When the Financial Conduct Authority published the findings of its frontier AI review on 2 September 2026, the most striking line was not about what AI can do. It was about what firms can absorb.
Firms testing frontier AI models for cyber resilience told the regulator that the value they get is determined less by which model they use and more by the governance, tooling, controls, human oversight and operational environment around it. The FCA summarises this as "the harness": the model is not the product. The system you build around it is.
That finding deserves attention well beyond financial services. It captures where many UK professional services firms now find themselves. The technology is ready, and the constraint has moved to the firm's own ability to act on what the technology finds.
What the review actually found
The FCA's multi-firm review, Frontier AI and cyber resilience, summarises what firms reported during the regulator's engagement. It is explicit that it introduces no new rules, guidance or regulatory expectations. That makes it useful in a different way: it describes what firms themselves are actually experiencing.
Three observations stand out.
The value comes from the harness, not the model. Firms reported that outcomes from frontier AI depend on governance, tooling, controls, human oversight and the operational environment the model sits in. Two firms running the same model can get materially different results.
Discovery is outpacing response. Frontier AI is changing the speed, scale and manner in which weaknesses in software, systems and infrastructure can be discovered. As the FCA puts it, a model can identify vulnerabilities, but the firm must be able to understand and act on those findings. Firms described bottlenecks in remediation teams, engineering resources and change management processes.
Prioritisation is being redefined. Traditional severity ratings are under strain. Models can combine multiple lower-rated weaknesses to reveal attack paths that no single finding would show, which pushes firms towards broader, risk-based assessment rather than chasing scores alone.
The review follows a joint statement from the Bank of England, the FCA and HM Treasury in May 2026, which warned that the cyber capabilities of current frontier AI models are already exceeding what a skilled practitioner could achieve, at significantly higher speed, greater scale and lower cost, and set out expectations on vulnerability management, third-party and supply-chain risk, and AI-enabled defences.
The government's September signal
On 7 September 2026 the government updated Parliament on rapidly advancing frontier AI capabilities, a series of recent incidents, and its response. Two points matter for every regulated firm.
First, the incidents were not exotic. The statement notes that existing best-practice security measures and technical controls for keeping agentic AI operating within intended limits, as advised by the National Cyber Security Centre, together with comprehensive monitoring, would almost certainly have prevented them. Ordinary controls, properly applied, remain the defence.
Second, the state is treating this as an operating reality rather than a hypothesis. The statement announced £115 million for two programmes under the Defence Investment Plan, one on AI biosecurity and one to build a UK Government agentic AI incident response capability, alongside the Cyber Security and Resilience Bill, the Government Cyber Action Plan backed by £210 million, and a further £90 million over three years for resilience across the wider economy.
Why this matters beyond financial services
Law firms, accountancy practices and consultancies hold some of the most commercially sensitive data in the economy and operate under duties of confidentiality, quality and professional judgement. The harness logic applies directly to any firm deploying AI in review, audit, diligence or advisory work.
An AI tool that flags five hundred anomalies is worthless, or worse, if nobody owns the triage. Unowned AI output is not efficiency. It is unmanaged risk with a timestamp on it, and it sits awkwardly against professional obligations that assume a human somewhere is accountable for what the firm relies upon.
There is a commercial side too. Clients increasingly expect their advisers to use AI, and to be able to explain the controls around it. A firm that can show named ownership, a working triage process and a real escalation route for AI findings has something genuinely reassuring to say in pitches and due diligence questionnaires. A firm that cannot is exposed by its own tooling.
What to do about it
The FCA's own questions to firms make a ready-made checklist, and they generalise cleanly to other professions.
- Name an owner. Who owns decisions about where AI is used in security and resilience work? Decision rights should be documented and sit at a senior level.
- Route findings to people. Are model outputs assessed by people with the relevant technical and business experience? The review is clear that human oversight remains critical regardless of automation.
- Build the escalation path. Important findings need a defined route to reach decision-makers quickly, with deadlines attached, not a dashboard nobody is accountable for reading.
- Re-baseline prioritisation. Review how risks are rated. A combination of low-scored issues can matter more than one high-scored issue, so assessment should be risk-based rather than severity-led alone.
- Fund the fix, not just the search. Discovery without remediation capacity leaves you with a longer list of known weaknesses, which is a worse position than a shorter one. Plan the throughput of fixes, including change management, before you scale discovery.
- Deploy controls before capability. Before introducing agentic AI anywhere in the estate, put the NCSC-advised limits and comprehensive monitoring in place, and make sure third-party and open-source dependencies are visible.
A useful stress test: run an exercise that assumes an AI-enabled sweep has just produced a large volume of findings across your systems, and see whether the process or the people break first. That tells you exactly where the harness needs work.
If you would like a practical view of where your firm stands, get in touch.
Sources
- FCA: Frontier AI and cyber resilience, multi-firm review (published 2 September 2026)
- Bank of England, FCA and HM Treasury: joint statement on frontier AI models and cyber resilience (15 May 2026)
- UK Parliament: written statement HCWS314, Artificial intelligence update (7 September 2026)
- TLT: AI Brief, October 2026