Data Academy · Tutorial 2 of 10 · Insurance

Governed AI in insurance

Insurers are putting agents to work in claims, underwriting and servicing, where knowledge-heavy work meets high volumes. The carriers that scale are the ones that redesign whole journeys and govern every settlement, referral and decline, because a claim decision has to be fair, explainable and defensible to the customer and the regulator.

01 What is changing

Where AI in insurance is heading

From summarising files to agents that triage a claim or submission, check cover and raise the next step in the claims or underwriting system.
From isolated tasks such as document reading to end-to-end redesign of a claims or underwriting journey.
From specialist roles doing data entry to handlers and underwriters supervising agents and spending their time on judgement and relationships.
From efficiency as the only case to growth, retention and technical profitability, with governance treated as part of the design.

02 Use cases

Three use cases on the governed path

Each use case runs the same path: a business question, governed context, a deterministic rule, specialist agents, a policy check, an action and a record. How the path works →

Illustrative: names and figures are invented to show the flow.

Use case 1

Motor claim fast-track

Simple motor claims wait in the same queue as complex ones, which frustrates customers and ties up handlers. Settling clear, covered claims within authority gives faster settlement and frees handlers for injury and disputed claims.

“Which of today’s new motor claims can we authorise straight away, and which need a handler?”Asked by a motor claims manager

Context: A first notification of loss in the claims system reports a rear-end collision; the third party has admitted liability and photos are attached.
The policy administration system shows comprehensive cover in force, premium paid and a £250 excess.
The approved repairer network returns an estimate of £1,640, leaving £1,390 payable after the excess.
Claims history shows no claims in 4 years, no injury is reported and the fraud score is below the referral threshold.
Rule: If the policy is in force, the peril is covered, the estimate is under £2,500, no injury is reported and the fraud score is below the referral threshold, authorise the repair.
Decision: Authorise repair Severity: Low
Agents: Intake agent Extracts the incident details and damage description from the notification and photos.
Coverage agent Checks the policy wording, endorsements and excess in the policy administration system.
Customer update agent Drafts the confirmation to the customer with the repairer booking and the excess due.
Policy: Within the fast-track team’s delegated claims authority. Any claim with injury, a disputed liability or a referral flag goes to a handler. Customer communication must meet the FCA Consumer Duty expectation of clear, timely information.
Action: The repair authorisation is released automatically in the claims system within policy, with the £250 excess recorded and recovery from the third-party insurer opened.
Data: Claims system: notifications, reserves and paymentsPolicy administration: cover, endorsements and excessRepairer network estimatesClaims historyFraud scores

Use case 2

Commercial property submission triage

Underwriters spend much of their day keying broker submissions and chasing information, so good risks wait while poor ones are worked. Triage against appetite gives faster quotes on risks the insurer wants and clearer declines on those it does not.

“Which broker submissions this week fit our appetite, and which need a senior underwriter before we quote?”Asked by a commercial property underwriting manager

Context: A broker submission in the underwriting workbench for Marlow Fine Foods Ltd seeks cover for a food processing site with a £18m sum insured.
The survey report shows a 1970s building with expanded polystyrene composite panels and no sprinkler system.
The underwriting guidelines limit risks with composite panels and no sprinklers to a £10m sum insured without senior referral.
The exposure management system places the site in a moderate flood zone, with 2 prior flood claims in the loss history supplied by the broker.
Rule: If composite panels are present, there are no sprinklers and the sum insured exceeds £10m, refer to a senior underwriter; no quote may be issued automatically.
Decision: Refer to senior underwriter Severity: High
Agents: Submission reader agent Extracts occupancy, construction, protection and loss history from the broker’s documents into the workbench.
Appetite agent Compares the risk with the written underwriting guidelines and shows which clauses it falls outside.
Risk improvement agent Drafts questions and possible conditions for the broker, such as fire risk improvements and flood defences.
Policy: Only an underwriter whose written authority covers the sum insured may quote or bind. Agents may draft but not price or issue terms. Accumulation in the flood zone is checked against the exposure limits that feed the insurer’s Solvency II capital calculation.
Action: A referral is raised in the underwriting workbench and held for senior underwriter approval; no quote is sent to the broker until it is approved.
Data: Underwriting workbench: submissions and referralsSurvey and risk engineering reportsUnderwriting guidelines and authoritiesExposure management and catastrophe dataBroker-supplied loss history

Use case 3

Claims fraud referral

Most claims are genuine, so fraud checks must not slow honest customers down, but paying organised or opportunistic fraud raises premiums for everyone. Connecting the signals across claims and policies stops leakage without accusing anyone prematurely.

