Eight in ten UK financial advisers are comfortable letting AI assemble the material behind an annual review. Fewer than half, 43%, are comfortable with it anywhere near a pension transfer. Those figures come from research by the platform technology firm GBST with the consultancy the lang cat, published this month from a survey of 178 advisers. Read as a governance result they are reassuring: the profession has drawn its own boundary and put it roughly where a regulator would. Read as a question about advice quality they measure the wrong thing. Every number in that survey describes what an adviser will let a system decide. None of them describes what happens to an adviser who spends two years approving decisions the system kept getting right.
Strategic Insight: The industry has spent two years governing AI on the accuracy axis, asking whether a tool is good enough to be trusted. The consequential question runs the other way: what a reliable tool does to the person whose independent view is the control.
The line advisers drew, and the axis it sits on
The GBST research is the clearest UK picture we have of where practitioners themselves want the boundary. Comfort concentrates in high-volume work with a defined right answer, and thins as decisions approach the client’s money. We covered the findings when they landed, in advisers welcome AI for admin but not for client money. Rob DeDominicis, GBST’s chief executive, reads the pattern generously and probably correctly: “Rather than resisting AI, advisers have drawn a sensible boundary around it.”
| Finding | Figure | What it actually measures |
|---|---|---|
| Assembling the pack behind an annual review | 80% comfortable | Appetite for delegating work with a verifiable right answer |
| Pension transfers | 43% comfortable, 29% neutral, 29% uncomfortable | Where advisers stop delegating the decision, not where they stop delegating the thinking |
| CIP switching and rebalancing | 53% comfortable, 22% neutral, 25% uncomfortable | A boundary drawn by consequence, not by cognitive load |
| Could not describe what agentic AI does | 31% | The boundary is being drawn partly in the dark |
Source: GBST and the lang cat, survey of 178 advisers, August 2026.
Every row of that table is an answer to the same question: how much authority does the machine get? It is a reasonable question and the answers are sensible. But authority is only one of two things being transferred, and it is the visible one. The other is the work of forming a view, and that transfers on a schedule nobody is tracking, in exactly the tasks where advisers said they were most comfortable.
Consider what “collating data for the annual review” means once the tool is good. The adviser no longer assembles the client’s position from the underlying documents. They read a summary of it. The suitability rationale arrives pre-articulated. The adviser’s contribution shifts from constructing an argument to inspecting one. Nothing about that is negligent, and on any given case it is probably better work than a tired human would produce at five o’clock. The cost is not paid on any given case.
Why does the compliance frame miss this?
Compliance asks three questions of an AI-assisted recommendation, and it can answer all three. Was the output correct? Was it reviewed by a competent person? Can we evidence that review? A firm with a decent file-check regime will pass on all three for years whilst the underlying problem develops, because the problem is not located in the output, the reviewer’s competence on paper, or the audit trail.
Deskilling asks a fourth question, and no file can answer it. Could the reviewer have produced the disagreement? Not: did they check. Did they retain the capability to have arrived somewhere else.
Critical Context: The Consumer Duty “requires firms to put their customers’ needs first” and to act to deliver good outcomes. It is written in the language of outcomes precisely so it cannot be discharged by process. A firm whose advisers reliably approve good recommendations they could not independently have reached is delivering good outcomes by a mechanism that has no defence when the mechanism fails.
This is where the professional standard and the technology part company. The fiduciary obligation never concerned itself with what produces a recommendation. It concerns who answers for one, and answerability that cannot actually be exercised is a formality. An adviser who cannot explain, with the screen turned off, why this allocation suits this client has not done the thing the standard exists for, whatever the file shows.
John O’Connell, founder and chief executive of The Oasis Group, put the mechanism plainly in Professional Wealth Management on 20 August. AI, he argues, is not taking the adviser’s place. It is “quietly recruiting them into rubber-stamping conclusions they would have challenged if a person had proposed them”. The distinction that matters in his piece is between an output treated as finished and an output treated as “a draft that requires the same scrutiny a junior analyst’s work would get”. Firms have built controls for the first. Almost nobody has built anything for the second, because the second is not a control at all. It is a capability.
The advisers who go first are your best ones
The intuitive risk model has this backwards. Firms worry about the inexperienced adviser leaning on the tool, and supervise accordingly: more file checks on new joiners, closer oversight in the first two years. That is sensible for a different risk.
