", 3 November 2021 (read from the archived copy). https://www.cisa.gov/news-events/directives/bod-22-01-reducing-significant-risk-known-exploited-vulnerabilities
S41 S41 Regulation (EU) 2016/679 (GDPR), Article 28, "Processor", text as reproduced at gdpr-info.eu. https://gdpr-info.eu/art-28-gdpr/
S42 S42 Regulation (EU) 2016/679 (GDPR), Article 32, "Security of processing", text as reproduced at gdpr-info.eu. https://gdpr-info.eu/art-32-gdpr/
S43 S43 Regulation (EU) 2016/679 (GDPR), Article 83, "General conditions for imposing administrative fines", text as reproduced at gdpr-info.eu. https://gdpr-info.eu/art-83-gdpr/
S44 S44 Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 50, "Transparency obligations for providers and deployers of certain AI systems", text as reproduced at artificialintelligenceact.eu. https://artificialintelligenceact.eu/article/50/
S45 S45 US Department of Health and Human Services, Office for Civil Rights, "Change Healthcare Cybersecurity Incident Frequently Asked Questions" (read from the archived copy). https://www.hhs.gov/hipaa/for-professionals/special-topics/change-healthcare-cybersecurity-incident-frequently-asked-questions/index.html
S46 S46 US Federal Communications Commission, news release "FCC Makes AI-Generated Voices in Robocalls Illegal", 8 February 2024. https://docs.fcc.gov/public/attachments/DOC-400393A1.pdf
S47 S47 US Department of Education, Protecting Student Privacy, "What is FERPA?". https://studentprivacy.ed.gov/faq/what-ferpa
S48 S48 CISA, "Secure by Design" (read from the archived copy). https://www.cisa.gov/securebydesign
S49 S49 RH-ISAC, "Retail and Hospitality Industry Insights: 2024 Verizon Data Breach Investigations Report Analysis" (Accommodation and Food Services figures from the 2024 DBIR). https://rhisac.org/wp-content/uploads/Verizon-DBIR-2024_Clear.pdf
S50 S50 IBM X-Force, "AI-powered adversaries and the enterprise risk challenge", 29 July 2026, on the Cost of a Data Breach Report 2026. https://www.ibm.com/think/x-force/2026-cost-of-a-data-breach-ai-adversaries-enterprise-risk
S51 S51 Visa, "What To Do If Compromised Document Revised", Visa Business News, 10 October 2019. https://www.visa.com.tt/content/dam/VCOM/global/support-legal/documents/what-to-do-if.pdf
S52 S52 Marsh, Global Insurance Market Index, third quarter 2026, release of 21 October 2026. https://www.marsh.com/en/about/media/global-commercial-insurance-falls-5-percent-q3-2026.html
S53 S53 UK Information Commissioner's Office, Data protection fining guidance, "The maximum amount of a fine under UK GDPR and DPA 2018". https://ico.org.uk/about-the-ico/our-information/policies-and-procedures/data-protection-fining-guidance/statutory-background/the-maximum-amount-of-a-fine-under-uk-gdpr-and-dpa-2018/
S54 S54 Office of the Australian Information Commissioner, Guide to privacy regulatory action, Chapter 7, civil penalties for serious or repeated interference with privacy. https://www.oaic.gov.au/about-the-OAIC/our-regulatory-approach/guide-to-privacy-regulatory-action/chapter-7-privacy-assessments
S55 S55 Help Net Security, "Omni Hotels suffer prolonged IT outage due to cyberattack", 4 April 2024, updated 15 April 2024. https://www.helpnetsecurity.com/2024/04/04/omni-hotels-suffer-prolonged-it-outage-due-to-cyberattack/