Third-party systems made up 53% of the high-materiality AI reported by UK financial firms in 2026, up from 35% in 2024, according to the Bank of England and FCA’s latest survey. For those bought-in systems, 24% of respondents said they relied on deep knowledge of how they work, against 77% for systems built internally, and 81% relied on vendor assurance and contracts instead. That gives a UK test for an argument made in ESG Today in May: that supplier AI is moving out of ESG and turning into an obligation for risk and compliance teams. On the documents we reviewed, the exposure is well evidenced, the obligations are patchy, and the market meant to supply independent proof is still developing.
Strategic Insight: The useful question is narrower than “is AI a supply-chain risk?”. It is which UK duties already require a buyer to show evidence about a supplier’s AI, and who can give that evidence if the supplier will not.
What does the supply-chain argument actually claim?
The ESG Today piece is a guest contribution from Simon Jaehnig, IntegrityNext’s president and chief strategy and innovation officer. He argues that buyers now carry exposure to decisions made by software they do not own or directly run, and that most third-party risk frameworks were not built with autonomous decision-making in mind. His forecast is that “the standard will shift from intent to evidence”: who is accountable, where controls sit, and whether decisions can be audited afterwards.
Two caveats matter. IntegrityNext sells supply-chain compliance software, so the author’s company works in the field he is writing about. The post also cites no figures and does not mention the UK. It should be read as a thesis to test, not as a finding.
It also addresses a slightly different case from the one the Bank and FCA survey measures. Jaehnig writes mainly about AI that suppliers use in their own operations, including tools deployed without approval that are hard to detect beyond tier-one relationships. The Bank and FCA figures concern AI systems that firms themselves report using, supplied by third parties. The documents we reviewed do not measure how far UK buyers can see into the first case.
The numbers from UK finance
| Measure | 2026 survey result | What it suggests |
|---|---|---|
| Third-party share of high-materiality AI systems | 53%, up from 35% in 2024 | More of the AI firms rate as most material now comes from outside |
| Reliance on vendor assurance and contracts for third-party systems | 81% | Understanding of these systems often rests on what the supplier says and signs |
| Vendor due diligence, against ongoing monitoring | Over 90%, against around half | Data risk controls are reported more widely before deployment than after it |
| Limited transparency as a major constraint | 34% | About a third of respondents report this barrier directly |
These results come from 196 firms in one regulated sector and should not be read across to UK business as a whole. Our earlier analysis of the same survey looks at what it says about AI agents and supplier concentration.
Reality Check: Vendor due diligence is close to universal among these respondents, so the gap is not that firms skip it. The weaker area is what happens afterwards: post-deployment monitoring, leakage tests and periodic reviews were each reported by only around half.
Where does UK law already ask for evidence?
There is no single answer, because the duties arrive through separate regimes with different reach and force.
| Route | Who it reaches | Force today |
|---|---|---|
| UK GDPR, as applied by the ICO | Organisations processing personal data in AI systems | Existing law; the ICO’s AI audit toolkit is under review after the Data (Use and Access) Act |
| Cyber Security and Resilience Bill | Medium and large managed service providers, data centres above thresholds, designated critical suppliers | Not yet law; Lords third reading listed for 2 November 2026 |
| AI cyber security code of practice | AI developers and system operators, including organisations that only use third-party AI components | Voluntary |
| Civil liability | Participants in the AI supply chain | Existing common law; product liability reform still at consultation |
Data protection is where the evidence argument is firmest. The ICO’s position is that signing a contract does not by itself fix whether a party is controller, joint controller or processor, because that depends on what happens in practice. It adds that other contracts between developers and deployers “will not override the fact-based assessment”. It also warns that deployers of closed-access models may lack meaningful control over all the processing at deployment.
The ICO’s audit toolkit on contracts and third parties sets out what it looks for in practice. That includes due diligence on accuracy, bias and trade-offs before procurement, contract clauses that allow audits, and in-life checks that agreed roles are being followed. It links those gaps to possible breaches of articles 5(2) and 28 of UK GDPR. The reach is limited to personal data, but within that limit it comes close to the shift Jaehnig predicts.
The Cyber Security and Resilience Bill goes after suppliers directly, but for cyber security rather than AI decision-making. The government’s summary says regulators will gain a power to designate critical suppliers, and that the regime “does not cover every UK organisation”. That summary was first published in November 2025 and revised in March 2026. The Bill has since been amended in the Lords, and the Lords amendments published on 3 November are now before the Commons, so this describes the policy as the government summarised it.
The AI code asks for something similar without compulsion. The code says a developer or system operator choosing an external provider “shall undertake a due diligence assessment”, but the code itself is voluntary.
Critical Context: Contract does different work in different regimes. Law-firm summaries of the UK Jurisdiction Taskforce’s legal statement say contract is generally expected to be the main basis for allocating AI liability along the supply chain. The ICO says a contract does not settle data protection roles. A buyer cannot assume that a well-drafted agreement answers both questions.
Liability sits alongside all this. The UK Jurisdiction Taskforce statement deems data protection out of its scope, along with intellectual property and AI use by public authorities. Burges Salmon’s summary reads it as signalling that “the use of AI does not insulate anyone from liability for foreseeable harm”. Burges Salmon’s head of AI advisory sat on the supporting expert group. Separately, the Law Commission’s consultation proposes treating software and AI systems as products in their own right.