Trust in a tool is earned the way trust in a colleague is earned, by repetition. An adviser with fifteen years of good outcomes from a system has a rational, evidence-based confidence in it. Rationality is the trap. The confidence is correct right up until a case the system’s assumptions do not cover, and by then the habit of hunting for that case has gone unused long enough to have faded. Seniority does not protect against this. It accelerates it, because seniority is what generates the track record that justifies the trust.
Cognitive load pushes in the same direction. O’Connell cites peer-reviewed research into how advisers interact with AI systems, which found that “trust and cognitive load move in opposite directions”: the more taxed the adviser, the more readily they wave a recommendation through unchecked. Which means review is weakest at exactly the hour when a full day of client meetings has finished and there is one more pack to sign off. Everything needed to skip a check is already present before the system has produced a single bad answer.
⚠️ Warning: Presentation does the work here, not content. A hedged wrong answer gets challenged. Lay the identical error out cleanly, state it without qualification, and it gets approved. Systems are consistently excellent at presentation regardless of whether the substance holds, which makes polish an inverse indicator of the scrutiny a recommendation receives.
Two adjacent mechanisms are worth separating, because firms routinely conflate them. Sycophancy, which we examined in AI sycophancy: the judgement risk your reviews miss, is the model flattering the reasoning the professional already brought. Deskilling is the reasoning capability itself thinning through disuse. The first can happen on day one to an adviser at the height of their powers. The second takes years and leaves the day-one adviser untouched. A firm can be fully exposed to both and see neither, because both are invisible to output review. The general version of the atrophy argument, with the MIT and Wharton evidence behind it, is in when AI thinks for us. What follows is the regulated-profession case, where the eroded capability is not a nice-to-have but the substance of the duty.
Who carries what
| Group | What changes for them | What they need | What tells you it is working |
|---|---|---|---|
| Advisers | They are no longer the first voice to offer a view on the case | Permission and time to reach a view before opening the output | They can articulate a rationale that differs from the system’s, and sometimes do |
| Compliance and the accountable senior manager | File checks certify a review that may be hollow | A second test aimed at capability, not correctness | Sampling that measures independent reasoning, not just sign-off |
| Training and competence supervisors | Behavioural bias training covers clients only | Coverage of automation bias, authority bias and confirmation bias in the adviser | T&C records show judgement assessed without tool access |
| Clients | Advice quality now depends on an invisible variable | Nothing they can see or ask for | Outcomes hold when a case falls outside the system’s assumptions |
What to change, and in what order
None of this argues for keeping AI out of the advice process, and firms that read it that way will lose to firms that do not. The high-volume, rules-bound work is exactly where the technology earns its keep, and DeDominicis is right that this is “necessary work, but it takes up time without adding visible value for clients”. The argument is against a governance model that stops at output accuracy.
Implementation Note: The three phases below are sequenced deliberately. You cannot test a capability you have not baselined, and you cannot hold one you have never tested.
A judgement-preservation framework
Phase 1: see it (4 to 6 weeks)
- Add automation bias, authority bias and confirmation bias to the T&C syllabus, framed as risks to the adviser rather than the client
- Identify which advice steps have moved from construction to inspection since AI adoption
- Baseline: on a sample of cases, ask advisers to state their view before opening the tool’s output
Phase 2: test it (one quarter)
- Introduce a periodic unaided assessment, where advisers reason a case without system access
- Change file-check sampling to record whether the adviser’s rationale is independent of the output’s
- Track disagreement rate, and treat a rate near zero as a finding rather than a success
Phase 3: hold it (ongoing)
- Require a three-question challenge on any AI-assisted recommendation before it reaches a client: what did the system assume about this client, what might it not have had, and would I reach this conclusion building it myself
- Report judgement-capability metrics to the board alongside adoption metrics
- Review annually against the FCA’s evolving position on decisions taken without a person in the loop
If you are just starting, the cheapest useful move is the baseline in Phase 1. It costs an hour per adviser and it is the only measurement that gets harder the longer you wait, because once the habit has formed you can no longer establish what was lost.
If you are already deployed at scale, look first at your most experienced advisers and your busiest hours, not your new joiners. That is where the exposure concentrates, and it is the opposite of where supervision currently points.
If you are running agentic workflows that plan and execute multi-step processes, the three-question challenge needs a defined owner per case. Distributed accountability collapses fastest in exactly the workflows where the system carries the sequence.
