A bank can run a high-materiality AI system for months without a single release on its own side, and still find that the system behaves differently on Tuesday than it did on Monday. The model, or the guardrails wrapped around it, belongs to a vendor, and the vendor shipped an update. Nothing in the firm's change calendar moved. The outputs did.
The Bank of England and FCA survey of AI in UK financial services, published on 26 October 2026, asked firms how they validate and reverse those updates. Among the respondents to that question, 25% reported always or usually allowing high-materiality vendor updates to propagate automatically into production. Only 48 firms answered it, and the report itself says the results should be treated with caution. That caveat travels with every update-control figure from section 4.3 of the report quoted in this guide.
This guide is written for the Chief Data and AI Officer or Head of Model Risk at a UK bank who has to sign off a change-control standard for vendor-supplied AI. It covers what counts as a vendor update, what the survey found, what PRA supervisory statements expect, and what a workable standard contains. It deliberately leaves three neighbouring subjects to other posts: ongoing monitoring of third-party models belongs to our guide on evidence for FCA AI live testing, exit and substitution planning belong to self-hosted models for regulated institutions, and the engineering of a regression suite and vendor retirement notice periods belong to golden-set regression testing.
This guide is shipped by the team behind DATS, the five-stage AI consulting system DILR.AI runs for enterprises, from Discover and Diagnose through to Scale and Run. Or see AI operating model, the engagement that covers governance, RACI and lifecycle, audit-ready by design.
What counts as a vendor model update in UK financial services?
A vendor model update, in UK financial services, is any change a third-party provider makes to a model or its guardrails that alters what a firm's AI system does, without the firm releasing anything itself. It includes new model weights behind an unchanged name, a new default version, revised safety filters and changed system behaviour. The Bank of England and FCA survey frames it as updates to vendor-provided models or their guardrails.
The distinction matters because most change-control frameworks in banks were built for software the firm deploys. A vendor update inverts that. The firm's code, configuration and prompts stay fixed while the component underneath them moves, so a release process that only fires when the firm ships will never see the change. Our enterprise AI consulting guide describes the wider placement and governance work this sits inside.
The PRA's own taxonomy helps here. In SS2/21 on outsourcing and third party risk management, off-the-shelf machine learning models appear in the list of examples of non-outsourcing third party arrangements, alongside purchased software and data. Paragraph 2.5 then says the PRA expects firms to assess the materiality and risks of all third party arrangements, whether or not they count as outsourcing. A vendor model consumed through an API is therefore not automatically an outsourcing arrangement, and it is not exempt from risk assessment either. Which side of the line a given contract falls on is a legal question for the firm, and the AI operating model is where that classification gets a named owner.
Three update types deserve separate treatment in any standard:
| Update type | What changes | Typical trigger | Why it escapes normal release control |
|---|---|---|---|
| Version upgrade | Model weights or architecture behind a deployment | Provider releases a new default version | The deployment name and API call stay the same |
| Guardrail change | Safety filters, refusal behaviour, content policies | Provider policy update | No model version number moves at all |
| Retirement migration | Forced move off a retired version | Provider retirement date | Arrives on the provider's timetable, not the firm's |
If you want that split applied to your own vendor estate, talk to us about a scoped review.
What did the 2026 Bank of England and FCA survey find about vendor updates?
The 2026 Bank of England and FCA survey found that most responding firms have some controls for validating and reversing high-materiality vendor AI updates, but not all. Among the 48 firms that answered, 48% always run regression testing on such updates and a further 25% usually do, 73% always or usually have defined rollback plans, and 25% always or usually let those updates propagate automatically into production.
Read those four figures carefully, because they answer different questions. The regression figures describe how often firms test an update before it reaches customers. The rollback figure describes whether a reversal plan exists. The propagation figure describes whether the update can bypass the firm altogether. A firm can sit in all three groups at once: it may usually regression test, hold a rollback plan, and still allow some high-materiality updates through automatically. The report does not cross-tabulate the answers, so nobody should infer that the 25% who allow automatic propagation are the same firms that skip testing.
The base matters as much as the percentages. The survey drew 196 responses, 101 from larger firms and 95 from smaller ones, and 105 respondents completed the full survey that carried the third-party questions. Only 48 of those answered the update-control question, which is why the report tells readers to treat the results with caution. We use the figures, as we did in our reading of AI regulatory constraints in financial services, as a signal of where practice stands, not as a market benchmark to quote in a board paper without that qualification.
