The North Sea at night under a low sky, one distant light on the horizon

Field notes Method

A supplier due-diligence questionnaire for AI in your supply chain

November 2026  ·  6 min read

Most supplier questionnaires were written before suppliers ran models over their customers' data, and it shows. They ask whether data is encrypted at rest and whether staff are vetted. They do not ask which AI provider receives your content, or whether the supplier's agent can change a record in your systems without anyone on your side agreeing to it.

The questions below close that gap. They are written to be sent as they are, to any supplier that holds or reaches your data: managed service providers, consultancies, outsourcers and SaaS vendors. Each group says which existing obligation it maps to, so the answers can sit inside your current supplier assurance rather than beside it. The answers protect the people whose data it is: your customers and your staff.

This is a working questionnaire, not legal advice; your counsel decides what your contracts require. And the mapping to UK GDPR Article 28 assumes the supplier processes personal data on your behalf. Where it does not, the questions still tell you what you need to know; the legal hook is different.

1. Access: what reaches our data

Maps to: UK GDPR Article 28(3)(a), processing only on documented instructions; Article 32, security of processing.

  1. List every identity, application and integration through which you, or software you run, can reach our systems or data. For each, state the roles or permissions it holds.
  2. Is any credential or identity used to reach our systems shared with any other customer? If yes, describe what stops it reaching another customer's data, and what stops another customer's work reaching ours.
  3. Can we see each of these grants in our own administration screens, and revoke them ourselves without contacting you? Name any we cannot.
  4. Where do you hold any credential of ours, and who and what can read it?

2. Models: what AI runs over our data, and where

Maps to: Article 28(2) and 28(4), sub-processors; UK GDPR Chapter V, international transfers.

  1. Which AI models process our data, from which providers, through which hosting service? Include models used inside your own tools and by your staff.
  2. For each, where is our data stored, and where is the model processing performed? If the processing location depends on a deployment type or routing choice, say which applies to our data today.
  3. Is each model provider on your sub-processor list, and were we given notice and the chance to object when it was added? Will you tell us before a model provider or processing location changes?
  4. For any hop outside the UK, what is the transfer mechanism (an adequacy decision, the International Data Transfer Agreement or the Addendum to the EU clauses), and where is your transfer risk assessment?
  5. Is personal data removed, masked or tokenised before our content reaches a model? Which flows, if any, send it untokenised?
  6. Is our data used to train or fine-tune any model, by you or by any provider you use? Cite the term that says so.
  7. Do your staff use AI tools outside your governed route, such as consumer chatbots or personal accounts, on our work? What prevents it?

3. Actions: what the AI may do

Maps to: Article 28(3)(a), documented instructions; Article 32.

  1. For each AI system that touches our data, does it read only, draft for a person to review, or change our systems or records?
  2. If any AI can change our systems, who on our side approves each change before it happens, and how is that approval recorded?
  3. How do you stop content read from our systems, such as text in a ticket or document, being treated by your AI as an instruction?
  4. Do you have a written rule set describing what your AI may and may not do with customer data? Please share it, or the parts that apply to us.
  5. Which named person at your organisation is accountable for what your AI does in our systems, and who reviews AI-prepared work before it reaches us?

4. Evidence: what we would hold if it went wrong

Maps to: Article 28(3)(h), information to demonstrate compliance and allow audits; Article 33(2), the processor's duty to notify the controller of a breach.

  1. Is every action your AI takes in our systems recorded? Does the record name the actor, distinguishing an agent from a person?
  2. Can the record be altered or deleted after it is written, by you or anyone else? How would we know?
  3. Can you give us an export of the record that we can verify independently, without your software or your help? What do we need to verify it?
  4. If we ask "who did this?" about an action in our systems, how quickly will you answer, and with what reference?
  5. If your AI produces something wrong that reaches us, how and when will you tell us?

5. Transparency: when AI did the work

Maps to: EU AI Act Article 50, transparency, where the Act reaches the supplier or its output; Article 4, AI literacy.

  1. Does our contract with you state that you use AI in delivering to us, and which tools touch our data?
  2. Do your deliverables say when they were prepared with AI assistance, and who reviewed them?
  3. If our staff interact with an AI system you operate, does it tell them it is an AI?
  4. Where you provide a generative AI system that reaches us, what is your position on marking its output, and by when?
  5. How do you train your people who use AI on our work?

The Article 50 duties fall mainly on providers of AI systems, with some on deployers, and only where the Act reaches the organisation or its output is used in the EU. Whether that includes your supplier depends on its facts. Ask the question anyway: a supplier that has not worked out its own position has not worked out yours. We set out why the marking date matters in 2 December is a marking deadline.

6. Exit: leaving with data and evidence intact

Maps to: Article 28(3)(g), deletion or return at the end of the service.

  1. At the end of the contract, what do we receive, in what format, and when?
  2. In what order is your access removed, who removes each part, and will you read back that each removal took effect?
  3. What do you keep after exit, for how long, and why?
  4. Do we keep the record of what you did in our name, in a form we can still verify after you have gone?

How to read the answers

Specific beats comprehensive. "We keep comprehensive audit logs" answers none of section 4. "Every action is written to a hash-chained ledger that names the agent or person, and we can give you an export that verifies with our public key" answers most of it, and you can test it.

Residency answers that are too clean. If a supplier says your data never leaves the UK while using a hosted model service, ask where the model processing happens and what deployment type governs it. Storage and processing are different questions.

Honest gaps are a good sign. A supplier that says "our AI does not change your systems today, and when it does, it will be behind your approver" is giving you a better answer than one that claims everything is in place. Ask which of their answers describe what runs today and which describe what is planned. A supplier that cannot tell the difference is the one to worry about.

Send it to us first. ORCA's managed service answers every section with something you can inspect: the grants our per-customer binding holds in your own portal and the manifest you approved; the label on every tool result and the taint window a hostile read opens; a ledger export you verify from the first entry with our public key alone (proof the chain is intact and signed by us; until signed heads reach you as written, which is planned, it cannot rule out re-signing); UK data at rest with the processing route for each hop in your terms; the AI-assistance line on our deliverables; the rules our AI works to; and the offboarding sequence. Writes behind your approver are planned; until then our service reads and does not write. Use it whoever you buy from. We also help run it across a supply chain and turn gaps into contract terms. Why it matters: Your suppliers are already running AI on your data.

All field notes

Tell us what you are accountable for.

The useful first conversation is usually about the thing that worries you, not about our capabilities. If it turns out we are the wrong firm for it, we will say so.

One reply from a person who has read what you sent.

Colophon

  • Plate 06, the sea at night: generated, Azure OpenAI gpt-image-1.5, 3 November 2026.

Every plate on this page is a generated stand-in for commissioned photography. The prompt that made each one is embedded in its file. How this site was made.