The copilot opens beside any check and answers questions about it the way an experienced account manager would brief a colleague: the answer first, then why, then what to do next. It reads the check through six read-only lookups, cites the document and page, and never takes an action.

The six lookups

LookupWhat it returns
check_summaryThe check at a glance: insured, line, carrier, stage and what blocks the next one, counts, controls, what the agents prepared.
proofing_resultsEvery check that didn’t pass, with both documents’ values, pages and how a person resolved it.
renewal_changesEach change against the expiring, broader or narrower, at quote, binder and policy.
document_fieldsWhat was read into the coverage model, which documents agree, labels unmapped.
correctionsRequests to the carrier: what, status, sent, due, the response.
search_documentsPassages from the check’s documents matching a query, each with its page.
sequenceDiagram participant U as Account manager participant C as Copilot participant T as Lookups (read-only, your scope) participant G as Figure check U->>C: What doesn't match the binder, and what goes to the carrier first? C->>T: check_summary C->>T: proofing_results T-->>C: failures with pages C->>G: draft answer G-->>C: unsupported figures replaced C-->>U: answer with sources, suggested next steps as links

What to ask

Ask the copilotLookups it callsWhat the answer covers
“What doesn’t match the binder?”check_summary, proofing_resultsEach failure with both values and pages, by severity
“What narrowed at renewal?”renewal_changesEach narrowing at quote, binder and policy, and what the summary must say
“Where is the carrier on our corrections?”correctionsEach request’s status, due date and response; which are overdue
“Does the policy mention a waiver of subrogation?”search_documentsThe passages, with pages
“Can this go to the client today?”check_summaryWhat blocks delivery, in the controls’ words

The guardrails

  • Scope: the lookups run with the asker’s scope — the same as the screens.
  • Figures: every figure in the answer is checked against what the lookups returned; an unsupported one is replaced, never shown.
  • No actions: the copilot suggests; the workspace is where a person acts.
  • No model: with the copilot’s model off, it answers from the facts directly.
Proofing
The Proofing zone on an illustrative general liability renewal
Proofing: the deductible, a waiver endorsement missing from the policy and a pollution exclusion the binder doesn’t have fail; the named insured and an outstanding subjectivity need review.

For the account team

Ask in your own words; get the answer with the page to check it against.

For broker IT

insight_assist’s tool-calling answerer, Claude Sonnet by default, prompts overridable in InsightPrompts.

For practice leadership

The copilot can’t send, sign or waive — by construction.

Questions we get asked

Which model does it use?

Claude Sonnet 5.5 by default through Bedrock in your account; configurable.

Keep reading