This article is by Micha Kiener, CTO, Flowable
Insurers want AI to do more of their repeatable work. Recent research found more than four in five leaders are on board with this, but just 6% would let a general-purpose model make high-consequence underwriting and claims decisions by itself.
We’re moving from AI that assists decisions to AI that increasingly participates in them. That’s a fundamentally different challenge for insurers. Underwriting and claims processes were designed around people making judgement calls, with technology providing support. They were not designed for AI systems helping to make decisions that affect risk, payouts and compliance.
Even the most capable model may have no idea what an insurer considers acceptable risk, which internal rules override general policy or who’s allowed to sign off a borderline case. It can only judge what it’s given, and in a lot of insurers that context is scattered through different systems and inboxes.
When these calls go wrong, the fallout can mean rejected claims, overlooked fraud, questions from the regulator and money out the door. And it’s the underwriters, claims teams and compliance leads who have to explain it.
Where automation gets stuck
A commercial property submission shows how this plays out. An AI model reads the broker’s documents, then pulls five years of loss runs out of scanned PDFs. It works out the loss ratio. Eligibility rules confirm the risk fits the insurer’s appetite. Then third-party data comes back showing the buildings in a wildfire zone. The submission needs a referral.
This is where rigid workflows alone fall short. Standard automation can struggle when a submission deviates from the expected path because new information can change the course of a submission, and the process needs to accommodate that exception without losing the context around it. Too often, the issue gets pushed out of the automated system into manual email threads and shared folders, leaving the underwriter to piece together what has happened. The result is reduced visibility and fragmented audit trails for underwriting and compliance teams.
Everything in one place
That’s where case management comes in because not every submission follows a straightforward path. Some need further investigation, follow-ups or human review before a decision can be made. So, a case can remain open while missing information is gathered, with the underwriter’s task reactivated once it’s available. This allows the process to accommodate exceptions without losing visibility of the work already done.
When AI works inside the case, it starts from the same picture as the underwriter. When it runs into something that it can’t resolve, say, a construction year that doesn’t make sense or a risk score sitting right on the approval line, the case sends it to somebody who has the authority to decide. They get the full history along with it, so nobody needs to go hunting through inboxes.
Back to that wildfire submission. Inside the case, the underwriter sees the loss runs the model extracted, the eligibility result and the third-party report that triggered the referral. Say they need a fire inspection report from the broker. The case requests it, pauses the service level clock and picks back up when the document arrives. Nothing has to be rebuilt from email.
Claims are no different. A claim might look perfectly routine on the form. But the case could show two other claims on that property in the past year, or a note from the handler who dealt with the last one. A fraud model that only sees the form won’t know any of that, and it’ll still hand back a confident score.
Proving the decision
There’s a compliance side to all this too. Under Consumer Duty, UK insurers have to evidence good outcomes for customers. So, when someone challenges an automated decision, you’ve got to point to the logic and the model version that made it. Even an ordinary pricing update on 1 June raises the question of which version priced a quote the day before.
When every step runs inside the case, that record is already there. Without it, teams end up reconstructing the audit trail each time a regulator asks.
That same record helps teams notice when something starts to drift. A model that suddenly begins to decline more applications from one region than it did last month might have picked up a bias. Or perhaps the risk there changed. With every decision logged in the case, the team can work out which before it carries on.
Where people come in
The same research showed that 86% think decisions with real consequences should stay with people. That’s the right instinct. People should step in at the points where judgement really matters, with the whole case in hand. AI takes care of the rest.
What now?
The priority is getting insurance context into the process, where AI and people can both use it. The models are already capable enough for a lot of this work. With the full case behind them, they’ll make better calls, and the people who answer for those calls will have good reason to trust them.
Be the first to comment