InsightProofing · The AI-first policy-checking workbench

Every policy checked by agents. Every difference cited to the page.

Before a binder, policy or endorsement reaches your client, InsightProofing reads it against what was quoted, bound and expiring. The engines decide pass or fail; nine agents explain each difference and prepare the correction and the client summary for a person to sign.

Request a demo See how it works

8policies in the demo book, across 7 lines

55checks on one general liability renewal

9agents, every action capped at ask first

8proofing controls in Rules Studio

Figures from the illustrative demo book: invented brokerage, carriers, insureds and documents.

The InsightProofing Check Workspace for an illustrative general liability renewal, Ferncastle Logistics, LLC, at Proofed: 50 of 55 checks matched, 3 failures open (3 high), 2 to review, 6 renewal changes (4 narrower), total cost $48.5K with premium up 17.7%; below, a proofing control blocks delivery until the high-severity failure is corrected or waived — **Ferncastle Logistics, proofed.** 50 of 55 checks match; three failures are open, and a proofing control holds delivery until they are corrected or waived. Illustrative data; click any screen to see it full size.

Proof, not promises

Three screens from the demo book, as they are

An invented brokerage is checking eight invented policies. This is what InsightProofing put in front of the account team without being asked.

A failed check on an illustrative general liability policy: deductible $5,000 on the binder and $10,000 on the policy, with chips citing page 2 of each — **A failure cites both pages**The binder says $5,000; the policy says $10,000. Each value opens the page it was read from.

A decision the proofing checker prepared: raise three failed checks with the carrier — the deductible, a waiver endorsement missing from the policy and a pollution exclusion the binder doesn't have — with Accept, Reject and Ask — **Prepared, never sent**Three failures drafted as one correction request. Nothing reaches the carrier until a person accepts.

The proofing controls that fired on an illustrative check: a high-severity failure blocks delivery; a narrowed cover is a heads-up for the client summary — **A block that means it**An unresolved high-severity failure blocks delivery; narrowed cover must be named in the client summary.

The problem

A policy that reaches the client unchecked is the broker’s E&O exposure

The binder said a $5,000 deductible; the policy says $10,000. An endorsement the client paid for is missing. Each is a line on page four, found by someone reading two documents side by side — or by the client, after a loss.

E&O

Differences found after a loss

The time to catch a dropped endorsement or a narrowed sublimit is before delivery. After a claim, it is an errors-and-omissions matter.

Volume

Hundreds of policies at 1/1

Renewals bunch up. Checking each one line by line against three other documents does not scale with the season.

Wording

Every carrier labels it differently

“Per Occ Limit” on the quote, “Each Occurrence” on the binder, “Each Occurrence Limit” on the policy — the same field, three ways.

Follow-up

Corrections that go quiet

A correction asked of the carrier in an e-mail has no due date and no owner, and the policy sits undelivered while it waits.

The product

Read, proof, compare, correct, deliver — in one workbench

Every document, every field and every difference on one page per policy.

Every document read

Expiring, quote, binder, policy and endorsements, read page by page into canonical fields.

Learn more

One field, any wording

‘Per Occ Limit’ and ‘Each Occurrence’ are the same field; a person confirms close matches.

Learn more

Proofing

Policy against binder, binder against quote, field by field. The engine decides pass or fail.

Learn more

Renewal comparison

Every change from the expiring policy, read as broader, narrower or costlier.

Learn more

Towers & underlying

Layers that attach without gaps, and a schedule of underlying that matches what was bound.

Learn more

Carrier corrections

Failures raised as tracked requests with owners and due dates; overdue ones surface.

Learn more

Client delivery

A summary of what changed and what was corrected, signed by the account executive.

Learn more

Rules Studio

Proofing controls a practice leader can read, test on a real check and publish.

Learn more

Take the product tour

How a difference is found

From the documents to a signed delivery

Press Check and nine agents run in order. Every pass, fail and change comes from the engines; every step after reading is something you can open and check.

Read

Each document into the canonical model, every value with its page and the carrier’s label.

Proof

Policy against binder, binder against quote, plus term, premium arithmetic and the tower.

Compare

The renewal against the expiring: broader, narrower or costlier.

4AI

Agents reason

Nine agents explain each result, cite both pages and draft.

5AI

Prepared

Correction requests, label confirmations and the client summary — never sent.

