Most AI policies govern the AI an organisation chose. That is the least dangerous AI it has. The AI that should worry a board is the AI its suppliers chose: the managed service provider reading its tenant, the consultancy drafting from its documents, the SaaS vendor that switched on an assistant in last month's release. None of it went through your approval process, because none of it was yours to approve.
Every supplier with access to your systems is under the same pressure you are to do more with AI, and the cheapest place to start is the work it already does for you.
Our position is that a board should be able to answer five questions about the AI in its supply chain, and that every one of them can be answered with engineering rather than reassurance. If a supplier cannot answer them, that is a finding. If you cannot answer them, that is a gap in your governance, whatever your own AI policy says. We run a managed service of exactly this kind, so we have an interest in the answer; the questions work whoever you buy from.
Why supplier assurance does not already cover this
Supplier due diligence was built for a world where a supplier's staff did the processing. It asks about encryption, certification, penetration testing and vetting. Those still matter. They do not tell you what an AI model running under your supplier's account did with your data last Tuesday.
Three things changed. First, the processing is now done partly by models the supplier does not own, run by providers you never contracted with, sometimes in regions neither of you chose. Second, the actor has changed: an agent with a consultant's access can read your helpdesk, finance system and cloud estate in the time a person takes to open one, and the logs on your side will often show the consultant's name, not the agent's. Third, the work product has changed. A report drafted by a model and lightly reviewed looks exactly like a report written by a person, and you are paying for the second.
A processor contract signed in 2023 almost certainly says nothing about any of this. It has no clause for an agent deciding what to read next.
When this goes wrong, it lands on people. The consultant whose name is on an action an agent took. The customer whose record moved to a model provider nobody on your side had heard of. The finance director who signed a board pack built on a figure no one can trace. The board that has to explain to a regulator what happened and finds the only record belongs to the supplier.
The question is not whether your suppliers use AI. They do, or they will by your next renewal. The question is whether anyone could reconstruct what that AI did with your data if you needed to.
The five questions a board should be able to answer
Each of these can be asked of every supplier that holds or reaches your data. Each has a good answer that sounds specific and a bad answer that sounds reassuring.
1. Which suppliers hold or reach our data, and through what? Not the contract list: the access list. Which identities belong to suppliers in your tenant, what roles they hold, and which of your systems their tools connect to. A good answer is a list you can check in your own admin screens. A bad answer is "our suppliers are all under NDA". Ask for: the list of identities, roles and integrations, which you then check against your own tenant.
2. What AI do they run over it, and where is it processed? Which models, from which providers, called from which region, and whether your content is sent as it is or with personal details removed first. A good answer names the provider, the hosting route and the processing location, and says plainly where the location depends on the deployment type. A bad answer is "your data never leaves the UK" from a supplier using a global model endpoint. Ask for: the model providers on the sub-processor list, with storage location, processing location and transfer basis for each.
3. What is that AI allowed to do: read, draft or change? Reading your systems, drafting something a person reviews, and changing a record are three different risk levels, and many suppliers have not separated them. A good answer says which of the three their AI does today, and for anything that changes your systems, who on your side approves it before it happens. A bad answer is "a human is always in the loop", with no record of which human or when. Ask for: the approval record for the last change the supplier's AI made, or a written statement that it makes none.
4. What evidence would we hold if it went wrong? If a record is changed, a figure is wrong in a board pack, or a regulator asks how a decision was reached, what record exists, who holds it, and can it be altered after the fact? A good answer is a record that names the actor, person or agent, for every action, and that you can verify without relying on the supplier's word. A bad answer is "we keep comprehensive logs", which are the supplier's logs, on the supplier's systems, editable by the supplier.
5. Can we leave with our data and the evidence intact? When the contract ends, do you get your material back, does the supplier's access actually end, and do you keep the record of what was done in your name? A good answer is an ordered offboarding with a read-back of every removed grant. A bad answer is a deletion certificate and nothing else, because a supplier that deletes the evidence on exit has deleted your ability to answer question four. Ask for: the offboarding sequence, and what you keep afterwards.
What good looks like in engineering terms
These questions have engineering answers, and we build them into our own managed service, which reaches customer systems through per-customer connectors. Here is the shape of a supplier that can answer all five, stated as properties you can test rather than promises you have to accept.
A binding per customer, not a key per supplier. The supplier's agents reach your systems only through an identity created for your organisation alone, holding grants you made and can see and revoke in your own portal without asking anyone. A single service account that serves every customer, holding a credential that could read any of them, fails this test however good the policies around it are.
Reads that are labelled, and content treated as data. A ticket in your helpdesk can contain text aimed at an AI. A governed supplier labels everything its agents read from your systems as customer content and designs the system around the model so that text cannot become the authority for an action. The strongest designs go further: when a read looks hostile, they open a taint window in which nothing proposed for your systems proceeds without a human check. Ask how that is enforced. "The prompt tells the model to ignore instructions" is not an answer.
Changes only behind your approval. If the supplier's AI ever changes your systems, it should do so only inside a change request your own approver authorised, with a rollback path where the change can be undone, and a record that ties the change to the approval. Until a supplier has that, its AI should not be changing your systems at all, and a governed supplier will put that boundary in writing.