What is not in doubt is the direction of travel. Third-party AI has become both more common and more material, which is what makes update control a board-level question rather than an engineering detail.
Third-party implementations accounted for 39% of reported AI systems in 2026, against 33% in 2024 and 17% in 2022. Among highly material systems the share was 53%, up from 35% in 2024. More than half of the AI that matters most to responding firms now runs on a component someone else can change. The same report found that 86% of respondents use AI, up from 75% in 2024, so the population of systems exposed to vendor updates is growing on both counts. Where those systems sit inside a bank, and where AI pays, is mapped in our AI for banking overview.
In a separate question in section 4.3, not limited to the 48 firms, the report notes that enhanced operational-resilience expectations, such as joint incident testing or exit planning, were reported by 27% of firms. Most respondents apply additional evaluation to third-party AI, with AI-specific risk identification and extra security and data-protection controls the measures respondents reported most often. Far fewer extend operational-resilience disciplines such as joint incident testing to those suppliers.
What does PRA SS1/23 expect when a vendor changes the model?
PRA SS1/23 expects firms in its scope to treat a vendor model as a model like any other: validated to the firm's own standards, monitored against the firm's own outcomes, and governed through a defined change process when it changes materially. SS1/23 applies to UK banks, building societies and PRA-designated investment firms with internal model approval for regulatory capital. Insurers, reinsurers and credit unions are outside its scope.
That scope point is the first thing to settle, because it decides whether the expectations below bind the reader or merely inform them. SS1/23, the PRA's model risk management principles for banks, came into effect on 17 May 2024. It does not apply to firms without permission to use internal models for regulatory capital, or to third-country branches, although the PRA says those firms may find the principles useful. A bank outside scope that adopts SS1/23 voluntarily is choosing a recognised benchmark, not meeting an obligation. For a firm in scope, SS1/23 expects a senior manager function holder to be accountable for the model risk framework, a theme we cover in AI accountability under the PRA's SMCR changes.
For firms in scope, Principle 2.6 deals with externally developed models directly. It says that, in line with SS2/21, "boards and senior management are ultimately responsible for the management of model risk, even when they enter into an outsourcing or third-party arrangement." It then sets three expectations for vendor models: satisfy yourself that they have been validated to the same standards as your internal ones, verify the relevance of vendor data and assumptions, and validate your own use of the product with ongoing monitoring and outcomes analysis using your own outcomes. What that monitoring looks like once a system is live is the subject of our guide to AI live testing evidence.
The change provisions are where vendor updates bite. Principle 3.3 expects model development testing to be conducted for material model changes, including changes over time in dynamic models, and to compare the model's output before the change with its output after the change against actual observations and outcomes. The PRA calls this parallel outcomes analysis. SS1/23 also expects the firm's policy to define the procedures that apply when a material change is made to a model and the materiality criteria used to judge prospective changes, and it expects the depth of independent review of a change to be commensurate with the change's materiality.
Put those together and a silent vendor upgrade sits awkwardly against SS1/23. If a high-materiality model's weights change without the firm's knowledge, no materiality assessment ran, no parallel outcomes analysis compared old and new behaviour, and no reviewer signed it off. For a firm in scope, an automatic propagation setting on a material model is, in our reading, hard to reconcile with a change process that has to run before the change takes effect.
Firms outside SS1/23 are not off the hook for outcomes. The FCA's Consumer Duty binds FCA-regulated firms, not their technology vendors, to deliver good outcomes for retail customers, and an update that degrades a customer-facing assistant is the firm's outcome problem whoever shipped the weights.
Why is automatic propagation the real control point?
Automatic propagation is the real control point because it decides whether a vendor update reaches production before or after the firm has had a chance to test it. Every other control in a vendor-update standard, from regression testing to rollback, assumes the firm sees the change first. Some hosted AI platforms expose this as a per-deployment upgrade setting, so the decision can sit in a configuration screen rather than a contract.
Microsoft's documentation for model versioning in Microsoft Foundry shows what that setting looks like in practice. For standard deployment types, Azure OpenAI deployments can opt out of automatic model version upgrades, in which case a new version needs a manual upgrade and the deployment stops working when its model is retired. They can upgrade once a new default version becomes available, or they can upgrade once the current version expires. The same page warns that customers might notice changes in model behaviour and compatibility after a version upgrade, and recommends testing applications and workflows with the new version.
