Why your RCSA does not pass the new operational resilience test

Welcome back to BuildMoat.

Some people have asked what BuildMoat really means. Simple answer: your moat is what keeps you in business. Not technology. Not speed. It's how well you orchestrate people and AI agents into a system that compounds over time. In March, that definition got a regulatory spine. The PRA published SS1/26 on operational resilience incident reporting, quietly redrawing what operational risk means when AI sits inside your important business services.

AI models, including ones you did not build, are now risk assets that must be mapped, tested against impact tolerances, and reported on when they fail. The policy takes effect March 2027. The accountability question is live now. This issue looks at the gap between that regulatory expectation and what RCSA programmes actually cover today. The tools to automate RCSA exist. The accountability layer does not. Every 2LoD function with an AI-assisted control in the first line now has a question to answer, and the FCA is asking it through supervision, not enforcement.

I use art in this newsletter. Not as decoration. As a second language for ideas that resist plain explanation. If you have a piece you think belongs here, send it through.

Right, let’s get into it.

On 18 March 2026 the Prudential Regulation Authority published Supervisory Statement SS1/26 alongside Policy Statement PS7/26, establishing how firms must report operational incidents and material third-party arrangements. The framework takes effect in March 2027. The market has read this as a resilience update. That reading is too narrow.

SS1/26 explicitly covers AI-driven systems where they support important business services. The PRA now expects firms to map AI dependencies, identify failure modes including model drift and data quality degradation, test those failures against impact tolerances, and report material incidents in a consistent, timely, and proportionate way. Read that list through a 2LoD lens and a specific control gap becomes visible. It is the gap between an AI model running in a first line workflow and a named individual who owns the risk of its output.

Start with the RCSA. Most firms treat AI as a technology risk captured under the IT or Cyber category. SS1/26 makes that categorisation insufficient. If an AI system is sitting inside the execution of an important business service, it is an operational risk in its own right, with its own control library, its own failure modes, and its own impact tolerance. That is a live RCSA update for every firm running AI in credit decisioning, customer onboarding, transaction monitoring, or complaints handling.

Now look at the tools. IBM OpenPages with Watson already offers AI-powered risk identification, incident classification, and automated reporting. MetricStream's AiSPIRE extends this with predictive risk identification, duplicate control detection, and cognitive recommendations written directly into the RCSA process. Both platforms can generate first-pass RCSA content faster and more consistently than a human risk analyst. Neither resolves the question SS1/26 forces firms to answer. Who is the named individual who reviewed the AI-generated RCSA output, applied judgement to it, and signed it off before it reached the risk committee.

Tech moves fast, but you're still playing catch-up?

That's exactly why 200K+ engineers working at Google, Meta, and Apple read The Code twice a week.

Here's what you get:

All delivered twice a week in just 2 short emails.

Join 200K+ engineers

This is the AI adoption gap in operational risk, made concrete. The productivity case for AI in RCSA is obvious. The governance case has not been built. Without the governance layer, the firm is faster at producing risk artefacts it cannot defend under regulatory scrutiny. That is not resilience. That is a documentation velocity problem.

The harder edge is control testing. SS1/26 requires firms to have tested AI failure modes against impact tolerances. A control that relies on an AI output without a documented human review step is not tested. It is asserted. When the FCA asks for evidence that controls operate effectively, or when the PRA asks for incident reporting against a tolerance breach, the firm needs to show an auditable decision trail. The Nationwide £44 million fine and Monzo £21 million fine both landed in late 2025 on the same underlying point. The FCA tested what was actually operating, not what the policy said should be operating.

The firms that will absorb SS1/26 cleanly are the ones treating AI-assisted RCSA and control testing as an accountability redesign, not a vendor selection exercise.

The FCA has quietly moved its regulatory exposure upstream. Formal enforcement investigations have dropped from 220 open in 2023 to 124 by October 2025. Voluntary Requirements, known as VREQs, have risen from 104 to 119 in the last year. Senior Manager attestations and S166 skilled person reviews are now the FCA's preferred tools. The FCA's January 2026 Enforcement Watch confirms the live escalation themes: Consumer Duty, financial crime controls, and culture and governance.

