My assertion: Most insurers have pointed their AI programmes at the expense ratio. I suggest they should be focussed at the loss ratio.
Two conversations in the last fortnight got me here. On 17 September I hosted a breakfast roundtable in London on the disruption of AI in underwriting and claims, with COOs, Heads of Customer Journey, Heads of Data and AI transformation leads from several of the biggest UK insurers and brokers. Last week I was on an INSEAD insurance panel and a completely different group brought me to a similar conclusion…
Ask an insurer what they’re doing with AI and the answer is claims. FNOL automation, document capture, triage, photo damage estimates, fraud scoring, settlement recommendations. It’s easy to justify because claims has clean KPIs: cost per file, cycle time, satisfaction, leakage. You automate a step, a number moves, you write the business case.
Claims data is structured and sits in one place, the metrics are simple, and a claims model gets through internal governance inside a year. They picked it because the challenge was understood, the benefits measurable, and it would get funded.
But it’s now the wrong order of priorities, and the business case was never as strong as it looked.
The claims savings are mostly on paper
Many claims business cases show time taken out of a process, not cost taken out of the organisation. Cycle times come down, touchpoints come down, straight-through processing goes up, and the same people are still on the payroll. Minutes saved per file only turn into pounds saved per year if headcount actually reduces or gets redeployed onto work that makes money. If it doesn’t, you’ve tidied up the workflow and left the economics where they were.
There’s also a regulatory problem with a programme built purely on cost. Consumer Duty gets very real when a triage model is deciding which claims a human looks at and which go straight to decline. Several of the customer-journey focussed people at the breakfast roundtable made the point that fast and fair are different things. Plain-English explanations, complaint-risk analytics and a proper audit trail aren’t things to add later. They’re what keeps the cost case standing when the FCA asks to see it.
Underwriting has barely moved
Ten years of predictive models and pricing tweaks, and almost nobody has rebuilt the actual workflow. Broker submissions still arrive as emails and spreadsheets, loss runs still get re-keyed, and underwriters still spend their first hour of the day on data entry.
That’s odd, because one point of improvement in risk selection does more for the combined ratio than several points of claims efficiency. We’ve put the AI into the smaller number because it had the better story attached.
Why underwriting doesn’t get the focus
Three reasons come up every time, and none of them is really about technology.
The first is identity. Claims is a process. Underwriting is a craft. A claims director will happily talk about straight-through rates. A chief underwriting officer hears “automated decisioning” and asks, quite reasonably, whose name is on the line when the model gets it wrong. That’s a governance question and it has governance answers: explainability, decision logs, human sign-off above agreed thresholds. But it has to be settled before anyone lets an agent anywhere near a binding decision.
The second is data. Underwriting data is broker emails, survey reports, third-party feeds, pricing tools and a policy admin system that predates all of it. Most underwriting AI programmes are plumbing projects with an AI label on them, and boards don’t get excited about plumbing.
The third is the feedback loop, which is broken almost everywhere. The most useful thing an underwriter could know is what happened to the risks they wrote: which claims came in, why, and what that says about the assumptions at inception. In most insurers that information goes through two departments, three systems and a quarterly reserving meeting before it gets anywhere near them. AI on either side of that wall is working with half the picture.
Make the next 12 months about decisions, not use cases
Every insurer already has a slide with forty use cases on it. What’s missing is a handful of decisions.
Move budget from claims to underwriting and be open about it. If more than two-thirds of your AI spend is in claims and operations, you’re over-indexed on the smaller number. Claims isn’t done, but the marginal return has moved. The next claims automation buys a fraction of a point of expense ratio. The first serious underwriting automation, in a line with reasonably structured submissions, buys underwriter capacity, faster quotes, a clearer view of appetite, and gets you into the loss-ratio conversation.
Pick one line and do the whole workflow. An intake agent that extracts data perfectly and hands it to someone to re-key into a pricing tool has changed nothing. Intake, enrichment, appetite triage and decision support for standard risks, with underwriters’ time moved onto the complex ones, is the size of change that shows up in the numbers.
Make the CUO the sponsor, not the CIO. If the person who owns the loss ratio doesn’t own the programme, it becomes a technology project that underwriters put up with rather than a change in how they work. Put explainability and human-override thresholds in the brief on day one. That’s the price of their sponsorship.
Treat the claims-to-underwriting feedback loop as a product. It’s the one investment that helps both sides, and in most firms nobody owns it.
Where is your programme concentrated?
The carriers that treat underwriting AI as a governance problem to solve, rather than a craft to protect, will price more accurately, quote faster and pick better risks. The ones that keep automating the back end of claims while the front of the business stays manual will end up very efficiently paying for other people’s better underwriting.
So the question is simple. Where is your AI programme concentrated, and does that match where your combined ratio is actually decided? I’d be interested to hear how others are answering it.
