- dora
- ict third-party risk
- ai agents
- llm vendors
- banking
- operational resilience
- ai governance
- eu ai act
- stress testing
- regulatory compliance
DORA and AI Agents: Managing LLM Vendors as ICT Third-Party Risk
AI agents run on LLM vendors, clouds and orchestration tools, and DORA makes banks fully responsible for all of them. Here's how Articles 28–30 apply, what the 2026 ESA frontier AI statement says and a practical checklist.
By Aodit
13 min read
A bank's AI agent rarely runs on the bank's own model. A customer-service agent, a KYC triage assistant or an invoice-processing agent is usually built on a large language model (LLM) accessed through an API, hosted on a hyperscale cloud and wrapped in orchestration software from yet another vendor. Each of those layers is an ICT service. Under the EU's Digital Operational Resilience Act (DORA), each of them is also ICT third-party risk that the financial entity remains fully responsible for.
DORA has applied since 17 January 2025. What has changed in 2026 is the attention supervisors now pay to AI. On 31 July 2026, the EBA, EIOPA and ESMA published a joint statement on ICT risks from frontier AI models (JC 2026 25), and the ESAs have begun embedding AI-related risks into their oversight of critical ICT third-party providers. For banks deploying AI agents, the question is no longer whether DORA applies to LLM vendors. It is whether the bank can evidence that it understands, contracts for, tests and can exit those dependencies.
This guide explains how DORA's ICT third-party risk rules apply to AI agents and LLM vendors, where AI creates new gaps, and a practical checklist risk and procurement teams can use now.
Why AI agents are a DORA problem, not just an AI Act problem
Most AI governance discussion in European banking focuses on the EU AI Act. That makes sense for use cases such as creditworthiness assessment, which Annex III of the AI Act classifies as high-risk. But the AI Act and DORA ask different questions:
- The AI Act asks whether an AI system is safe, transparent, accurate and subject to human oversight, with obligations allocated between providers and deployers.
- DORA asks whether the financial entity can keep its critical or important functions running when an ICT dependency fails, changes or becomes compromised.
An AI agent sits squarely in DORA's scope because it is delivered through ICT services. DORA Article 28(1) is explicit: financial entities "shall manage ICT third-party risk as an integral component of ICT risk" and "remain fully responsible" for compliance. Outsourcing the model does not outsource the accountability.
Agents also differ from earlier AI in ways that matter for operational resilience:
- They act, not only predict. An agent that can open tickets, draft customer communications, update CRM records or trigger payments workflows has a larger operational blast radius than a scoring model.
- Their behavior depends on a remote, changing model. When an LLM vendor updates or deprecates a model version, the agent's behavior can change without any code change at the bank.
- They have deep supply chains. A single agent may depend on a model provider, its cloud host, a vector database, a retrieval pipeline and an orchestration framework, several of which subcontract further.
That combination of autonomy, silent change and layered dependency is exactly what DORA's third-party framework was designed to make visible.
What the 2026 ESA statement on frontier AI says
The ESAs' joint statement of 31 July 2026 is important context, but it should be read accurately. It focuses on frontier AI as a cyber threat: highly capable models that could "rapidly discover and exploit vulnerabilities", target shared infrastructure, and compress the time between vulnerability discovery and exploitation. It does not set out rules for AI agents that banks themselves deploy.
Its expectations are nonetheless directly relevant to any institution running AI agents:
- Governance: management bodies should be committed to mitigating frontier-AI-driven cyber risks, governance structures should be put in place without delay, and risk appetite frameworks should be reviewed to add relevant metrics and tolerance thresholds, including in relation to internal use of such models.
- Inventories: prevention starts with asset inventories that include AI/ML components.
- Testing: operational resilience testing should evolve toward AI-enhanced threat scenarios, with more frequent scanning and testing.
- Supply chain: cybersecurity standards should be assessed, monitored and enforced across the whole supply chain.
- Proportionality: measures should scale with the entity's size, risk profile and complexity, consistent with DORA Article 4.
On the oversight side, the ESAs, acting as Lead Overseers, say they engaged critical ICT third-party providers on AI-related risks, are embedding those risks into their oversight examination methodology, and expect them to be reflected in 2027 oversight examinations.
