Why a Successful Claims AI Pilot can Still Fail in Production

This piece is by Adam Gaca, a specialist at Future Processing

Future Processing’s June 2026 AI Scaling Paradox report examines why many enterprise AI programmes stall between pilot and production. Its central argument is that the decisive gap is usually not model capability, but whether the organisation has validated the business case, data, operating process and governance before committing to build.

The insurance sector shows the paradox particularly clearly. McKinsey’s 2025 State of AI identified only a small group — around 6% of organisations — that combined significant value from AI with more than 5% EBIT attribution. Yet the potential upside in insurance is already measurable: WTW reports that analytics leaders among large carriers operate with combined ratios around six points below the market, while EXL cites operating-cost reductions of up to 21% in AI-enabled insurance operations.

Claims triage is a good example of why pilots can mislead. A pilot may prove that a model can classify or prioritise claims on a curated dataset. Production requires a different level of evidence: reliable data lineage, integration with the claims platform, clear exception thresholds, a named decision owner, human review, monitoring, and an audit trail showing how a decision was reached and corrected.

In other words, a pilot tests whether the model can produce an answer. Production tests whether the insurer can rely on that answer.

The appropriate level of automation also depends on the consequence of error. For high-volume, low-value claims, such as straightforward motor claims below $1,000, straight-through processing can materially reduce settlement time; market leaders report reductions of up to 80%. In specialty claims, where a single loss may run into tens or hundreds of millions, the more credible use of AI is decision support: surfacing relevant information, identifying anomalies and improving reserving or coverage analysis, while keeping accountable human judgement in the process.

The UK market adds another problem. DSIT’s 2025 AI Adoption Research found that 71% of UK firms had not identified a concrete AI use case. That suggests board-level ambition is moving faster than use-case definition.

Adam Gaca, Managing Director UK&I and Vice President of Innovation at Future Processing, said:

“UK insurers rarely lack AI ideas. What is usually missing is a baseline, a named business owner and a clear definition of what better looks like. If those are absent, a pilot may show that the technology works without proving that the business should deploy it.”

The practical lesson is to separate two decisions before delivery starts: is the use case economically justified, and is the organisation ready to scale it? A credible assessment must also allow for a NO-GO outcome. Stopping a weak use case early is not a failure of innovation; it is disciplined investment and risk control.