ML Underwriting Automation

About the Project

Risk selection and pricing intelligence that gives commercial underwriters back their judgement

Sixty submissions arrived on Monday morning. By Friday, the underwriting team had reviewed every single one , and written eighteen. The other forty-two had been read carefully, researched, priced in the underwriter's head, and ultimately declined. None of that effort was wasted exactly, but it was effort spent on outcomes that, in hindsight, had been predetermined from the start.

“This wasn't an efficiency problem,” the Head of Underwriting at our client , a mid-market commercial insurer writing SME property, liability, marine, and trade credit risk , told us. “It was a quality problem. My best underwriters were spending forty percent of their week on work with a predetermined outcome. The cases that deserved their full analytical capability were getting whatever time was left after that forty percent was gone.” That conversation became the starting point for the platform Kreate Technologies would go on to build.

The Challenge

The insurer had invested in real actuarial talent and pricing tools, but the bridge between those models and the decisions underwriters made on individual risks remained informal , communicated through rate tables and tribal knowledge rather than anything systematic or auditable. Accumulation risk surfaced only in quarterly reviews, long after the relevant decisions had already been bound.

Challenges

60 weekly submissions reviewed to find just 18 worth writing
Pricing and risk-selection models communicated informally, with no audit trail
Accumulation risk surfaced only in retrospective quarterly reviews
No structured feedback loop between claims outcomes and underwriting decisions

Our Solution

Kreate Technologies built a fi ve-layer ML platform on a single founding principle, agreed with the underwriting team in the first week of the engagement: the system's job is to give underwriters better information at the right moment , never to make the decision for them. Submissions are triaged the instant they arrive, so the underwriters' attention lands exactly where their judgement is genuinely needed.

Implementation Phases

Phase 1 , Discovery:

eight weeks mapping risk logic and five years of loss data by class.

Phase 2 , Model Build:

loss-propensity and dynamic pricing models trained per line of business.

Phase 3 , Triage Engine:

appetite-matching rules calibrated for genuinely safe straight-through risk.

Phase 4 , Platform & Integration:

underwriting workbench connected to broker portal and core systems.

Phase 5 , AMC & Retraining:

quarterly performance reporting and continuous model recalibration.

Results

Business Outcomes

A year in, the underwriting floor looked different , not because the people had changed, but because what reached their desks had. Routine submissions moved themselves through the system. The complex, judgement-heavy cases arrived already briefed, already contextualised, and ready for the kind of thinking only an experienced underwriter can do.

Operational Impact

45% — Less Time on Routine Work
3 Hrs — Quote Turnaround (vs. 3.5 Days)
2.8% — Loss Ratio Improvement
28% — Submissions via STP
38% increase in broker-facing submission capacity
Portfolio concentration profile materially closer to target at year-end

Key Takeaways

The goal of automation is attention, not headcount , freeing experts for what needs them

Conservative triage thresholds build underwriter trust in straight-through decisions

Portfolio context at the point of decision prevents accumulation surprises

Insurers already own the data ML needs , most simply aren't using it systematically