Our reading (opinion, not regulatory text): if supervisors expect AI/ML components in asset inventories and AI-aware resilience testing for threat purposes, it is a short step to expecting the same discipline for AI agents the bank itself operates. Institutions that already treat agents as inventoried, tested ICT assets will find those conversations easier.
How DORA's third-party rules map to AI agents and LLM vendors
1. The register of information must include AI dependencies
Article 28(3) requires financial entities to maintain a register of information covering all contractual arrangements on the use of ICT services, at entity, sub-consolidated and consolidated levels. The templates are set out in Commission Implementing Regulation (EU) 2024/2956.
For AI agents, the practical gaps are usually:
- LLM API subscriptions bought by business units outside central procurement.
- AI features embedded in existing SaaS (CRM, helpdesk, document management) that silently route data to a model provider.
- Orchestration and retrieval tooling treated as "developer tools" rather than ICT services supporting a business function.
If an agent supports a function that the institution has classified as critical or important, the underlying model and hosting arrangements need to be recorded as such, including the supply chain behind them.
2. Pre-contract assessment: "critical or important" is the key question
Before entering an arrangement, Article 28(4) requires the financial entity to assess whether it covers ICT services supporting a critical or important function, identify and assess all relevant risks, undertake due diligence on the prospective provider, and assess conflicts of interest.
For an AI agent, the classification question should be driven by what the agent can do, not by what the model is. The same LLM can sit behind an internal drafting assistant (probably not critical) and an agent that handles customer complaints or prepares payment instructions (quite possibly supporting a critical or important function). A good due diligence file therefore documents the agent's permissions and actions, not only the vendor's security certifications.
3. Contracts: Article 30 meets standard AI terms of service
Article 30(2) sets minimum contract content for all ICT services, including a clear description of services, the locations where data is processed (with advance notice of changes), data integrity and confidentiality provisions, access to and return of data on termination or insolvency, service level descriptions, incident assistance, cooperation with authorities and termination rights.
Where the service supports a critical or important function, Article 30(3) adds:
- precise quantitative and qualitative performance targets;
- notice periods and reporting obligations for developments that could materially affect service delivery;
- business contingency plans and security measures;
- participation in threat-led penetration testing (TLPT);
- unrestricted access, inspection and audit rights; and
- exit strategies with a mandatory adequate transition period.
The AI-specific friction points are predictable:
| DORA requirement | Why it is hard with LLM vendors | What to negotiate or evidence |
|---|---|---|
| Service descriptions and SLAs (Art. 30(2)(e), 30(3)(a)) | Uptime is measurable; output quality and behavior usually are not covered | Uptime/latency SLAs plus contractual notice of model changes; your own behavioral acceptance thresholds |
| Notice of material developments (Art. 30(3)(b)) | Model versions are updated, re-tuned or deprecated on vendor timelines | Minimum notice before deprecation; access to pinned versions; change logs |
| Data locations (Art. 30(2)(b)) | Inference and logging may occur across regions | Region pinning, data retention terms, no-training commitments where relevant |
| Audit and access rights (Art. 30(3)(e)) | Hyperscale AI providers resist individual audits | Pooled audits, certifications and reports, plus your own independent testing of the deployed agent |
| Exit strategies (Art. 28(8), 30(3)(f)) | Prompts, tools and evaluations are tuned to one model | Documented fallback model or manual process, and re-testing before switching |
4. Subcontracting chains
Commission Delegated Regulation (EU) 2025/532 specifies the elements a financial entity must determine and assess when ICT services supporting critical or important functions are subcontracted. For AI agents this matters because model providers commonly rely on cloud infrastructure from another provider. Understanding that chain is also part of assessing concentration risk (Article 29): if several of your agents, and your core cloud estate, depend on the same hyperscaler, a single outage can take out both your primary systems and the AI layer meant to help you respond.
