AI for insurance: 42% faster claim decisions without taking the adjuster out of the loop.
Adjusters do not need a faster decision engine. They need the file assembled, the coverage checked and the evidence shown before they open it.
Nordwind processes 3,000 claims a month. Each one crossed five systems before an adjuster saw it: the intake portal, a document store, the policy system, a spreadsheet of exceptions and an email thread. Adjusters spent most of the day assembling context rather than deciding anything, and the backlog was growing 18% a year.
The standing plan was to hire more adjusters, which is the reason the team called us instead. Twelve weeks later, claim decisions were 42% faster, each adjuster was closing 2.1 times as many claims a day, and the backlog was 63% smaller by week eight.
Nothing about that outcome required an autonomous system. The agent reads the file, checks coverage against the policy and drafts a recommendation with citations back to source documents. A person accepts, edits or rejects it in one click, which is exactly where the judgement should sit.
Insurance at a glance
- Usual first project
- One claim type, end to end, in the adjusters' tool
- Time to production
- 12 weeks including hardening and training
- Where it runs
- Your cloud tenancy, zero-retention model endpoints
- Price band
- $50K to $150K
- Cases on this page
- Nordwind Insurance · Veyra · MedArc
The bottleneck is assembly, not adjudication
Time a claim from first notice to decision and the deciding part is minutes. The rest is finding the policy version in force on the loss date, locating three documents, reading an adjuster's note from a previous claim, and checking whether an exclusion applies.
Nordwind's adjusters spent about forty minutes per claim on that assembly across five systems. The workbench collapsed it into one screen that is already populated when they open it, which is where the 42% came from. We did not make anyone decide faster.
Citations are what make a recommendation usable
An insurance recommendation without a source is an opinion, and adjusters are right to ignore opinions. Every output in the Nordwind workbench carries citations to the exact clause and document page it came from, so verifying a recommendation takes seconds and disagreeing with one is a normal, logged action.
That design choice is also the compliance answer. When a regulator or a reinsurer asks why a claim was paid, the audit trail contains the evidence considered, the clause applied, the recommendation, the human decision and the time between them.
It is why adoption reached 95% in week one rather than month three. Adjusters trusted the tool because it showed its work, not because anyone mandated it.
The queues after claims
First notice of loss triage, subrogation, complaints handling and renewals all share the shape that makes this work: scattered context, a written policy, repeatable judgement, and a person who should stay accountable for the outcome.
Subrogation is usually the second project because the payback is direct: recovery opportunities missed because nobody had time to read the file are money already on the books. Nordwind scoped theirs at $40,000 to $60,000 over eight weeks, reusing the same agent layer and workbench.
Regulators, model risk and the questions your audit committee will ask
Which model, where does the data go, how do you know it is accurate, what happens when it changes, and who is accountable. These are reasonable questions and they have build-time answers.
The model layer sits behind one interface, so a change of provider is configuration plus an evaluation run. Nordwind's evaluation suite is 1,200 historical claims with known outcomes and it runs on every change. Below-threshold cases route to a human by default rather than guessing, and every claim file records what the agent saw and what the human did with it.
Everything runs inside the insurer's own cloud account. No claim data leaves the tenancy, and model calls use zero-retention endpoints. The security review happens in the audit week, before code exists to argue about.
Adoption is a design problem
Most claims AI projects fail at the adjuster's desk, not in the model. If the tool is a second window that duplicates half the work, it gets used for a fortnight and quietly abandoned.
So the workbench is the adjusters' primary screen, built to look like the tool they already knew, with the queue, the file, the recommendation and the actions in one place. Ten years of interface work is the actual reason the number in week one was 95% and not 30%.
Budget, timeline and an honest fit test
A production claims workbench with an agent layer, evaluations and training is a $50,000 to $150,000 project over ten to twelve weeks with a team of about four. A single narrower queue, such as FNOL triage, sits lower.
The fit test is whether a competent adjuster can write down the rule they apply. Where the rule is written and the evidence is in documents, this works. Where the decision is genuinely discretionary and every case is unlike the last, it does not, and the audit week is designed to find that out before you commit a budget.
What changed, measured
Nordwind Insurance, first 90 days in production against the previous quarter, from their claims system.
Insurance work that shipped
Where to go next
Contract and clause review
Policy wording checked against a playbook with deviations flagged and sourced.
Questions we get from insurance teams
No. It reads the file, checks coverage and drafts a recommendation with citations. An adjuster accepts, edits or rejects it. Cases below the confidence threshold are routed to a human without a recommendation at all, and no payment is issued by a machine.
With an evaluation suite of historical claims with known outcomes, run on every change. Nordwind's is 1,200 claims. Reports show accuracy by claim type and the rate of human overrides, which is usually the number a committee cares about most.
In your own cloud account. No claim data leaves your tenancy, model calls use zero-retention endpoints, and access is scoped to the workbench service. The security review is part of the audit week rather than a step at the end.
The model layer sits behind one interface, so switching is a config change plus an evaluation run. During an outage the workbench still assembles the file and shows the history; it simply presents no recommendation rather than a stale one.
Nordwind reached 95% adoption in week one, because the workbench replaced their existing screens rather than sitting beside them. Training is two sessions and a written guide, and the two-week hardening phase exists to fix what the first week of real use exposes.
The same pattern fits any queue where evidence is scattered and the rule is written, and submission triage is the common starting point. We would still scope it separately, because underwriting appetite changes far more often than claims policy and the evaluation set has to keep up.
Have a insurance problem shaped like this?
Thirty minutes on a call answers fit and gives you a rough estimate. The audit week that follows produces a written plan and a fixed price, and you keep the plan either way.