Two further details on that page matter to a model risk function. Azure maintains the previous major version of a model until its retirement date, so a firm can switch back if needed, which makes a tested rollback a real option within that window. And Azure upgrades deployments during the scheduled upgrade window even where the new version is not yet available in a region. A deployment set to follow the default will move on the provider's date. We do not take a view here on which policy any particular firm should pick, only that the choice is a governed decision and should be recorded as one.
The survey's challenge data points the same way. Respondents cited information security or operational resilience concerns as the most frequent major constraint on third-party AI, at 40%, followed by limited transparency into third-party systems at 34%, the problem that model supply chain assurance tries to narrow. Counting major and moderate constraints together, difficulty negotiating contractual protections, which the report illustrates with liability, audit rights, data usage and advance notice of changes, was the second most cited challenge. If notice cannot be secured by contract, the deployment setting is the control the firm actually holds.
The same logic underpins our AI execution office, embedded delivery in which the production placements, and the change controls around them, are owned by the client.
What should a vendor update change-control standard contain?
A vendor update change-control standard should contain five elements: a materiality classification for every vendor-supplied model, a rule that high-materiality deployments do not upgrade automatically without a recorded risk acceptance, a parallel outcomes test before release, a named approver with staged traffic, and a rehearsed rollback path. Classification, parallel testing and approval map directly to SS1/23 expectations; holding propagation and rehearsing rollback make them workable for a vendor model.
Classify before the update arrives. The materiality tier of each vendor model should be set when the model enters the inventory, not when an update lands. A credit-decision support model and an internal search assistant should not share a change path. Our guide to decision support and agent autonomy in financial services sets out how autonomy level feeds that tiering.
Hold propagation on anything material. For a high-materiality deployment the default should be pinned to a version, with upgrades pulled deliberately. Where a platform does not allow pinning, the firm should record that as a risk acceptance with a named owner, because it means the firm cannot meet a change process that runs before the change.
Test parallel outcomes, not just accuracy. SS1/23's parallel outcomes analysis compares old and new outputs against what actually happened. For a generative system that usually means replaying a fixed, versioned set of real cases through both versions and scoring the differences that matter to customers and to controls. The mechanics of building that set are covered in golden-set regression testing.
Approve with independence scaled to materiality. The reviewer should be independent of the team that wants the upgrade, with depth matched to the change. A guardrail tweak on a low-tier tool can take a lightweight sign-off. A new model version under a customer-facing complaints assistant should not. Once approved, release to a small share of traffic first and widen only when outcomes hold.
Rehearse the rollback. A rollback plan that has never been run is a document, not a control. The 73% of survey respondents who always or usually have defined rollback plans should be asking whether those plans have been exercised. Version pinning and bundled releases are the practical tools here, and our piece on release management and change control covers how to pin model and prompt together.
The same discipline sits behind one of the six named enterprise AI solutions on our site: evaluation and observability for production agents.
What should the contract say about vendor model changes?
The contract with an AI vendor should give the firm notice of model or guardrail changes early enough to test them, a way to stay on a known version while it does, and a right to object or exit where a change materially increases risk. PRA SS2/21 sets a comparable expectation for sub-outsourcing and material changes in outsourcing agreements, which firms can use as a drafting reference for AI contracts.
The precise text is worth reading rather than paraphrasing loosely. In its sub-outsourcing section, SS2/21 expects outsourcing agreements to require the provider to inform the firm of planned sub-outsourcing or material changes, and says firms should be informed sufficiently early to carry out a risk assessment and object before the changes come into effect. That provision is written for outsourcing arrangements and sub-outsourcing, not specifically for model updates, so a firm should not claim that SS2/21 imposes a model-update notice duty on every AI vendor. It is, however, the clearest statement of what a regulator regards as adequate notice in a third-party relationship. A March 2026 update to SS2/21 takes effect from 18 March 2027; what changes then for material third-party AI is covered in our guide to self-hosted models for regulated institutions.
Retirement notice is not the same as change notice. Providers publish deprecation timetables for retiring whole model versions, and those periods are compared in our guide to golden-set regression testing. What a bank's model risk function usually needs is different: notice of behavioural change inside a version family, including guardrail and safety-policy changes that move no version number. That is the gap a contract should close, and the survey suggests it is a hard one to negotiate.
A practical drafting checklist for the commercial and legal teams:
- A defined notice period for any change to model weights, default version or guardrails on the contracted service.
- A right to remain on the current version for a stated period after a new one is released.
- Release notes detailed enough to scope a parallel outcomes test.
