CAIRNCYBER ADVISORY, homeContact

Insights

Your model provider is an ICT third party. Most registers don't say so.

· 6 min read · Gary Johnston

An LLM behind a business process is an ICT third-party dependency under DORA and the UK outsourcing rules. It usually arrives without a procurement event, so it is usually missing from the register, the contract review and the incident process.

A firm adds a summarisation step to client reporting, or an assistant that drafts responses to operational queries. It works, the pilot becomes the process, and a year later a model API sits in the path of something the business depends on.

Ask where that dependency is recorded and the answer is usually nowhere in particular. It didn’t arrive through procurement. It arrived as a feature in a SaaS product already under contract, as a line on a cloud bill, or as a pilot that never had a go-live decision. So the third-party framework the firm has spent years building has never looked at it.

Which rules apply

For EU financial entities, DORA has applied since 17 January 2025. Plenty of UK groups are in scope through an EU management company or subsidiary. For UK-regulated firms, the same questions come from the PRA’s outsourcing and third-party risk management expectations (SS2/21) and the operational resilience rules for important business services. The vocabulary differs, “critical or important function” against “important business service” and “material outsourcing”, but the work is the same. Find the dependency, decide how much it matters, contract for that, and be able to survive losing it.

Find it: model providers arrive by three routes

  1. Direct. You hold an API contract with a model developer. It is visible, but often signed on click-through terms by a team with a credit card.
  2. Through a cloud platform. The model is consumed through your hyperscaler’s AI service. Your contract is with the cloud provider, and the model is one more service on an existing agreement, so nobody reassesses it.
  3. Embedded in a supplier’s product. Your SaaS vendor calls a model provider on your behalf. The vendor is your ICT third party, and the model provider is their subcontractor, processing your data.

The third route is the one registers miss most, and it is the one DORA’s register of information cares about. For services that support critical or important functions, the register has to record the subcontracting chain, not just the supplier you pay. The question for every existing supplier is short: do you use a model provider to deliver this service to us, which one, where is it hosted, and what do you send it?

Classify the function honestly

The model itself isn’t critical or important. The function it supports might be. There is a strong pull towards classifying anything AI as a productivity tool, because that keeps it light-touch. The honest test is operational: if the model stopped tomorrow, or started answering differently, which process would degrade, and would that breach an impact tolerance?

Revisit the answer as usage grows. The classification made at pilot stage is almost never right a year into production.

Contract for how models actually change

For services supporting critical or important functions, DORA Article 30 expects, among other things, service levels with precise quantitative and qualitative performance targets, notice periods and reporting obligations, rights of access, inspection and audit, and exit strategies with a mandatory transition period.

Compare that with the standard terms of most model providers. Versions are deprecated on the provider’s timetable. Behaviour can change without a version change. Availability is committed to, if at all; output quality is not. That gap won’t close by itself, so write it down and work it from both ends:

  • Negotiate where you have leverage: minimum notice before deprecation, a pinned version for the life of a release, notification of material changes to safety layers, and clarity on whether your data is retained or used for training.
  • Compensate where you don’t: an evaluation suite run against the production configuration on a schedule, so that a change in behaviour is detected by you rather than by a client. Record the residual risk formally, as a risk acceptance, not a footnote.

Make “up but wrong” an incident

Most incident processes recognise an outage. Few recognise a model that is available and returning worse answers than it was last week, or a prompt that sent data somewhere it should not have gone. Both are ICT incidents. Whether either reaches the threshold for a major incident report depends on its impact, and that assessment can only happen if the event is recognised in the first place. One line in the incident taxonomy and one runbook entry covers most of it.

Concentration, and what designation does not do

Concentration risk looks different here. The model provider and the cloud provider are often the same company, or the model runs on the cloud platform the rest of the estate already depends on. An exit plan that moves the model to another service on the same platform removes less risk than it appears to. The practical controls for model change and exit are covered in Evidence, not attestation.

Supervisors are moving towards the providers directly. The European Supervisory Authorities published their first list of critical ICT third-party providers under DORA in November 2025, and the major cloud providers are on it. In the UK, HM Treasury designated AWS, Google Cloud, Microsoft and Oracle as critical third parties with effect from 13 July 2026. The Treasury Committee had already called for the major AI and cloud providers to be designated by the end of 2026.

None of that moves responsibility off the firm. DORA says financial entities remain fully responsible for their obligations whatever they outsource. The Bank of England said the same when announcing the UK designations: the regime complements, and does not replace, firms’ own responsibility for due diligence, risk management and contingency planning. Oversight of your provider is not a control you operate.

What to have before a supervisor asks

  • Every model dependency in the register, including the ones inside suppliers’ products, with the chain recorded.
  • A classification rationale for each, dated and revisited since the pilot.
  • A contract gap list against Article 30 or your outsourcing policy, with each gap either negotiated or accepted as a risk.
  • An incident taxonomy that recognises “up but wrong”.
  • An exit plan that names the alternative model and the date it was last tested.

None of these needs a new framework. They need the framework you already run to be pointed at a dependency that arrived through the side door.

Written by Gary Johnston, Principal Security Architect at Cairn Cyber Advisory Ltd. CISSP, CCSP. Nothing here describes any client's environment.

More insights

Start a conversation

Send me an outline of the problem and your timescale. I normally reply within one working day and will tell you directly whether I think I can help.