A person decides

Raise, waive with a reason, or sign and deliver. A high-severity failure blocks delivery.

Model switched off or failing: the engines’ checks stand on their own, and every prepared decision says it reasoned by rules.

Every answer, waiver and sign-off is signed on the check’s decision record. The full architecture is on How it works.

The Check Workspace

One page per policy, with the agents’ work on top

The hero says what the policy is, where it stands and what blocks it; the stage strip moves it through Received, Read, Proofed, With carrier, Ready to deliver and Delivered. Seven zones hold the rest: Findings, Documents, Fields, Proofing, Comparison, Corrections and Delivery.

The Proofing zone on an illustrative general liability renewal: 55 checks, 50 match, 3 fail and 2 need review; the deductible failing at $5,000 on the binder and $10,000 on the policy, a waiver of subrogation endorsement on the binder but missing from the policy, a total pollution exclusion on the policy that the binder doesn't have, each with Raise with carrier, Accept and Waive, and both pages cited — **Proofing.** Each failure with the binder’s and the policy’s value side by side and a chip that opens the cited page.

The agents

Nine policy-checking agents, each with one job and a hard limit

Each agent keeps the engines’ results as its facts and reasons over them with the model. The actions they may prepare — asking for documents, confirming labels, raising corrections, sending them, delivering, signing off — are capped at ask first.

Document intake

Sorts what arrived — expiring, quote, binder, issued policy, endorsements — and drafts the request for what is missing.

Asks; never sends

Coverage extraction

Reads every document into the canonical coverage model for the line, each value with its page and the carrier’s own label.

Confirming a label is yours

Proofing checker

Policy against binder, binder against quote, field by field: pass, fail or needs review, cited on both documents.

Never decides a pass

Renewal comparison

The renewal against the expiring as quoted, bound and issued: broader, narrower, costs more, costs less.

Forms & endorsements

Endorsements dropped or added, edition changes, and restrictive forms nobody agreed to.

Tower & underlying

Layers that attach without gaps, and a schedule of underlying that matches what was bound.

Discrepancy drafter

Drafts the correction request to the carrier from the failed checks and follows each one up.

Capped at ask first

Client delivery writer

Writes the client summary — the coverage, what changed, what was corrected — once every failure is settled.

The AE signs and sends

Supervisor

Runs the proofing controls and prepares a practice leader’s sign-off when one fires.

Prepares; never decides

The autonomy dial

The level for an action is the lowest of the agent’s level for that line, the action’s hard cap and the brokerage’s ceiling. In InsightProofing the cap is L2 for every action.

Draft

The agent drafts; nothing is put in front of anyone else. Where every agent starts.

CeilingL2

Ask first

The agent prepares the decision; a person accepts, rejects or asks. Sending to a carrier or a client always stops here.

Act and tell

Locked. InsightProofing caps every action at L2 whatever the dial says.

Act

Locked. What reaches a carrier or a client is a person’s decision.

Agreement is still measured: an agent moves up a notch at 90% agreement over 50 answers and down at once below 70% over 20 — within the cap.

Two agents in the catalogue: the renewal comparison agent, which never invents a change, and the tower and underlying reviewer, which checks that layers attach without gaps — **One job, one limit**Each agent says what it reads, the one action it may prepare, and what it never invents.

The autonomy dial for the proofing checker: raise failed checks as corrections for the carrier, at L1 draft, with L2 ask first available and L3 and L4 locked — **The dial, with a ceiling**L1 draft and L2 ask first are open; act-and-tell and act are locked for every proofing action.

Product rationalization

One field, however the carrier words it

Proofing and comparison run on canonical fields, so a renamed label never reads as a field dropped and another added. The cascade: a label a person confirmed before, then the taxonomy’s synonyms, then a close match — used, but flagged for a person to confirm.

The general liability taxonomy: each canonical field — each occurrence limit, general aggregate, products-completed operations aggregate, personal and advertising injury, damage to premises rented, medical expense — with the carrier labels read as it — **One field, however the carrier words it**‘Per Occ Limit’, ‘EO_LIMIT’ and ‘Each Occurrence’ are all the each-occurrence limit.

The citation drawer: page 4 of an illustrative general liability policy with the cited line, a total pollution exclusion endorsement, marked — **Every citation opens the page**The cited line is marked; the document itself says it is illustrative.

