This piece is by Andy Male, Strategy Consulting Lead for FS&I, Valtech
Mobile began with an app for customers. Cloud started with infrastructure. Digital was all about the front end. In each case you could name a first move, build a case around it, and get on with things.
AI offers no such starting point. It spreads across customer experience, core systems, data platforms and operations at once, and with roughly equal plausibility in each. That is why so many insurers have spent heavily and can still point to so little. The capability is no longer the constraint. What separates firms that are getting returns from those still running pilots is the discipline to choose where to apply it.
Too many good ideas at once
Sit in on the prioritisation exercise at any large insurer, and the difficulty becomes obvious. Claims can make a case for AI. So can underwriting, pricing, distribution, fraud, servicing and finance. Every case is credible on its own terms, argued by people who know their function better than anyone else in the room, and there is no shared unit of measurement that lets you rank them against each other.
So they don’t get ranked. Boards are presented with a dozen options, all of which seem reasonable, with no basis for choosing among them. The path of least resistance is to fund several and see which survive.
That is a hedge, and it produces what hedges produce: pilots multiply, two functions build the same document-ingestion capability without knowing about each other, and budget spreads thin across a wide surface. Little gets past the demonstration stage. The organisation looks busy while the P&L stays quiet.
There is a subtler cost, and it points the wrong way. AI is bought to simplify the business and speed it up, a sprawl of pilots does the reverse. Each experiment is another integration to maintain, another model to monitor, another control surface to own, a layer of complexity added by the very technology meant to remove it. That governance layer is permanent, it grows with every new test, and it almost never appears in the business case that funded the pilot. So, firms weigh build cost against benefit and leave out the standing cost of running the estate, the number that actually decides whether any of it was worth doing.
Nobody in particular made a bad decision. Each approval was defensible on its own. The failure happens in the aggregate, and it is exactly the kind of failure no single meeting is set up to catch.
Rank the options on more than one axis
The firms that break the deadlock do it by giving the choice a structure. Not a single value-versus-effort score, which is what produced the unrankable list in the first place, but a small set of dimensions applied to every candidate: the value genuinely at stake, the readiness of the data underneath it, the true end-to-end complexity of changing it, and the regulatory exposure it carries.
Scored that way, a dozen plausible ideas resolve into a short list you can defend. The point is not the scoring model itself. It is that the firm now has a stated reason for what it funds and, just as importantly, for what it declines.
The scrutiny is catching up faster than the returns
This would be a manageable problem if nobody were watching, but they are. In January, the Treasury Select Committee published its report on AI in financial services, warning that the current approach from regulators and the Treasury risks exposing consumers and the wider system to serious harm. It asked the FCA to publish practical guidance by the end of this year on how the Consumer Duty applies to AI and on the assurance senior managers must provide under the SMCR for harm caused by AI use. Evidence to the Committee put AI adoption at more than 75% of UK financial services firms, with insurers among the heaviest users.
The direction is clear. The burden of proof sits with the firm. You will be asked what your AI systems do, which decisions they influence, who owns them, and how you know they are producing fair outcomes. A scattered portfolio of pilots cannot answer that. Thirty experiments across six functions, with no one owning the whole picture, end the same way every time: someone in a supervisory review asks who is accountable, and there is no good answer.
Why the use-case list keeps failing
The instinct in most firms is to reach for a use-case list: rank by effort and value, pick the top few, start. But a use-case list treats AI as a series of contained bets, whereas the value of an AI deployment lies mostly in the layer beneath it, the data it draws on, the systems it writes back to, and the operational process it changes.
This is why the firms that get it right map the whole service before they scope the build. A use-case list cannot show you the downstream handoff that quietly cancels the gain; a service blueprint can, which is why end-to-end impact belongs on the scorecard rather than in a post-mortem.
What the firms getting value actually do
The pattern among firms making measurable progress is narrower ambition, properly applied. They choose one or two value pools and go deep, not one or two use cases, but one or two parts of the business where AI could change the economics. Then they fix everything standing in the way: data, process, controls, the lot.
They establish the baseline before anything gets built. Whatever the measure, cost per claim, cycle time, loss ratio, they know the starting number, because without it there is nothing to show when someone asks whether it worked. They design governance into the build rather than bolting it on after a supervisory letter arrives. And they decline.
This is the part firms find hardest, and it matters most. A credible AI programme is defined at least as much by the good ideas it turns down as by the ones it funds.
Be the first to comment