5. Critical ICT third-party providers (CTPPs)
On 18 November 2025, the ESAs published the first list of 19 designated critical ICT third-party providers, including major cloud providers such as AWS, Google Cloud and Microsoft. Designated CTPPs are subject to direct EU oversight by a Lead Overseer. Two points matter for AI agent deployments:
- Many LLMs are delivered through, or hosted on, designated CTPPs. Direct oversight of the provider does not relieve the bank of its own Article 28 obligations.
- Standalone AI model providers were not among the publicly highlighted designations at the time of writing. Banks should not assume that a provider outside the CTPP list is low-risk. The absence of a designation says nothing about the provider's importance to your critical functions.
Where DORA leaves a gap: behavior
DORA was drafted around availability, integrity, confidentiality and recoverability of ICT services. It gives financial entities strong tools to ask: Is the service up? Is the data protected? Can we leave?
It is less direct about a question that AI agents make urgent: Does the service still behave the way we approved it to behave?
Consider three scenarios that do not register as classic ICT incidents:
- Silent model update. The vendor ships a new model version. Uptime is 100%. But the complaints-handling agent now concedes refunds it previously escalated, because the new model is more "helpful" under customer pressure.
- Prompt injection via retrieved content. An agent reading supplier invoices follows instructions embedded in a document and changes bank details in a draft payment file. No system was breached in the traditional sense.
- Degraded escalation judgment. After a retrieval pipeline change, a KYC triage agent stops flagging borderline cases for human review. Every component reports healthy.
Each of these is an operational resilience event for a critical or important function. Each is invisible to uptime SLAs, SOC 2 reports and infrastructure monitoring. DORA's general testing requirements (Articles 24 and 25 require a risk-based digital operational resilience testing programme, including scenario-based and performance testing) provide a legal hook, but they do not tell you how to test an agent's judgment.
This is where behavioral stress testing becomes the missing control. AODIT provides independent behavioral risk ratings for AI agents: it puts an agent through structured adversarial conversations and scores how it behaves under pressure. You can see what that looks like in a two-minute free live demo, with no account required.
A practical DORA checklist for AI agents
Use this as a starting point with your ICT risk, procurement, model risk and compliance teams. It reflects our recommendations, not a regulatory checklist.
Inventory and classification
- Every AI agent in production or pilot is recorded as an ICT asset, with its model provider, hosting provider and orchestration components.
- Each agent is mapped to the business function it supports, and classified against your critical or important function criteria based on its permissions and actions.
- LLM and AI-feature arrangements appear in the register of information, including relevant subcontractors.
Contracts
- Contracts for agents supporting critical or important functions are reviewed against Article 30(3).
- Model change and deprecation notice periods are agreed, or the gap is documented and risk-accepted.
- Data processing locations, retention and use of your data for training are explicit.
Testing and monitoring
- Each agent has a behavioral baseline: how it handles normal requests, grey-zone requests, manipulation attempts and escalation triggers.
- That baseline is re-tested after every model version change, prompt change, tool change or retrieval change.
- Testing includes adversarial scenarios such as prompt injection, authority spoofing and social engineering, consistent with the ESAs' call for AI-aware threat scenarios.
- Results are documented in a form your second line, internal audit and supervisors can review.
Exit and continuity
- A fallback exists for each critical agent: an alternative model, a degraded mode or a manual process.
- The fallback has been tested, including behavioral re-testing of the alternative model before it goes live.
- Concentration risk across agents and core cloud services has been assessed.
How AODIT supports DORA-aligned AI agent oversight
AODIT does not replace your DORA framework, your register of information, your contract reviews or regulatory threat-led penetration testing. What it adds is independent, repeatable evidence about agent behavior, which is the layer DORA's traditional third-party controls do not see.
The AODIT-6 methodology assesses agents across six dimensions: Reliability, Integrity, Confidentiality, Judgment, Resistance and Resilience. Its Resistance dimension covers jailbreaks, prompt injection, social engineering, authority spoofing and adversarial reframing. Its Judgment dimension covers escalation and blast-radius decisions. Each scenario follows an eight-turn protocol that moves from a baseline request through grey-zone requests, hard pushes and manipulation, and then checks whether the agent recovers.
For DORA purposes, that gives risk teams three useful things:
- A behavioral baseline to approve an agent against before it supports a critical or important function.
