An adjuster picking up a file wants to ask it questions: is the reserve adequate, what is overdue, what did we tell the insured, how do claims like this usually close. The claims copilot answers those questions about the claim on screen, cites the parts of the file it read, and checks every figure it writes against them.

How it answers

sequenceDiagram participant A as Adjuster participant C as Copilot participant T as Read-only lookups A->>C: "Is the reserve adequate?" C->>C: Plan which parts of the claim to read C->>T: reserves_and_payments, decision_record T-->>C: Reserve by feature, basis, history, agent findings C->>C: Answer with numbered sources C->>C: Check every figure against what was read C-->>A: Answer, sources, suggested actions as links

The model plans, calls read-only lookups on the claim, and answers with numbered citations. If a figure in the answer is not in what it read, or the answer cites nothing, it gets one repair round; a figure still unsupported after that is replaced with “‹not in the record›”. It runs at most six rounds of lookups per answer.

The ten lookups

LookupWhat it returns
claim_overviewHeader facts, stage, assignee, clocks, narrative, next step, closure checklist
policy_and_coverageThe policy at the loss date: limits, aggregate left, trigger, retro date, forms; coverages and decisions
reserves_and_paymentsReserve now, suggested and ultimate with basis, by feature; authority; payments and holds; recoveries; settlements
notesNotes — privileged notes are never returned
diary_and_clocksOverdue items with days late; clocks with hours left or late
documentsWhat extraction read; privileged or confidential content withheld
decision_recordVerdict, rule hits, fraud score and indicators, what the agents prepared
litigation_recoveries_siuSuits, recoveries and SIU cases
parties_and_approvalsParties and approvals
similar_closed_claimsStatistics over the book; named examples only from the asker's own book
The claims copilot answering a question about a claim, with numbered sources, on demo data
The claims copilot on a general liability claim (demo data) answering “Is the reserve adequate?”: where the reserve stands and why it looks short, each figure with a numbered source.
The claims copilot in dark mode, on demo data
The copilot in dark mode: a real answer from the model against demo data.
The claims copilot on a phone, on demo data
The claims copilot on a phone, answering what is overdue on the claim.

The guardrails

  • Scope. The claim opens only if it is in the asker's book; an approver reaches it through an approval routed to their role. Every lookup re-checks scope and runs as the asker's organisation.
  • Reads, never acts. Suggested actions — confirm coverage, run Analyze, review what the agents prepared — are links into the workspace. The copilot cannot take them.
  • Privilege. Privileged notes and privileged or confidential document content are never returned.
  • Fallback. With no model configured or a model failure, it falls back to a plain facts summary and says so. A kill switch turns the model off.
  • Traces and prompts. A Lense span per answer; the prompt from InsightPrompts when released there, else the local text.

The copilot is built on the platform's tool-calling assistant, so InsightUW and InsightCede can use the same machinery over their own records.

Keep reading

How InsightCW works Rules Studio for claims Use cases, end to end Auto collision with subrogation