AI adoption outpaces assurance

Where this pressure most often appears in the customer lifecycle

Onboarding

Ongoing monitoringPrimary

Exit & de-risking

Firms are weighing up more than AI.

Machine learning, graph analytics, behavioural analytics, robotic process automation and agentic AI are all moving into detection and monitoring, raising the same question: not only whether the technology performs, but whether the firm can prove it does.

July 2026 was the month UK financial services regulation stopped circling AI and committed to a direction. Within a fortnight, the FCA’s Mills Review (6 July) and HM Treasury’s Financial Services AI Adoption Plan (14 July) set out how the UK intends to govern and accelerate AI in the sector, with seventeen recommendations between them and one clear reframing of the regulatory question: from “What rule applies?” to “What can you evidence?”

There is no standalone AI rulebook. The existing framework (the Consumer Duty, SM&CR, model risk management, operational resilience) is being extended to AI, and firms are being asked to demonstrate how it is being met as AI takes on more of the work. The Bank of England and FCA’s 2024 survey found 75% of UK financial services firms already using AI; a 2026 global survey cited by the Mills Review put it at 81%, with 40% at more advanced stages of scaling or transformation. Agentic AI, barely present in surveys two years ago, is now being piloted or deployed by more than half of industry respondents.

Our AI Assurance Framework is adapted from the NIST AI Risk Management Framework and operationalised through our Risk Taxonomy and G3MA Assurance Framework (Govern, Map, Measure, Manage, Assure), the control framework we operate and validate every AI deployment against.

Assurance itself now operates at three levels. Provider assurance covers the security, transparency, resilience and foundational capability of the model or platform itself, the layer the Adoption Plan’s proposed third-party AI assurance scheme would standardise. Implementation assurance covers everything the firm builds around it: prompts, retrieval and data, orchestration, access controls, human review and integration. Outcome assurance covers the system in operation: fairness, accuracy, consumer impact, drift and continuing compliance. The scheme addresses the first level. Firms retain full responsibility for the second and third. A certified model can still sit inside an unsafe use case.

From “What rule applies?” to “What can you evidence?”

That is the position we put to the FCA in our AI Input Zone submission and to IOSCO on its AI Supervisory Toolkit: assurance credible to customers, regulators and counterparties requires independence, external challenge, reproducible tests and evidence others can verify. Internal checks are essential but not sufficient: you cannot mark your own homework.

Accountability therefore keeps resurfacing in a different form. A control used to mean a system flags something and a person investigates it. Now a model may prioritise and recommend before anyone approves it. The more consequential the decision, and the less time available to reach it, the harder it becomes to distinguish genuine judgement from endorsement.

The Mills Review’s autonomy spectrum (operator, collaborator, consultant, approver, observer) is about to become supervisory vocabulary. The pressure point the Review identifies is the shift into Approver and Observer, where opacity and model drift could open a gap between AI-mediated outcomes and any identifiable responsible party.

Agentic AI sharpens this further. Prompt-level guardrails can be bypassed under adversarial conditions; control has to be enforced at system and orchestration level, in the harness, not primarily in the instructions given to the model. Telling an agent not to do something is not the same as preventing it. The Adoption Plan’s Know Your Agent protocols, agentic payments framework, and Critical Third Parties regime (under which AWS, Google Cloud, Microsoft and Oracle became the first designated CTPs on 13 July 2026, with AI providers the logical next candidates) describe the network-layer response. The firm-level foundations most institutions have not yet built are the harness, the evaluation harness that re-runs on every model change, and the accountability model that survives them.

Firms increasingly buy this capability rather than build it. The less capacity a firm has to test a model itself, the more of that assurance question it inherits from the provider. Buying the technology does not buy the accountability: the firm still answers for how the model reaches its conclusions and what happens when a vendor changes it, often through providers now formally recognised as critical to the sector.

That is not a reason to hold the technology back. It is why the accountability conversation has to happen before deployment, not after.

It is also why trust in AI has to be engineered in, not retrofitted, the position we set out with our partner ExoBrain in our 2025 joint response to the FCA’s AI Live Testing proposals, and re-affirmed at our AI Tipping Point briefing with senior leaders in June 2026.