Who can supply the proof?
Here the supply-chain thesis runs into a practical gap. Jaehnig points to cyber security and sanctions compliance as areas that ended up with common baselines and minimum controls. AI has not reached that point in the UK on the evidence we reviewed.
DSIT’s third-party assurance roadmap, published in September 2025, describes a market worth over £1 billion whose quality “is currently unclear”. It said existing AI assurance certifications were hard to judge and that none was issued by UKAS-accredited organisations. It also said technical standards were still developing. We have not checked whether those points still hold 14 months later.
The roadmap also records the access problem from the other side. Assurance providers name poor access to information on AI systems as a barrier, and audited firms may be unwilling to share it because of commercial confidentiality, among other reasons. Limited transparency into third-party systems was a major constraint for 34% of Bank and FCA respondents, and difficulty negotiating contractual protections such as audit rights ranked second among challenges once moderate constraints were counted.
Hidden Cost: An audit clause is only as useful as the supplier’s willingness to open its systems, and a third-party assurer is only as useful as the standard it assures against. On DSIT’s 2025 account, both were unsettled.
Who is affected, and how?
Financial firms provide the clearest evidence of the exposure in what we reviewed. No area of governance uncertainty or implementation difficulty was cited more often than oversight of third-party AI: 54% selected applying model risk principles to vendor-supplied models, and 49% selected oversight of systems that third parties supply or materially shape. Moody’s has made a related warning about banks’ dependence on AI suppliers.
Data protection officers and procurement teams at any organisation using AI on personal data sit within the ICO’s expectations, whatever the sector. Jaehnig’s point about ownership fits here: he says the exposure is shared across six functions, from procurement and legal to enterprise risk.
Suppliers face the change from the other direction. Managed service providers would be regulated directly under the Bill, with the Information Commission as regulator. Generative AI developers told the ICO they accept being controllers at the development stage but do not want controllership of downstream deployment. The ICO kept its position that joint controllership remains likely in closed-access deployments, whilst stressing that this does not mean joint and equal responsibility for everything.
What remains unresolved?
Visibility beyond the direct supplier
The ICO toolkit expects organisations to consider the whole supply chain, and its governance section suggests questionnaires and direct meetings with key partners to map information flows. That covers personal data. For AI that a supplier uses in its own logistics or production planning, with no personal data involved, the data protection route falls away, and the government frames the Bill around cyber security. Jaehnig’s concern stands as a question, but the documents we reviewed do not measure its scale in the UK.
Change after the contract is signed
Jaehnig argues that point-in-time assessments are no longer enough. The Bank and FCA data gives that some support. 25% of respondents said they always or usually allow high-materiality vendor updates to propagate automatically into production, though the regulators caution that only 48 firms answered that question. Stronger resilience expectations, for example exit plans or incident tests run jointly with the supplier, were reported by 27%.
Bargaining power
One industry body told the ICO that a party with a dominant negotiating position could unfairly push obligations onto the other in joint controllership. The Bank and FCA survey reports high concentration too, with the three largest providers accounting for 79% of cloud and compute relationships and 54% of model relationships. Whether a mid-sized buyer can obtain audit rights from a large provider is a commercial question that these documents flag but do not answer.
Warning ⚠️: Evidence from one sector should not be taken as the state of UK business. Nothing we reviewed measures how firms outside regulated finance manage third-party AI.
What to watch
Four checkable signposts would show whether the “intent to evidence” shift is arriving in the UK or remains a forecast.
- The Cyber Security and Resilience Bill. If it becomes law, watch the first critical supplier designations. A designation that reached an AI model or hosting provider would bring AI suppliers directly into this regime. Designations limited to traditional IT suppliers would leave the AI question where it is.
- The ICO’s updated guidance. The regulator says its generative AI positions will feed into revised guidance, and respondents asked for practical examples of controllership. Worked examples for closed-access deployment would give buyers something concrete to apply. A restatement of principle would not.
- The Law Commission’s product liability consultation. It closes on 14 January 2027, with final recommendations expected in late 2027. Whether AI systems stay within the proposed definition of a product, and how the proposed presumption of defectiveness fares, will shape who can be held liable when an AI product is defective.
- The next Bank and FCA survey. Earlier rounds ran in 2019, 2022, 2024 and 2026. The figures to compare are the 24% relying on deep knowledge of third-party systems and the roughly half reporting ongoing monitoring.
Jaehnig’s framing is useful, even though it comes from a company with a stake in the answer. The UK record supports his description of the exposure, and in data protection it already supports his point about evidence. Elsewhere, duties are proposed, voluntary, confined to cyber security or left to general civil liability. For now, the UK’s stated expectation of proof is running ahead of the means of producing it, and a buyer relying on vendor assurance alone is relying on a party with a commercial interest in giving it.
Source: AI Governance Is Becoming a Supply Chain Risk, Not Just an ESG Issue (ESG Today, 2026), guest post by Simon Jaehnig of IntegrityNext. Additional sources: Artificial intelligence in UK financial services - 2026 (Bank of England and FCA, 2026); Allocating controllership across the generative AI supply chain (ICO); Law Commission product liability consultation (2026); AI liability under English law (A&O Shearman).
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching whether the first critical supplier designations and the ICO’s revised guidance turn this expectation of evidence into something buyers can act on. Read more analysis at Insights, or get in touch.