The starting position
The insurer had everything that should make AI enablement easy. Decades of historical claims data. A sophisticated actuarial function. Strong vendor relationships across the policy admin and claims platforms. An executive team that understood that AI was going to redesign their industry and wanted to be on the right side of the curve.
What they had built was less impressive than what they had in principle. The claims operation ran on a workflow designed in 2009 with successive layers of automation grafted on top. A chatbot deployment had handled 14% of inbound contacts but the overall cost-to-serve had not moved meaningfully. Claims cycle times on routine property claims were running at 28 days against a market average around 14. The actuarial function had built models that the operations function couldn't operationalise in real time. The customer NPS on closed claims was below the firm's competitive cohort.
The COO sponsored the engagement. The Chief Underwriting Officer was sceptical at the start. The Chief Claims Officer was the one who actually had to make it land.
What we found in the diagnostic
The first six weeks confirmed what the COO suspected: the claims function was in classic augmentation mode. Multiple AI tools were in production. None of them changed the workflow shape. Claims handlers were doing the same work they had always done, just with marginally faster individual steps. The cost-per-claim had improved by single-digit percentages over two years; the structural ceiling had been reached.
The deeper finding was about decision rights and feedback loops. Claims handlers overrode the system's triage recommendations approximately 28% of the time — a rate so high it suggested either the model was wrong or the handlers had context the model couldn't see. We sampled the overrides. Both were true: the model was missing critical context (loss circumstances, policyholder relationship history, broker patterns) and the handlers were applying judgment the model hadn't been trained on. The override information was being captured in free-text fields and never fed back into the model. The flywheel was broken.
We also found something that surprised us: the actuarial function was a much stronger ally than we expected. They had been quietly frustrated for years that their work didn't reach the operational layer in real time. They became one of our most important internal sponsors.
What we redesigned
The blueprint phase produced a target-state operating model for property claims built around the five enablement pillars:
Production function: Claims rebuilt in BPMN 2.0 with first-notice-of-loss triage as the structural pivot. Routine claims (homeowner, low-complexity, no fraud signals, no liability dispute) flow through the system end-to-end with human review at material thresholds and full audit trails. Complex claims, fraud edge cases, and customer-facing situations escalate to claims handlers with rich context already assembled. The handler role shifted from "process the claim" to "judge the case the system flagged."
Data layer: A new claim wide-row built around the policyholder, the loss event, the broker relationship, the policy history, and the relevant external signals (geographic patterns, broker submission patterns, fraud indicators). Captured at first notice of loss in the form the workflow needed, not reconstructed from policy admin overnight.
Decision rights: A formal decision rights matrix mapping which claims the system handles end-to-end, which escalate at which threshold, and who owns the outcome. Material claims thresholds were preserved (Lloyd's regulatory requirements and reinsurance triggers don't move) but the routine workflow simplified dramatically.
Embedded governance: A second-line risk specialist embedded with the team, working alongside the actuarial function on model validation. The model risk file was built to PRA SS1/23 standard with the actuarial function as a co-author. The PRA conversation moved to constructive dialogue early.
Feedback flywheel: Override decisions captured as structured signal — pre-defined override categories with optional free text — and routed back into a feedback curation pipeline. The model retrains monthly. The override rate is monitored against a target band of 6–10%.
How the delivery ran
Phase 2 ran for 12 weeks with the embedded second-line risk specialist and weekly working sessions with the actuarial function. Phase 3 (activation) ran for 10 months.
The hardest part was Phase 3, month 3 to month 6, when the new workflow was running in parallel with the legacy workflow. Claims handlers were sceptical. Some genuinely struggled with the role change (handling fewer cases but each more complex). Three handlers chose to leave during this period. The COO was clear with the team and clear with us: this work would not be slowed down to avoid that pain. It was the right call but it was uncomfortable.
By month 6 the new workflow was processing 35% of claim volume. By month 9 it was at 75% and the legacy workflow was being decommissioned. By month 10 the team had been retrained, the role design had stabilised, and the metrics had moved.
What the outcomes look like
Ten months after activation:
- 58% of claims handled end-to-end by the redesigned workflow with no human touch beyond the at-threshold review. Up from approximately 12% under the legacy model.
- Average cycle time on routine property claims dropped from 28 days to 11. That number understates the customer experience improvement because the variation also dropped — customers no longer experience the worst-case three-month outliers that drove the legacy NPS down.
- Customer NPS on closed claims rose by 22 points in the rolling 90-day window after the new workflow stabilised. The team is being honest about which part of that is the workflow vs. other factors, but the directional move is unmistakeable.
- 100% of material decisions reconstructable on demand with full lineage from policy admin systems through to the model that made the call. The PRA conversation is now constructive rather than defensive.
- The claims handler team is about 30% smaller than it was 18 months ago, but each role is more demanding and each handler is dealing with cases their judgment actually changes. Career paths have been redesigned.
- The actuarial function is now a primary user of the operational data rather than working from quarterly extracts. The feedback loop between operations and actuarial is continuous.
Why this was harder than it sounds
Three structural challenges defined this engagement:
The actuarial function was the critical enabler — and was almost an obstacle. Actuaries are model-fluent and rigorous, which makes them excellent partners on AI work. They are also justifiably suspicious of "AI" that does not respect their model risk discipline. The right move was to engage actuarial leadership in the first conversation and treat their model risk framework as the foundation rather than the obstacle. We did. It worked. Most engagements that get this wrong fail.
The claims handler role change was the hardest single thing. Telling experienced claims handlers that their job was changing — fewer cases, more judgment per case, smaller team — required organisational courage from the executive team. Some handlers left. Others thrived in the new role. The change management piece was substantial and not a side workstream.
Long-tail decisions matter. Claims decisions made today have consequences that play out over years (reserving accuracy, fraud uplift, customer lifetime value, broker retention). The data flywheel is slower to spin up than in fraud or KYC. The temptation to over-optimise for short-term cycle-time signal had to be resisted. We did this by setting outcome quality SLOs alongside throughput SLOs from day one.
What changed for the insurer
The technical outcomes are real and they matter. The deeper change is the operating pattern: the insurer now has a working example of an AI-native workflow in one priority value stream, and the data layer, governance, and role design are being reused as the foundation for redesigning underwriting next. The actuarial function is integrated into operational decision-making in real time rather than via quarterly cycles. The competitive position has structurally improved.
Two years from now, this insurer will be running at materially different cost-to-serve and customer experience levels than competitors who have not done the work. That is the compounding case. It only works if you start.