Forecast Autonomous Claims Handling
Design an anonymous, iterative expert elicitation that improves reasons and preserves material disagreement.
26 minutes
Scenario
An insurer needs a planning range for the share of low-value claims that could be handled autonomously in three years without exceeding its error and complaint thresholds. Internal data is incomplete and expertise is distributed across claims, fraud, compliance, data science and customer advocacy.
- Your role
- Independent Delphi facilitator
- Method
- Delphi Method
Evidence pack
Claims data
42% of last year's claims were below EUR 500; 18% of those required manual document clarification.
Pilot
A rules-plus-model pilot processed 28% autonomously with 1.9% rework in a non-representative two-month sample.
Risk threshold
The board requires confirmed overpayment below 0.5% and no increase in upheld complaints.
Panel
Candidates include two claims leaders, a fraud analyst, a compliance officer, two data scientists and a customer advocate.
Dependency
New digital identity rules are expected but the implementation date and permitted checks are uncertain.
Baseline disagreement
Pre-interview estimates range from 20% to 75%; high estimates assume near-perfect document extraction.
Time
Three asynchronous rounds can be completed before the investment review.
Decision use
The forecast will size a staged capability programme, not authorise autonomous production decisions.
Constraints
- Individual responses remain anonymous to other panel members.
- The facilitator must report distribution, reasons, assumptions and confidence—not only a mean.
- Stopping rules are defined before round one and cannot require consensus.
Case steps
Work through each prompt using the evidence pack. Answers and rubric weights stay protected in the interactive flow.
Write the forecasting question with population, horizon, conditions, outcome definition and confidence request.
Select the panel and map which perspective each member contributes. Identify one missing perspective or dependency owner.
Design three rounds, including the statistical summary and anonymised reason themes returned after each round.
Define stopping rules and a final report that preserves the material low and high scenarios.