Skip to main content

Insurers use AI, and insure it. Both need evidence.

Pricing, underwriting and claims triage decide about policyholders. At the same time, insurers are being asked to cover the losses AI causes for their clients. In both cases the question is whether a person really answers for the decision.

AIC.Karoo Mutual (Demo)

Decision log: claims triage

Sample data

  • CLM-20931Held for review
  • CLM-20932No review needed
  • CLM-20933No review needed
  • CLM-20934Held for review

For

  • Underwriting and pricing
  • Claims and fraud
  • Conduct and compliance
  • Cyber and liability underwriters

Where AI decides, and what goes wrong

  • Decides about: Applicants and policyholders

    Pricing and accepting cover

    Premiums that track a proxy for a protected group rather than the risk.

  • Decides about: Claimants

    Triaging and settling claims

    Valid claims slowed or short-paid because a model scored them as unusual.

  • Decides about: Claimants

    Flagging claims as possible fraud

    An honest claimant treated as a suspect with no person reviewing the flag.

  • Decides about: Policyholders

    Deciding which policies to renew

    Cover quietly withdrawn from people a model rates as unprofitable.

A claim is flagged as possible fraud

Fair treatment in claims is measured by what happens to the claimant, not by how the flag was raised.

  1. 1

    The flag is raised

    The triage model marks a vehicle claim for investigation and holds it, rather than declining it.

  2. 2

    A handler reviews it

    The claims handler who answers for the model sees the reason it gave and the evidence behind the flag.

  3. 3

    The outcome is recorded

    The handler clears the claim, with a reason. The override joins the log against the model.

  4. 4

    The pattern is visible

    If one model's flags are overturned again and again, it shows on the risk register before it shows in complaints.

Rules that already apply

A summary, not legal advice about your organisation.

POPIA section 71

Applies to automated decisions about policyholders and claimants that profile them. Read the guide

What it asks of youSafeguards including human intervention and a chance to make representations.

Policyholder Protection Rules

The fair treatment rules for long-term and short-term insurance, including how claims are handled, whoever or whatever handles them.

What it asks of youOutcomes that are fair to the policyholder, which a model's results are judged by as much as a person's.

King V

Boards are expected to oversee how technology and AI are governed, including accountability for outputs. Read the guide

What it asks of youEvidence for the board that AI is governed, not only a policy saying it should be.

EU AI Act

Lists AI used for risk assessment and pricing of individuals in life and health insurance as high-risk, from 2 December 2027, where its output is used in the EU. Read the guide

What it asks of youHuman oversight, logging and a risk management system for the systems in scope.

Questions

Can AIC evidence be used in underwriting?

Yes, with the client's permission. A client can share a read-only extract or a sealed evidence pack. AIC gives no rating or recommendation; pricing and acceptance remain the insurer's decisions.

Does AIC replace our own model governance?

No. It records who answers for each decision a model makes about a policyholder, and keeps the evidence. Your model governance decides whether the model is sound.

Is fraud detection exempt from the automated decision rules?

Not automatically. Where a flag profiles a claimant and has a substantial effect on them, the safeguards apply. Holding flagged claims for a person, rather than declining them, is the usual design.