Ninety-three percent confident, thirty-five percent blind
Here's a number that should make any financial services board uncomfortable: 93% of firms believe their AI tools have improved outcomes for vulnerable customers. Here's the number that should make them more uncomfortable still: 35% of those same firms don't separately track whether that's actually true.
That's the headline finding from the Payments Association's AI Inclusion Readiness Report 2026, published in late September after surveying 110 senior decision-makers across financial services, retail and wholesale. It's a confidence survey dressed up as a readiness report, and the gap it exposes between what firms believe and what firms can prove is exactly the kind of gap Consumer Duty was designed to close.
The numbers behind the confidence
The report used the FCA's own four drivers of vulnerability — health, life events, resilience and capability — to categorise who it was asking about. Firms rated their own readiness at an average of 4.1 out of 5. That's a high score for a problem over a third of them admit they can't actually measure.
The support itself is uneven in a telling way. Sixty-seven percent of firms offer specific support for older customers, a group that's visible, well understood and easy to design for. Only 34% offer equivalent support for people going through recent adverse life events — redundancy, illness, bereavement — the kind of vulnerability that arrives suddenly and doesn't show up on a standard customer profile until it's too late.
And the consumer side of the story is blunter still: 17% of vulnerable consumers couldn't complete a banking or payment task because a digital or AI tool failed them — more than three times the 5% failure rate for everyone else.
The regulator found the same gap a month later
This would be a useful but self-contained piece of trade body research if the FCA hadn't, a few weeks later, published its own good and poor AI practice findings and landed on almost exactly the same fault line. Among the poor practice it identified in firms it tested this year: excluding vulnerable customers from AI journeys by default, and launching automated customer journeys without adequate testing. Among the good practice: firms that built live, evolving lists of vulnerability signals rather than relying on a static profile or a once-a-year review.
Two independent pieces of work — one from an industry body surveying its own members and one from the supervisor testing firms directly — arrived at the same conclusion within weeks of each other. When that happens, it's not a coincidence worth a passing mention. It's a pattern worth acting on.
Confidence is not evidence
Here's the distinction I'd want every board I advise to sit with. A confidence survey asks a firm's own staff whether they think something is working. An outcomes measurement asks whether it actually is, using data the firm can point to. The Payments Association report shows that the industry is answering the first question generously and the second question barely at all.
Thirty-five percent not tracking outcomes for vulnerable customers isn't a rounding error. It's over a third of the market running on belief. Consumer Duty doesn't accept belief as evidence. It asks firms to monitor outcomes, act on what the monitoring shows, and be able to demonstrate both. A board that approves an AI rollout on the strength of a 4.1 out of 5 internal confidence score, without a corresponding outcomes metric for vulnerable customers, has approved a feeling, not a control.
What boards should do about it
Three moves, none of them complicated and none of them requiring new technology.
- Add a specific vulnerable-customer AI metric to your board MI pack, separate from general AI performance metrics, before your next reporting cycle. If you can't produce one, that's your answer to whether the 93% confidence figure applies to you.
- Close the life-event gap deliberately. Support for older customers is relatively easy because the group is identifiable in advance. Support for someone who was fine last month and isn't this month requires a different kind of design — triggered by signals rather than static profiles — and it's currently the weaker half of most firms' provision.
- Stop treating "AI helps vulnerable customers" as a settled fact in internal reporting. Treat it as a claim that needs renewing with fresh evidence every reporting period, the same as any other Consumer Duty outcome.
The bottom line
If your board pack says AI is helping vulnerable customers but can't show the outcome data behind that claim, you don't have evidence. You have a hope, and hope has never been a defence in a Consumer Duty review.