Under DORA, a financial entity that relies on a foundation model provider must manage it like any other ICT third-party service provider. That means recording it in the register of information, assessing concentration and substitutability before contracting, including the Article 30 contract terms, and keeping a tested exit plan. The entity stays fully responsible. No model provider appears on the EU's or the UK's critical provider lists so far, so supervisors will judge your own controls.
What DORA requires of you
Regulation (EU) 2022/2554, the Digital Operational Resilience Act, has applied since 17 January 2025 under Article 64. Article 28(1)(a) says financial entities remain fully responsible for meeting their obligations when they use ICT third-party service providers. Nothing in the text exempts a provider because its service is new or because it is an AI model.
Article 28(2) requires an ICT third-party risk strategy, including a policy for services that support critical or important functions and, where applicable, a multi-vendor strategy. Article 28(4) requires a pre-contract assessment that covers concentration risk under Article 29 and due diligence. Article 28(6) calls for risk-based audit planning, and Article 28(7) requires termination rights.
For a model provider, the first decision is classification. If a model supports a critical or important function, such as credit decisioning, fraud screening or customer communications that your business cannot pause, the stricter rules apply. If it supports only an internal drafting tool, the obligations are lighter. Record the reasoning either way, because a supervisor will ask why you classified it as you did.
The register of information and the model layer
Article 28(3) requires a register of information on all contractual arrangements with ICT third-party providers, distinguishing those that support critical or important functions. Entities report it to their competent authority at least yearly, and they must inform the authority of planned arrangements that support critical or important functions.
The register is where AI dependencies tend to go missing. Consider a hypothetical insurer whose claims system is a SaaS product. The SaaS vendor calls a foundation model through an API, and that model runs on a hyperscaler's infrastructure. The insurer has one contract, but its service depends on three providers. Article 29(2) asks entities to consider subcontracting chains, third-country providers, insolvency law and data recovery, so the chain is the unit to map, not just the contract you signed.
For each AI service, capture the provider, the model or model family, where it is hosted, which functions rely on it, and which subcontractors sit beneath it. Ask your vendors to tell you when the underlying model changes. A silent model swap changes the behavior your controls were tested against.
Contract terms that matter for model providers
Article 30(2) lists terms every contract must carry, including data locations, return of data on insolvency or termination, and service levels. Article 30(3) adds terms for services that support critical or important functions: full service level descriptions with quantitative targets, notice and reporting obligations, business contingency testing, participation in threat-led penetration testing, access, inspection and audit rights, and exit strategies with a mandatory, adequate transition period.
Standard terms from model providers may not meet these. Check each item against the actual contract:
- Data location and return: can you say where prompts, outputs and any fine-tuning data are stored, and get them back in a usable form?
- Service levels: are availability and latency targets quantitative, and do they cover the model endpoint, not only the platform?
- Audit and inspection: does the provider accept audit rights, or will it offer only its own certifications?
- Change notice: will the provider tell you before it retires or replaces a model version?
- Transition period: does the contract give you time to move, and on what terms?
Keep the evidence of each negotiation, because Article 28(6) expects audit planning to follow risk, and a provider that refuses audit rights should draw more of your attention, not less.
If a provider will not accept a term that Article 30 requires for a critical or important function, that is a finding for your risk committee, not a negotiation footnote.
Concentration risk and substitutability
Article 29(1) tells entities to consider whether a provider is not easily substitutable, and whether they have multiple arrangements with the same or closely connected providers. Both questions apply directly to models. Substitutability depends on more than switching an API endpoint. Prompts, evaluation suites, fine-tuned weights, retrieval indexes and downstream approvals are all built around one model's behavior. That is analysis, not a regulatory finding, but it is where the effort sits.
Supervisors are watching this. In a letter to the UK Treasury Committee dated 1 April 2026, the Bank of England's Sarah Breeden said the Bank and FCA AI Consortium, set up in May 2025, work on concentration risks including those from third-party model providers. A multi-vendor strategy under Article 28(2) is the response DORA anticipates, but it carries costs. Two models mean two sets of evaluations and two sets of contracts.
A workable test is to run your main use cases on a second provider's model, on a sample, and measure the gap. The result tells you whether substitution is a realistic exit or only a line in a policy.
