For Insurance

Protecting Underwriting Judgement While Using AI in Insurance

Your best underwriters built their judgement through thousands of case decisions. When Guidewire AI or SAS AI automates these decisions, that institutional knowledge disappears into a black box. The risk is not just regulatory exposure when you cannot explain why a claim was denied, but the loss of the human discretion that once caught the actuarially correct decision that was ethically wrong.

These are suggestions. Your situation will differ. Use what is useful.

Download printable PDF

Keep Underwriting Judgement in the Loop

AI tools like SAS AI and Azure AI can score risk quickly, but underwriters who have spent years learning to spot context that models miss should still see and approve high-value or ambiguous cases. Your senior underwriters have pattern recognition that goes beyond what training data shows. When you remove them from the decision, you lose their ability to catch the edge cases where the model fails or produces a fair but cruel outcome.

›Set decision thresholds so that any underwriting decision worth more than a set premium goes to a person for final approval
›Route cases flagged as unusual or borderline to your most experienced underwriters, not to a queue for faster processing
›Document why underwriters override model recommendations, then use that data to understand where your models need refinement

Audit for Bias That Models Hide

When you move underwriting to Guidewire AI or IBM Watson, bias that was once visible in individual underwriter decisions becomes systemic and scaled. A model trained on historical underwriting data will replicate historical unfairness at machine speed. You need to test whether your AI tool denies coverage to certain groups at higher rates than others, even if the model does not use protected characteristics as inputs.

›Run monthly fairness audits that compare approval rates and premiums across demographic groups for the same risk profile
›Ask your SAS or Azure AI vendor for their bias testing methodology and what protected-class proxies they tested for
›Keep a cohort of manually underwritten cases in parallel to your AI system so you can compare outcomes and catch drift

Demand Explainability Before Implementation

Regulators now expect you to explain underwriting and claims decisions. Models from Watson or Azure AI that use deep learning or ensemble methods often cannot tell you why they said no. Before you deploy any AI tool for underwriting or claims triage, test whether you can produce a plain-language reason for each decision that would satisfy a regulator or a customer complaint. If you cannot, you have a compliance problem on day one.

›Request SHAP values or similar explainability outputs from your AI vendor and test whether they make sense to your underwriters
›Build a rule-based approval flow as your baseline, then only add AI to cases where the model can articulate its reasoning
›Store the model explanation alongside every AI-driven underwriting or claims decision so you can defend it later

Protect Claims Discretion in Fast-Moving Cases

Claims assessment tools powered by Guidewire AI or ChatGPT can triage thousands of claims and suggest approvals at speed, but claims adjusters have always used discretion to find fairness in ambiguous situations. A catastrophe claim with missing documentation, a dispute over coverage, or a case with partial liability all require human judgment. Speed of processing is not the measure of good claims handling if you are denying fair claims because the model did not know how to handle complexity.

›Route claims with coverage ambiguity or high severity to your most experienced adjusters, not to the fastest-processing path
›When AI recommends a claims decision, require the adjuster to confirm they agree before payment or denial goes out
›Track which AI recommendations adjusters override and why, then use that to retrain your model or to change your system rules

Build Feedback Loops Between AI and Actuaries

Your actuarial team has built loss models based on years of claims data and underwriting outcomes. When you deploy AI for underwriting or fraud detection, those models will change behaviour, and that changes the data that feeds your loss models. If your actuarial team does not see the decisions the AI is making, they cannot track whether experience is aligning with projections. Disconnect between model predictions and actual claims creates hidden risk.

›Give your actuarial team monthly reports on approval rates, claim denials, and fraud flags generated by your AI tools
›Compare AI-underwritten cohorts against your loss projections to spot early whether the AI is under or over-selecting risk
›Create a monthly forum where your actuaries can challenge AI decisions that do not align with underwriting intent or expected loss ratios

Key principles

1.Underwriters and claims adjusters should approve high-stakes or ambiguous decisions even when AI recommends an outcome.
2.Fairness audits must run monthly and compare approval rates across demographic groups to catch systemic bias that individual decisions hide.
3.Every AI underwriting and claims decision must come with a plain-language explanation that would satisfy a regulator or customer.
4.Actuarial teams must see the decisions AI is making so they can verify that experience matches loss projections and catch hidden risk.
5.When AI speeds up claims or underwriting, protect the cases that need human discretion instead of moving them faster.

Key reminders

Before deploying SAS AI or Guidewire AI, run a parallel manual underwriting cohort for three months so you can compare AI outcomes against your best underwriters.
Ask your AI vendor to show you cases where the model was most confident but the outcome was wrong, so you understand its failure modes.
Create a simple rule that any underwriting decision that contradicts an underwriter's recommendation must be documented with the model reasoning attached.
Set up a quarterly review where your compliance team, actuarial team, and underwriting leadership examine a sample of AI decisions to spot patterns of fairness concern.
When you implement fraud detection through Watson or Azure AI, keep a small team of manual investigators doing the same work so you can validate whether the AI is catching real fraud or just noise.