“Is there anything about this contents claim that should stop payment until our investigators have looked at it?”Asked by a home claims team leader

Context: A home contents claim for theft of electronics worth £7,800 is logged in the claims system.
The policy administration system shows the policy incepted 41 days ago, and the contents sum insured was raised from £20,000 to £45,000 12 days before the loss.
The knowledge graph links the claimant’s phone number to an earlier claim under a different name.
The police reference supplied has not yet been confirmed with the force.
Rule: If a claim is made within 60 days of inception or within 30 days of a mid-term sum insured increase, and a contact detail is shared with another claim, hold payment and refer to the special investigations unit.
Decision: Hold payment pending investigation Severity: High
Agents: Link analysis agent Shows the shared contact detail, the related claim and the timeline of policy changes.
Document check agent Checks receipts and photos for consistency and lists what still needs verification.
Holding letter agent Drafts a neutral update to the customer that the claim is being reviewed, without any allegation.
Policy: Agents may not decline a claim or allege fraud; only the special investigations unit can. Processing for fraud prevention must have a lawful basis under GDPR, and the customer must still receive timely, fair updates under the FCA Consumer Duty.
Action: Payment is blocked pending the investigation check in the claims system and a referral is raised to the special investigations unit; the claim stays open.
Data: Claims system: claims, payments and notesPolicy administration: inception and mid-term changesKnowledge graph of people, contacts and claimsFraud database and police reference checksDocument and image store

03 The foundation

What the agents need to understand

Core entities in the ontology

PolicyholderPolicyCoverageRiskSubmissionClaimLoss EventReserveBrokerThird Party

Systems they come from

Policy administration system (PAS): policies, cover, endorsements, premiums and mid-term changes
Claims management: notifications, reserves, payments and recoveries
Underwriting workbench: broker submissions, referrals, quotes and authorities
Exposure management: locations, accumulations and catastrophe exposure
Billing and collections: instalments, arrears and refunds
Supplier networks: repairers, loss adjusters and estimates

04 Guardrails

The controls that let it scale

Rules decide, agents explain

Settlements, referrals and payment holds come from written authorities and guidelines, while agents read, compare and draft.

No automatic declines

Declines and fraud findings always rest with a named handler or investigator, never with an agent.

Fair outcomes

Automatic decisions are tested for consistent treatment of customers in line with the FCA Consumer Duty and its fair value expectations.

Personal and special category data

Health and criminal offence data in claims is processed only where a lawful condition under GDPR applies, and agents see only what the case needs.

Capital-aware underwriting

Underwriting actions respect exposure limits that feed the Solvency II capital view and the reserving that feeds IFRS 17 reporting.

05 Rollout

From the first use case to many

1
Choose one journey
Start with a high-volume journey with clear authorities, such as motor fast-track claims or property submission triage.
2
Build the insurance model
Connect policies, claims, submissions and parties into one governed model, so cover checks use the same facts the handler would.
3
Shadow the handlers
Run agents alongside handlers and underwriters, compare outcomes and tune the rules before any automatic release.
4
Release within delegated authority
Allow automatic settlement only inside written authority limits, with every other outcome held or referred.
5
Extend across the value chain
Reuse the model and policies for servicing, renewals and recoveries rather than building each journey separately.

06 What to measure

Outcomes, not activity

Claims cycle timeStraight-through settlement rateQuote turnaround timeClaims leakageComplaint rateDecisions overturned on review

07 Pitfalls

What usually goes wrong

Tools without redesign. Adding document readers to an unchanged claims flow leaves the queues in place; redesign the journey around triage and authority.
Under-investing in people. Handlers and underwriters who were not involved in the design do not trust the outputs; build the roles and incentives alongside the agents.
Unclear authority. If delegated limits are not written down, agents cannot be governed; codify authorities before allowing any automatic action.
Legacy data left as it is. Agents reading inconsistent policy and claims records give inconsistent answers; reconcile the core entities first.

08 Diagnostics

Questions to ask your team

1
Are our claims and underwriting authorities written precisely enough for a rule to apply them?
2
Which outcomes may an agent release automatically, and which must always be held or referred?
3
Can we explain any settlement, decline or payment hold to the customer and the regulator from the record alone?
4
Do policy, claims and broker data describe the same customer and the same risk in the same way?

09 Keep going

Frequently asked

Can an agent decline a claim?

No. In the governed path agents can recommend, hold or refer, but declines and fraud findings rest with a named person. This protects the customer and keeps the insurer’s decisions defensible.

How is this different from straight-through processing we already have?

Traditional straight-through processing handles fully structured data. Agents add the ability to read notifications, surveys and broker documents, while the decision still comes from the same written rules and authorities.

Where should an insurer start?

With one journey where authorities are already clear and volumes are high, such as motor fast-track claims. The policy, claim and party model built there carries over to underwriting and servicing.