Controls

Rules a practice leader can read, and a block that means it

Rules Studio

Proofing controls as sentences

Eight starter controls: an unresolved high-severity failure blocks delivery; a waived high-severity failure, an exclusion never agreed or a large account refers to a practice leader; narrowing, a premium jump, late delivery and an overdue carrier warn.

Decision record

Every answer signed

Each control that fires, each prepared decision and each person’s answer is written to the check’s decision record — the E&O file a broker can show.

Learning

Labels people taught

A label someone confirms is learned and used first next time; the taxonomy page lists them and lets a lead retire one.

Rules Studio: the proofing controls that fired most in the last 30 days, each in plain words — a high-severity check failed and unresolved blocks delivery; cover narrowed at renewal warns; a correction request overdue at the carrier; a policy not delivered 30 days after inception — **Eight controls ship with it**The busiest in the last 30 days, each readable as a sentence.

A follow-up the discrepancy drafter prepared: a correction request to the carrier past its due date with no correction yet — **Corrections that never go quiet**A request past its due date comes back as a prepared follow-up, with Accept, Reject or Ask.

The broker suite

One broker platform, from placement to the policy the client keeps

One identity, one agent framework and one decision record across the Insight broker products.

InInsightPlacementThe binder, quote and expiring policy from placement

InInsightXtractCarrier PDFs read into fields, with page and confidence

InInsightPulse 360The client’s company profile, one lookup away

InsightProofingPolicy checking: read, proof, compare, correct, deliver

OutThe clientThe checked policy and a summary of what changed, signed by the AE

OutInsightAdvocateThe policy as checked is the record a claims advocate argues from

AlongsideInsightRollupPolicies on an acquired book, proofed before they reach the client

Governance

Built so an E&O committee can say yes

The decision record

Every run, control hit, agent finding and person’s answer on the check’s record — what was decided, by whom, on what.

Grounded figures

An answer with a number that isn’t in its facts is thrown away and the engine’s finding stands. No fact-id markers reach a screen.

Scope

Account executives and managers see their own book; practice leaders and compliance see every check. Every row belongs to an organization.

Traces and prompts

A trace per agent step with model, tokens and cost; prompts versioned in InsightPrompts and overridable without a release.

Budgets

A daily budget per agent; over it, the fast model answers. Every step records what it cost.

Identity and audit

Sign-in through the Insight identity platform, roles, and an audit row for every change — including what an agent did and on whose behalf.

Deployment

Your cloud, your database, your model account

Runtime

One API, one UI

A Python API and a React workbench on the shared Insight kits — the same shell, workspace and agent pages as the rest of the suite.

Security & deployment

Data

PostgreSQL

One schema on the platform Postgres, versioned with Alembic. Documents go to your object store when one is configured — never local disk.

Models

Claude on Bedrock

Haiku-class models for agent steps and Sonnet for the copilot, through your own AWS account. With the model off the engines’ findings stand and say so.

From the blog

Worked examples

Policies followed from the binder to the client, on the illustrative book.

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

Use case

A $5M gap in an excess tower

Layer 3 attaches at $15M where layer 2 ends at $10M — losses in between are uninsured.

October 29, 2026 · 14 min read Read article

All posts

FAQ

What account teams ask first

Does an agent send anything to a carrier or a client on its own?

No. Every action is capped at L2, ask first, whatever the dial says. The discrepancy drafter prepares the correction request and the delivery writer prepares the client summary; an account manager sends the first and the account executive signs the second. In this build carrier e-mail runs as a dry run until a brokerage connects its mail.

Who decides whether a field passes?

The engines. Proofing is a deterministic comparison on canonical fields with the page and line on both sides. The model explains and drafts over those results; an answer that carries a number not in its facts is thrown away and the engine’s own finding stands.

Our carriers word everything differently. Does that break it?

That is what the coverage taxonomy is for. Each carrier label maps onto one canonical field: a label someone confirmed before, then the taxonomy’s synonyms, then a close match that is used but flagged for a person. What people confirm is learned for next time.

Do we have to use the whole broker suite?

No. InsightProofing stands on its own. With InsightXtract it reads PDFs through Xtract’s extraction contract; with InsightPlacement the quote and binder can come from the placement record; with InsightPulse 360 the client’s company profile is one lookup away.