Resource Reality: The syllabus change and the sampling change are perhaps two days of design work for a mid-sized network, plus an hour per adviser per quarter for the unaided assessment. The genuine cost is the disagreement rate, because a healthy one means advisers occasionally taking longer and occasionally being wrong where the system was right.
Four things this breaks that nobody costs
The efficiency case double-counts. AI business cases assume the adviser’s review is free, because the adviser was always going to be there. Preserving judgement means paying for a review that is deliberately slower than approval, and that cost lands against the same benefit line that justified the tool. Mitigation: cost the review honestly at business-case stage rather than discovering it in year three, and scope the tool to the tasks where the saving survives an honest review cost.
Your best evidence of safety is also your risk signal. A near-zero override rate reads as a well-performing tool and a well-supervised process. It is equally consistent with advisers who have stopped generating alternatives. Mitigation: never interpret override rate alone. Pair it with an unaided assessment, which is the only instrument that distinguishes the two readings.
Firm-level judgement decays faster than individual judgement. The senior adviser who can still reason a case unaided learned it constructing cases by hand. The joiner who arrives into an AI-assisted workflow never does that apprenticeship, so the capability is not merely thinning, it is failing to form. Mitigation: build unaided casework into the competency pathway deliberately, the way trainee pilots still fly manually.
The remediation exposure is retrospective and large. If a systemic assumption in a tool proves wrong, the firm’s defence is that a competent professional independently sanctioned each case. If that sanction was a formality, the exposure is the whole book of affected cases, not the individual file. Mitigation: keep contemporaneous evidence of independent reasoning, not just of review, because the difference between the two is what a past business review would turn on. The parallel is playing out in law, where the SRA has extended supervision duties to AI-assisted work: see AI hallucination is a supervision problem.
Reality Check: This is not a twelve-week programme with a completion date. Judgement capability is a standing condition, like liquidity, and it degrades quietly when nothing is measuring it. Expect the first unaided assessment to be uncomfortable and treat that discomfort as the finding.
The takeaway
The Mills Review, as reported alongside the GBST findings, asked the FCA to keep watch on the drift towards open-ended decisions taken with nobody in the loop, and to adapt its framework accordingly. That is the right thing to watch, and it is not the whole exposure. A person can be involved at every decision point, documented, competent, contractually accountable, and still not be exercising the judgement the standard assumes, because the capability that made their involvement meaningful has thinned under exactly the conditions the firm designed for efficiency.
Three things separate firms that keep the capability from firms that discover they lost it:
- Measuring judgement, not just adoption. Adoption metrics are collected everywhere. Capability metrics are collected almost nowhere, and only one of the two tells you whether your controls are real.
- Supervising seniority, not just inexperience. The exposure sits with the advisers whose track record earned rational trust in the tool, which is the inverse of where oversight currently points.
- Protecting the disagreement. A profession whose practitioners never differ from the system has not achieved alignment. It has stopped generating the second opinion the whole structure depends on.
Strategic Insight: Treat AI output as a draft from a capable junior, not a finished recommendation. That single reframing does more governance work than most AI policies, because it restores the one thing accuracy controls cannot: a professional who is expected to have an opinion of their own.
O’Connell ends on the competitive point, and it is the right one. Tooling sophistication will not decide the next ten years of this. The firms that come out ahead, he writes, “will be the ones whose advisers still know how to disagree with them”.
Next steps
This month
- Run the pre-output baseline on a sample of live cases
- Check whether your T&C syllabus covers bias in the adviser or only in the client
- Pull your override rate and ask what else it is consistent with
This quarter
- Add an unaided reasoning assessment to the T&C cycle
- Change file-check sampling to record independence of rationale
- Cost the honest review time back into the AI business case
This year
- Build unaided casework into the trainee competency pathway
- Report judgement-capability metrics to the board next to adoption
- Review the position against FCA developments on human involvement in decisions
Sources and attribution
Source: When advisers stop thinking for themselves by John O’Connell, founder and chief executive of The Oasis Group, published in Professional Wealth Management, 20 August 2026. Survey figures from GBST and the lang cat, reported by The Intermediary in August 2026.
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will keep watching how the FCA treats human involvement in AI-assisted advice decisions. Read more analysis at Insights, or get in touch.