Who Actually Understands Your Third-Party AI?

UK financial firms now source slightly more than half of their important AI from third parties. For third-party AI, 81% of them lean on what the vendor tells them and what the contract promises.

That is the plain reading of the survey the Bank of England and the Financial Conduct Authority published on 26 October 2026, the fourth in their series, with 196 responding firms. The adoption rate is not the finding that matters for an AI lead. The finding that matters is the gap between owning third-party AI and understanding it.

What the survey found about third-party AI

Bought in, and material

Of the responding firms, 86% reported using AI. Third-party implementations made up 39% of all the AI systems firms reported, and 53% of the systems they classed as high materiality. Firms decided for themselves which systems counted as high materiality, since the survey sets no formal definition. Even so, the direction is plain. Slightly more than half of the systems firms themselves rate as important now come from third parties.

Assurance in place of knowledge

The sharper numbers sit in the governance section. 77% of respondents reported deep knowledge of the AI systems they had developed in-house. For third-party AI systems, the figure was 24%. Instead, 81% relied on vendor assurance and contractual arrangements.

Meanwhile, the paperwork is in place. 94% of respondents reported firmwide AI policies, principles or standards, and 84% reported centralised AI or model-risk governance fora. In other words, the policies exist. Understanding of the bought-in systems those policies govern is far thinner.

That is not a failure of effort. It is the predictable result of buying a system whose inside the buyer never sees.

Why a contract is not the same as knowing

A contract tells you what a vendor has promised. It does not tell you what the system does on a Tuesday afternoon after the vendor has changed it.

The survey shows where that bites. The top governance challenge, selected by 54% of respondents, was applying model risk management principles to third-party or vendor-supplied models. Limited transparency into third-party AI systems was a major constraint on adoption for 34%, behind information security and operational resilience at 40%.

Updates are the quiet risk

Vendors change their models, and the buyer inherits the change. 48% of respondents said they always run regression testing on high-materiality updates, and a further 25% said they usually do. Only 48 firms answered that question, so the Bank and the FCA ask readers to treat the result with caution. Read it as a signal rather than a benchmark: even among the firms willing to answer, testing every material update is not universal. Asking a supplier for notice of upstream changes, one of the three requests in our piece on model provenance, helps only if someone then tests what arrives.

Concentration raises the stakes

The top three cloud and compute providers accounted for 79% of the relationships firms reported, and 98% of respondents used one of the three largest. Critical third-party dependencies were rated a high or severe systemic concern by 40% of respondents today, a share they expect to reach 63% within one to three years. When so much third-party AI rests on a few suppliers, a gap in one supplier’s assurance can become a gap shared across the sector. The Anthropic case showed the same pattern from the buyer’s side: dependence on one provider became the vulnerability.

What EU law already expects of third-party AI

For EU financial entities, the Digital Operational Resilience Act settles who carries the risk. Article 28(1)(a) of Regulation (EU) 2022/2554 says that a financial entity with contractual arrangements for ICT services “shall, at all times, remain fully responsible” for its obligations. The contract allocates work. It does not move responsibility.

DORA also writes the remedy into the contract. Where an ICT service supports a critical or important function, Article 30(3) requires the arrangement to include the right to monitor the provider’s performance on an ongoing basis, and exit strategies with a mandatory adequate transition period. Where third-party AI runs as such a service, those two clauses are the floor, not a negotiating win.

Outside financial services, no statute says it this bluntly. The logic still holds for a mid-market firm whose credit checks, pricing or customer screening run on a bought-in model. The vendor’s assurance is an input to your judgement, not a substitute for it.

Four questions for every third-party AI system you rely on

Before the next renewal, put four questions to the owner of each third-party AI system that feeds a decision you care about:

1. Can we explain what the system does, and where it fails, without quoting the vendor’s own material?
2. Do we test vendor updates before they reach a decision that matters, or do we find out afterwards?
3. What would we see if the system’s performance drifted, and who would see it first?
4. If we had to leave this vendor, how long would the transition take, and what would run in the meantime?

The first question is the hardest, and on this survey’s evidence a weak answer is common. Our filter for reading AI evaluation reports covers the evidence side of it: what was tested, what was claimed and what was left untested.

The second and third questions turn assurance into evidence. A vendor’s statement that an update is safe is a claim; your own test result on your own cases is knowledge. Similarly, a monitoring right in the contract is worth only what someone inside the firm actually looks at.

The fourth question is the easiest to put off, because asking it feels like planning a divorce at the wedding. DORA makes financial entities answer it for critical or important functions.

The lesson travels beyond finance

The survey is a snapshot of one regulated sector in one country. Nevertheless, its lesson travels, because the gap it measures is not specific to finance. You can sign for a third-party AI system you do not understand, and you still own what it decides.

Which of your bought-in AI systems could you explain tomorrow without the vendor in the room?