Critical third-party regimes in the EU and UK
Article 31 of DORA lets the European Supervisory Authorities designate critical ICT third-party providers, known as CTPPs, for direct oversight. On 18 November 2025 the ESAs published the first list. Their press release says designation followed the registers of information and a criticality assessment covering systemic importance, role in supporting critical or important functions, and substitutability.
The list names 19 providers, including Amazon Web Services EMEA, Google Cloud EMEA, Microsoft Ireland Operations, IBM, Oracle Nederland, SAP, Accenture and Bloomberg. No foundation model provider is on it. Check the ESAs' current list, because it may have changed since November 2025.
The UK took a similar route. The Financial Services and Markets Act 2023 inserted section 312L into the 2000 Act, and the Critical Third Parties (Designation) Regulations 2026 (SI 2026/777), made on 8 July 2026 and in force from 13 July 2026, designate four providers: Amazon Web Services EMEA SARL, Google Cloud EMEA Limited, Microsoft Ireland Operations Limited and Oracle Corporation UK Limited. HM Treasury's 10 July 2026 announcement says firms remain responsible for managing their third-party risk. The Bank of England's CTP page describes joint oversight by the Bank, the PRA and the FCA, under rules published in November 2024.
The Bank's April letter recorded the Treasury Committee's recommendation that HM Treasury designate the major AI and cloud providers by the end of 2026. It said a provider of AI services could be designated if it were judged to pose financial stability risks. Treat that as a signal, not a forecast. Designation oversees the provider's systemic services. It does not take away your own duties.
Exit plans that can be tested
Article 28(8) requires exit strategies for critical or important functions. Plans must be comprehensive, documented, tested sufficiently and reviewed periodically. They must identify alternative solutions and transition plans, so that the entity can exit without business disruption, without limiting regulatory compliance and without harming service continuity to clients.
UK firms face parallel expectations. The PRA's Supervisory Statement SS2/21 on outsourcing and third party risk management has a version published on 18 March 2026, effective 18 March 2027, following Policy Statement PS7/26. Its Chapter 10 says exit plans should cover stressed exits and be documented and tested as far as possible. The PRA does not prescribe a form of exit. It focuses on the outcome: continued delivery of important business services. Viable options it names include bringing the service back in-house or moving it to an alternative provider.
SS2/21 also asks firms to assess the ability to substitute a provider, including estimated cost, operational impact, risk and timeframe in both stressed and non-stressed scenarios. It sets out an annual register of material third party arrangements, filed through FCA RegData with a 90-day window, and expects the register to support concentration risk analysis. For a model provider, a stressed exit might be a sudden withdrawal of a model version or the provider's insolvency. Write the plan for both. Name who triggers it, which fallback model or in-house option takes over, what data must move, and how long the business can run in a degraded mode.
Where to start
Start with facts you can gather in a few weeks:
- List every AI service in use, including those embedded in SaaS products, and map each to its provider, model, hosting and subcontractors.
- Classify each against your critical or important functions, and record the reasoning.
- Update the register of information so each model dependency appears, with the supply chain beneath it.
- Compare each contract against Article 30(2) and 30(3), and log the gaps for the risk committee.
- For each critical use, test a second model on a sample of real tasks and record the quality gap, the effort and the time to switch.
- Write a stressed exit plan covering model withdrawal and provider failure, then run a tabletop exercise against it.
Work through these steps by function, not by vendor, because the exit question is about whether the business service survives. Where you cannot answer it, you have found the work.
References
- European Union: Regulation (EU) 2022/2554 (DORA), EUR-Lex
- ESMA, EBA and EIOPA: European Supervisory Authorities designate critical ICT third-party providers
- European Supervisory Authorities: List of designated critical ICT third-party providers
- HM Treasury: UK financial system strengthened with new safeguards for major technology providers
- UK Government: The Critical Third Parties (Designation) Regulations 2026, SI 2026/777
- Bank of England: Letter to the Treasury Committee on AI in financial services, 1 April 2026
- Bank of England: Critical third parties
- PRA: SS2/21 Outsourcing and third party risk management (March 2026 update)