- Access to the previous version for rollback until a stated date.
- A right to object, and an exit route, where a change materially increases risk to the firm or its customers.
None of these clauses replaces the internal standard. A firm with perfect notice but no test, approval and rollback path has simply been told in advance about a change it cannot govern. Building that path inside the firm is what an embedded AI execution office is for.
How does vendor concentration change the update risk?
Vendor concentration changes the update risk because a single provider's update can reach many firms, and many systems within one firm, on the same day. The 2026 Bank of England and FCA survey found that the three largest model API or hosting providers accounted for 54% of all reported model relationships. A change at one of them is a correlated event across the market, not an isolated release.
Within a firm the effect is similar. When several high-materiality systems sit on one provider's model, an automatic upgrade does not move one system, it moves all of them together. That is a strong reason to stagger upgrades across systems even when each one has passed its own test, and to keep the update-control register at the level of the underlying model as well as the business application.
The survey also suggests that most reported model relationships are seen as replaceable. Of the 226 reported model API or hosting relationships, only 14% were deemed both high criticality and hard to substitute. In our reading that is reassuring for exit planning, but it does not help with updates: a provider that is easy to leave can still change behaviour overnight while the firm is using it. Substitution and exit planning are covered in depth in the self-hosted models guide.
What is the best way to control vendor AI updates in 2026?
The best way to control vendor AI updates in 2026 is a firm-owned standard that pins high-materiality deployments, tests parallel outcomes before release and rehearses rollback, backed by contract terms that give notice of behavioural change. The right partner to build it depends on scale: a firm-wide model risk transformation, a single-platform estate and a targeted governance build each favour a different kind of help.
We would judge any approach, including our own, on four criteria. Does it start from the firm's own outcomes rather than the vendor's benchmark scores? Does it produce change records a model risk reviewer can read without the build team in the room? Does it treat guardrail changes as rigorously as version changes? And does it fit the firm's size and the materiality of its AI estate? These are the same tests our enterprise AI consulting guide applies to any adviser.
On those criteria, a large bank running a firm-wide model risk transformation across hundreds of models may, in our view, be better served by a large advisory practice such as Deloitte, KPMG, PwC or EY, with the scale to staff a programme of that breadth. A firm whose AI estate runs almost entirely on one hyperscaler, such as Azure, AWS or Google Cloud, may get a long way with the platform's own version controls and evaluation tooling before it needs outside help. Specialist AI consultancies such as Faculty are another route for firms that want outside data science capacity.
DATS fits a narrower case: a bank or financial firm that needs the change-control standard designed and the first high-materiality placements put through it, by senior practitioners who ship code, not decks. Our approach starts from a diagnostic that produces a ranked roadmap of where AI belongs and where it does not, and the operating model then sets governance, RACI and lifecycle for the placements that survive. If your vendor-update question is part of a wider review, talk to the team about scope.
Does SS1/23 apply to insurers using vendor AI models?
SS1/23 does not apply to insurers or reinsurers. The PRA's model risk management principles cover UK banks, building societies and PRA-designated investment firms with internal model approval for regulatory capital, and the statement excludes insurers, reinsurers and credit unions in terms. An insurer can still adopt the principles as a benchmark, and its wider third-party obligations come from other PRA and FCA requirements.
For an insurer, the practical consequence is that the change-control standard in this guide is good practice rather than a supervisory expectation under SS1/23. The governance case for pinning, testing and rollback does not depend on that scope, and our about page explains how we work across regulated sectors.
Is a vendor AI model API an outsourcing arrangement under SS2/21?
A vendor AI model API is not automatically an outsourcing arrangement under PRA SS2/21. The statement lists off-the-shelf machine learning models among examples of non-outsourcing third party arrangements, but whether a specific AI service is outsourcing depends on what the provider actually performs for the firm. SS2/21 expects firms to assess the materiality and risks of every third party arrangement either way.
The classification changes which contractual expectations apply, so it is worth recording explicitly for each AI vendor, with the reasoning, in the third-party register. Our enterprise AI consulting guide shows where that register fits in a wider AI operating model.
Want this in production? See how DATS AI consulting places governed AI inside regulated firms, read about our approach to placing AI inside enterprise systems, explore the six enterprise AI solutions, or browse more industry AI guides.
Govern the change you did not ship.
30-min scoping call · No deck · Confidential. We will tell you where vendor AI updates already reach your customers untested, and what a standard to stop that looks like.
Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.