InsightProofing’s nine agents run in order when someone presses Check. Each one has one job, reads the engines’ results as numbered facts, reasons over them with the model, and prepares at most one kind of decision for a person. None of them decides a pass or a fail, and none of them sends anything.

flowchart LR subgraph Read["Read the file"] I["Document intake"] --> X["Coverage extraction"] end subgraph Check["Proof and compare"] P["Proofing checker"] --> C["Renewal comparison"] --> F["Forms & endorsements"] --> T["Tower & underlying"] end subgraph Resolve["Correct and deliver"] D["Discrepancy drafter"] --> W["Client delivery writer"] --> S["Supervisor"] end X --> P T --> D

The nine, and what each may prepare

AgentIts jobMay prepareNever invents
Document intakeSorts what arrived; finds what is missing.Ask the carrier or client for missing documentsA document nobody sent
Coverage extractionReads every document into the canonical model.Confirm carrier labels onto canonical fieldsA value
Proofing checkerPolicy vs binder, binder vs quote, derived checks.Raise failed checks as correctionsA pass or a fail
Renewal comparisonRenewal against expiring: broader, narrower, cost.—A change
Forms & endorsementsDropped, added, edition changes, restrictive forms.—A form
Tower & underlyingLayers stack; underlying matches.——
Discrepancy drafterWrites and follows up correction requests.Send the correction requestA correction
Client delivery writerWrites the client summary once settled.Deliver the policy to the clientA coverage term
SupervisorRuns the proofing controls.Practice leader sign-off—
AI & Agents · Catalogue
The agent catalogue: nine policy-checking agents in three steps
The agent catalogue: nine agents in three steps, the action each may prepare and what each never invents.

One agent step, from the inside

sequenceDiagram participant E as Engines participant A as Agent participant M as Model participant G as Grounding check participant P as Person E-->>A: results as numbered facts [F1]…[Fn] A->>M: role prompt + facts + fixed answer shape M-->>A: summary, findings citing fact ids, open questions A->>G: every number in the facts? alt grounded G-->>A: keep the model's reading else a number not in the facts G-->>A: discard, the engine's finding stands end A-->>P: prepared decision at L1/L2

Every step is recorded: the facts, the model used, tokens, cost and seconds, and what was prepared. With the model switched off, each agent’s own engine findings stand, and every prepared decision says From the engines.

AI & Agents · Run trace
The proofing checker's recent runs and one run's trace
The proofing checker’s runs and one trace on an illustrative umbrella check: the engines’ facts, the model step with tokens and cost, and the decision prepared.
AI & Agents · Activity
Agent activity on the illustrative book
Agent activity: every agent step on the book, what it prepared and what is waiting on someone.

For the account team

Each card in Findings tells you which agent raised it and whether the engines or the model wrote it.

For broker IT

Agents run on the shared insight_agents pipeline with a registry, an autonomy service and a run store per product.

For practice leadership

Nine narrow agents are easier to govern than one broad one: each has one action, one cap and one record.

Questions we get asked

Do the agents run on their own when a document arrives?

They run when someone presses Check. The catalogue shows each as When asked.

Keep reading

How it works

Engines decide, models explain

If the number is not in the facts, the sentence does not ship.

October 29, 2026 · 10 min read Read article
Use case

A property renewal with a halved flood sublimit

Not a proofing failure — a renewal change the client must hear about in plain words.

October 29, 2026 · 15 min read Read article
Use case

A missing waiver and an exclusion nobody agreed

The binder lists the waiver the client's contracts require. The policy doesn't.

October 29, 2026 · 15 min read Read article