The FCA published two documents this year. Read separately, they’re interesting. Read together, they tell you where advice is going.
Starting with the Mills Review and the engagement around it, there were 140 written submissions (Marloo included) covering four distinct themes: AI evolution; how this affects markets and firms; how this affects consumers; and how regulators need to change.
While this is a long-term vision paper, it’s also reactive. AI is already embedded deeply within thousands of practices nationwide and the landscape is similar among consumers: roughly one in five UK adults is already receptive to AI making financial decisions for them. And so, while this paper is no doubt forward looking, it reflects what’s happening in the UK market today.
Its outcome? Seven recommendations to the FCA Board that could, ultimately, take years to fully land. Read together, they reiterate what has been happening since consumer duty. Simply put, prescriptive rules are going down, and continuous oversight is going up.
More revealing than its exact recommendations is the review’s AI-autonomy spectrum that helped shape them:
- Operator: The human uses AI purely as a tool. AI does nothing without direct instruction each time (eg drafting letters, with every word checked before use).
- Collaborator: The human and AI plan and act together. It’s a back-and-forth process, with AI contributing but the human still actively shaping the outcome.
- Consultant: AI recommends; the human decides. AI does the analysis and proposes suggestions, but humans make a call before anything happens.
- Approver: AI prepares the action, and the human authorises it. The work is done by the system (eg a drafted suitability letter or a flagged review); the human’s role narrows to sign-off.
- Observer: AI acts within agreed limits; humans monitor outcomes rather than individual actions. This is the most autonomous stage with oversight happening after the fact on results not on each decision.
Turning to CP26/10, released in March 2026 ahead of the Mills Review. On the face of it, it’s a consultation paper addressing the simplification of pension and investment advice rules with no reference to AI.
Comparing CP26/10 to the Mills Review is not a headline grabber. One talks about tidying up COBS and reviewing the delivery of ongoing advice; the other discusses the future of AI-augmented advice.
What is interesting is the proposal to replace fixed annual suitability reviews with periodic reviews triggered by a client’s circumstances, not a fixed review date. CP26/10 doesn’t force anyone off the annual review but gives the option to move to a circumstance-triggered frequency. The decision, and the burden of justifying it, still remains with the firm.
Circumstance-triggered reviews should provide much better client outcomes than defined annual review periods but they only really work at scale if “something” or “someone” is watching for circumstance changes between review dates. Without the infrastructure to support that judgement, most firms won’t change and, if anything, may go the other way. For example, does someone far from retirement and in accumulation need to be told annually that nothing has changed?
See also: Vanguard: Over half of advisers are using AI for non-administrative tasks
That’s exactly where Mills becomes relevant again. Most firms sit at operator or collaborator on the autonomy spectrum: AI summarising meetings, drafting advice and, maybe, automating their tasks resulting from a meeting. That’s a monumental shift in productivity, but not what effective circumstance-triggered reviews need which is closer to the observer: AI watching continuously across a whole client bank, within agreed limits, surfacing changes rather than waiting to be asked.
Once a firm can genuinely operate at that end of the spectrum, the circumstance-based review CP26/10 proposes, but never says how to deliver at scale, becomes possible.
Mills is honest in arguing that as AI moves toward the observer stage, meaningful human control and with it, consent, accountability and evidence of good outcomes get harder to prove. That’s true in general, and it’s the right note of caution for a longer-term vision paper. But I’d argue the opposite holds in this case. A fixed, annual suitability review produces one file note as evidence and as CP26/10’s own drafting makes clear, “annual review conducted, no changes” is thin evidence.
An AI operating at observer, watching a client bank continuously, produces the opposite: a running, timestamped record of exactly when circumstances change, what was flagged, and why a review was or wasn’t triggered. That’s not weaker evidence of control. It’s stronger evidence than the process it’s replacing ever generated.
What changes isn’t the principle. It’s what firms are expected to have in place to prove they’re meeting it. Under consumer duty, that meant evidencing good outcomes rather than just following a checklist. Under CP26/10, that bar gets more specific again, not more advice, but more justification per decision.
It is no longer enough to show a review happened annually; a firm must show why that frequency was right for that client, and if it can’t, it loses the right to charge for the service at all. Under Mills, that same individualised burden of proof extends further still, to a system acting continuously rather than a person acting on a fixed date.
Read each of these individually, and the bar keeps rising. Three separate documents, each asking firms to prove something more specific than the last. Read them together, alongside what AI enables, the opposite becomes true. It becomes easier, not harder, to meet FCA’s requirements, because AI doesn’t just help firms clear a higher bar, it does the evidencing for you. Continuous, timestamped records isn’t something a firm has to build to satisfy a regulator.
It’s what falls out of an efficient system with evidencing as a by-product of the advice process. The direction of travel hasn’t changed since consumer duty. What’s changing is that firms no longer have to choose between doing right by the client and proving it because AI can help you do both at once.
Jack Allington is product and GTM lead at Marloo