For operational risk teams this shift has three consequences. First, regulatory exposure arrives earlier, through supervision, before any formal enforcement process opens. Second, individual accountability surfaces sooner, via SMF attestation requests rather than Final Notices. Third, remediation is binding through VREQ mechanisms, not discretionary through enforcement negotiation.

The practical effect on RCSA and control testing is sharper. A VREQ on a control weakness identified in supervision does not wait for the next RCSA cycle. It requires immediate remediation and evidence. If the control in question depends on an AI-assisted process, the firm has to demonstrate, in real time, who owns the AI output, what judgement was applied before it was relied on, and what monitoring is in place for drift and degradation.

For 2LoD functions, the implication is that the 2026 supervisory environment rewards firms who can produce an auditable, named-owner decision trail on demand. It punishes firms running AI-assisted controls with governance designed for a deterministic system.

The PRA's SS1/26 and the joint FCA/PRA/BoE operational resilience package require firms to treat AI as a mapped operational dependency, tested against impact tolerances, with material incidents reported. This forces a direct update to the operational risk control library. AI-assisted controls in the first line need an explicit governance layer covering named ownership, pre-deployment approval, drift and performance monitoring, and an incident pathway. The RCSA template itself needs a control attribute recording whether AI is in the execution path and who the accountable SMF is.

Leading firms have already separated AI-generated RCSA inputs from human-approved RCSA outputs, with a mandatory sign-off step in between. Laggards are still allowing AI-assisted RCSA outputs to flow directly into risk committee packs without a named reviewer. By summer 2026, expect supervision letters to probe this distinction directly.

"For every control in our RCSA where AI is in the execution path, show me the named individual who reviewed the most recent AI-generated output, the date they reviewed it, and the specific judgement they applied before it was accepted."

Hair dye fixes gray. It also gives you roots, a smell, and a hairline that looks painted. Particle Anti-Gray Serum works at the root to gradually restore your natural color — hair and beard — no dye, no mess. Thirty-day guarantee. Get 20% off now with the code BH20.

Risk-Free & 20% Off Today!

US Treasury releases FS AI Risk Management Framework with 230 control objectives (March 2026). The Cyber Risk Institute published a Financial Services AI RMF covering bias testing, drift monitoring, explainability thresholds, and evidence requirements across the AI lifecycle. US-focused but directly relevant for UK firms benchmarking governance maturity. Link

FCA, PRA and BoE issue joint Policy Statements on operational resilience (March 2026). The tripartite package covering operational incident reporting and third-party arrangements takes effect March 2027. Firms must report material third-party arrangements, AI vendors included, under the new framework. Preparation obligations start now. Link

Nationwide fined £44m for financial crime systems failures (December 2025). FCA Final Notice highlighted customer due diligence and monitoring weaknesses. Framework existed on paper, operation was deficient. Direct read-across for any firm relying on AI-assisted monitoring controls.

Monzo fined £21m for inadequate controls during rapid growth (late 2025). Repeated breaches of a restriction on higher risk onboarding, with the FCA calling out governance of controls during scale. Lesson: the supervisor tests what operates, not what is documented.

FCA AI Live Testing expanding under AI Lab programme (ongoing). Supervised testing of AI systems with direct regulatory engagement. Feedback from the programme is expected to shape formal guidance in 2026. Firms considering how to demonstrate good practice to the regulator should track this closely. Link

Kennedys — FCA Supervision and Enforcement Trends in 2026. The clearest summary available on the FCA's shift from enforcement to supervision-led remediation. Read for the VREQ and S166 analysis.

Bank of England — SS1/26 Operational Resilience Incident Reporting. The primary source. Do not rely on secondary coverage, the implementation detail is in the supervisory statement itself.

Kennedys — Deploying AI in UK Financial Services: FCA and Data Protection Considerations. Clean map of existing framework obligations applied to AI deployment. Useful for any firm updating its AI governance policy.

The BuildMoat Newsletter is written for professionals who need practical intelligence on how AI is actually landing inside regulated workflows. Reply to this email with a challenge, a tip, or a case you want covered. Forward it to a colleague if something in here is useful.

Next Issue lands Monday 4 May. Topic: Model Risk, SR 11-7, and who signs off on LLM outputs when the model risk team cannot.

Until then.