- Regression evidence when an LLM vendor changes or deprecates a model, so a model update becomes a controlled change rather than a silent one.
- Exit readiness: the same test suite can be run against a fallback model to show that switching providers will not introduce unacceptable behavior.
Because DORA is fundamentally about reducing dependency risk, it matters that AODIT itself does not add a new data exposure. AODIT supports on-premise and air-gapped deployment, so test data and agent interactions do not need to leave your environment. Results are advisory and are meant to inform, not replace, your institution's own risk decisions.
If you are preparing AI agent evidence for a DORA review, an AI Act conformity process or an internal audit, request an evaluation and we will walk through how a behavioral assessment would fit your framework.
How DORA and the AI Act fit together for AI agents
For many banks, the same AI agent will fall under both regimes:
- DORA applies to the ICT services and dependencies behind the agent, regardless of the use case.
- The AI Act applies additional obligations if the agent is, or is part of, a high-risk AI system, for example in creditworthiness assessment. Under the Digital Omnibus amendments in Regulation (EU) 2026/1744, obligations for Annex III high-risk systems apply from 2 December 2027. Deployers of high-risk systems must, among other things, monitor operation, assign human oversight and keep automatically generated logs for at least six months. See our overview of EU AI Act obligations for AI agents.
The two regimes reinforce each other. Behavioral testing evidence produced for one is usually reusable for the other. For Swiss institutions, which are not directly subject to DORA, FINMA's expectations on third-party AI, testing and independent review are covered in our article on FINMA Guidance 08/2024. For a closer look at one of the attack scenarios above, see our prompt injection testing guide for banking AI agents.
Frequently asked questions
Does DORA apply to AI and LLM providers?
DORA applies to financial entities and, through them, to the ICT services they use. LLM APIs, AI features in SaaS products and AI hosting are ICT services, so the financial entity must manage them as ICT third-party risk under Articles 28 to 30. AI providers are only directly overseen by the ESAs if they are designated as critical ICT third-party providers.
Are AI model providers designated as critical ICT third-party providers under DORA?
The first list, published on 18 November 2025, designated 19 providers, including major cloud providers such as AWS, Google Cloud and Microsoft. Many AI models are hosted on these platforms. The ESAs review designations periodically, so institutions should monitor future lists rather than assume the current one is final.
Does an AI agent always support a critical or important function?
No. The classification depends on the function the agent supports and what it can do. An internal drafting assistant may not qualify; an agent that handles customer complaints, prepares payments or makes KYC decisions might. Base the assessment on the agent's permissions, actions and potential impact.
Is behavioral testing required by DORA?
DORA does not use the term "behavioral testing". It does require a risk-based digital operational resilience testing programme (Articles 24 and 25), and the ESAs' July 2026 statement calls for testing to evolve toward AI-enhanced threat scenarios. Behavioral stress testing is one way to apply those requirements to AI agents. That is our recommendation, not an explicit legal obligation.
What happens when an LLM vendor updates its model?
From a DORA perspective, a model update can be a development that materially affects service delivery. Banks should seek contractual notice of model changes where possible and re-test the agent's behavior against its approved baseline before or immediately after the change.
Next step: make your AI agents evidence-ready
DORA has made one thing clear: when a bank relies on an ICT service, it must understand it, contract for it, test it and be able to leave it. For AI agents, the hardest part to evidence is behavior.
AODIT gives your teams an independent behavioral risk rating for each agent, a repeatable baseline for model changes and a report your second line and auditors can review. Try the free demo to see an adversarial run in about two minutes, or request an evaluation to assess the agents that support your critical functions.
Sources
- Regulation (EU) 2022/2554 (DORA), EUR-Lex
- EBA, EIOPA and ESMA: ESA Statement on ICT risks from frontier AI models (JC 2026 25), 31 July 2026
- ESMA: European Supervisory Authorities designate critical ICT third-party providers, 18 November 2025
- Commission Delegated Regulation (EU) 2025/532 on subcontracting, EUR-Lex
- Commission Implementing Regulation (EU) 2024/2956 on register of information templates