We are vendor-agnostic. We keep live knowledge of the AI-led financial crime tech market across monitoring, screening, KYC and fraud, because we can only assure what we actually understand. That is the difference between a governance opinion and a technical validation.

How we support you

We help firms identify, evaluate and deploy AI where it can improve economic crime controls, and work on the assurance that lets them stand behind a system’s output, from configuration and tuning through to vendor selection and ongoing monitoring.

Typical engagements include:

Decompose the system: open the AI pipeline and list every independent part inside it (router, retriever, generator, judge, reviewer) and where each part hands to the next.

Name the job: each part is a familiar problem (classification, information retrieval, generation, sequential decisions, judgement), each with an established evaluation method.

Measure: build a labelled ground-truth eval set, set a pass bar per part tied to the cost of that part failing, then measure each part in isolation and the chain end-to-end. Errors compound: four parts averaging 95% multiply to roughly 81% end-to-end, so a strong-looking system can still fail one answer in five.

Monitor: turn the eval set into the live dashboard. Re-score every release before it ships, watch for drift as inputs change, and alert on regressions with the failing part already named.

Give every part a bar and an owner. Accountability is a number plus a name, and both are required before deployment. That is what turns ‘we use AI’ into ‘we can defend AI’.

Most AI assurance still evaluates the model as one thing. We evaluate every part. That is what lets a senior manager point at a specific number, from a specific part, tested against a specific bar, in the sign-off pack. It is also why we catch what a single end-to-end accuracy score hides: strong parts still compound into weaker end-to-end results, and the fix is only ever in one of them.

See our full list of services

The same discipline we apply to your AI, we apply to our own. Our delivery model runs on persistent Plenitude-methodology context, engineered guardrails, and named accountability for every output, so when we assure your AI, we do it from a delivery bench that runs on the same standard.

Our experience

We were the lead delivery partner on the FCA’s Synthetic Data in Anti Money Laundering initiative, delivered with the Alan Turing Institute and Napier AI, the industry’s own test bed for exactly the assurance question this page is about.

A UK wealth and investment manager

Why we were brought on: a financial crime transformation programme spanning several of its wealth and investment businesses needed to address where technology and AI sat inside the target operating model.

What we did: we led current-state validation, designed the target operating model, and worked through what independent oversight reliance on AI-enabled tools actually needed.

Outcome: a phased transition roadmap that addressed the AI question inside governance design from the outset.

A UK asset manager

Why we were brought on: its financial crime standards needed uplifting across AML/CTF, CDD, screening, fraud and transaction monitoring, and delivery had to work within the capabilities and upgrade schedule of its outsourced technology platform.

What we did: we led the platform-specific workstream, designing and mobilising the standards changes so they matched what the vendor’s system could actually support, including screening configuration and CDD process changes.

Outcome: standards uplifted and implemented directly on the vendor’s platform, with governance and reporting embedded in it.

A UK regulator

Why we were involved: the regulator identified access to real financial data as a barrier to testing new economic crime technology safely, given legal and privacy constraints on using real customer data.

What we did: we delivered as a partner alongside the Alan Turing Institute and Napier AI, helping build a synthetic dataset containing realistic money laundering scenarios that firms can test new technology against without exposing sensitive data.

Outcome: a shared testing environment regulator and firms can both use to work through the exact assurance question this page is about, without either side risking real data.

17%to33%

of firms using third-party AI, in two years

75%

of UK financial services firms already using AI, as of November 2024

34%

reported a complete understanding of how their own AI systems worked

The Bank of England and FCA’s own joint survey of the industry shows AI adoption running well ahead of understanding. As of November 2024, 75% of UK financial services firms were already using AI, including for anti-money laundering, fraud and cybersecurity, but only 34% reported a complete understanding of how their own AI systems worked. Understanding fell further wherever the model came from a third party rather than being built in-house, a growing pattern, given third-party AI use had risen from 17% to 33% of firms in the space of two years. The industry was not short of adoption. It was short of firms who could explain what they had adopted.