A signed record you can verify without the supplier. Every action, read or write, written to an append-only, hash-chained ledger that names the actor and is signed as it is written. The test is simple: can the supplier hand you an export that you can verify from the first entry using only a public key, without their software and without their help? If verification requires trusting the supplier, it is a log, not evidence. Be precise about what this proves: a hash chain shows that nothing was changed after it was written, not that everything was written. Ask what would make an omission visible to you, such as signed chain heads delivered to you as they are written, or a record on your side to reconcile against.
Residency stated per hop, never rounded. Where data is stored and where it is processed are different questions, and every model call is a hop. On the major cloud model services, the region a resource sits in and the place a given call is processed can differ, depending on the deployment type the supplier chose. A governed supplier tells you each hop, says which parts are fixed and which depend on a deployment type, names the transfer mechanism for any hop outside the UK, and tells you before any of it changes. Tokenising personal details before a model call is good practice, and it is worth asking about; it is pseudonymisation, so the content usually remains personal data in law and the transfer rules still apply.
Disclosure when AI did the work. In the contract, on the deliverable, and in the record. You should never have to discover that AI was involved in work you paid for, and if you ask an agent whether it is an AI, it should say so.
A written rule set. Somewhere, in writing, the supplier says what its AI may and may not do with customer data: what it may read, where it may send content, what it may store and where, what it may never do. A rule set the supplier cannot show you is a rule set you cannot hold them to.
Each of these is a design decision that is cheap at the start and expensive to retrofit. The suppliers who made them early can answer you. The rest will send you a policy document.
What to do this quarter
None of this needs a purchase. It is work a CIO, CISO or head of procurement can start with what they have.
Build the access list. List every supplier identity in your tenant, every third-party application with consent to your data, and every connector or integration a supplier runs into your systems. Set it beside your contract list. The gaps between the two are your first findings.
Send the five questions in writing. Start with the suppliers that reach your most sensitive systems. Ask for written answers, not a call. A supplier that can only answer verbally is telling you something about its evidence. We have published a questionnaire you can send as it is.
Ask for one real record. Pick an action a supplier's AI took in your systems in the last month and ask for the record of it: who or what took it, under which identity, what it read, and what it produced. How long the answer takes will tell you more than any questionnaire.
Read the sub-processor list for model providers. Your suppliers' sub-processor notices are where AI usually enters your supply chain on paper. Look for model providers, the regions they process in, and whether you were given the chance to object.
Run one exit drill. With your service owner's agreement, pick a low-risk supplier grant outside any critical window, revoke it from your own admin screens, and see what breaks and how long the supplier takes to notice. If you cannot revoke a supplier's access yourself, you have found something more urgent than AI.
Put the questions into renewals. Every contract that comes up for renewal is a chance to add terms on AI use, drafted with your counsel: disclosure, the processing locations and transfer basis, notice before a model provider changes, approval for changes to your systems, access to the record, and the evidence you keep on exit.
Where this is going
Regulation is moving the same way. As of November 2026, the EU AI Act's transparency duties apply from August 2026 where the Act reaches an organisation or its output, and AI literacy has applied to providers and deployers since February 2025. UK GDPR already makes you answerable for what your processors do with personal data. None of these lets an organisation say that the AI was its supplier's and therefore not its problem. The direction is towards organisations being expected to know, and to show, what AI did with their data, wherever in the chain it ran.
The organisations that do well here will not be the ones with the longest AI policy. They will be the ones that can produce, on the day it is asked for, a record of what their suppliers' AI did, signed by something other than the supplier's word. That is also what lets them give their suppliers more to do: a supplier whose AI you can see is one you can trust with more of the work.
What we will show you
ORCA's managed service is built to answer all five questions with something you can inspect. Ask us for any of these:
- Per-customer federated bindings. Our agents reach only what your organisation bound, through an identity created for you alone, with the tools fixed in a manifest you approve by digest. You see every grant in your own portal and can remove it without asking us.
- Labelled tool results and taint windows. Everything read from your systems is labelled as your content, and a read that looks hostile opens a window in which anything proposed for your systems is raised to a human check.
- A signed evidence ledger you verify yourself. Every read and action is chained, signed and attributed to a person or a named agent. The export verifies from the first entry with our public key alone; today we produce it on request, and signed heads delivered to you as they are written are planned.
- Governed writes that wait for your approver. Planned. Until it ships, our managed service reads and does not write.
- UK data at rest. Stored in Microsoft Azure UK South in a partition of its own, with personal details tokenised before any model call and processing locations set out in your terms.
- Disclosure when AI did the work, on the deliverable and in the record, with a named director accountable for everything we deliver.
- A written rule set saying what our AI may and may not do with your data, and which rules are enforced in code.
- Exit with data and evidence intact: export first, credentials purged, a final signed record.
The standard behind each is in What a governed managed service owes its customers. We also help organisations put the same questions to every supplier: mapping who reaches which data, running the questionnaire, and setting the contract terms. Start with the supplier you are least sure about. Talk to us; we will show you ours first.