Document Intelligence — Policies and Other Documents

A policy is sixty pages of which four matter, and which four depends on the loss. Estimatics reads the document, pulls the values that bear on the claim, and shows you the page each one came from.

Nothing it extracts is written to the job until you confirm it, field by field.

---

What you can upload

Eight document types, each analyzed differently:

Attachments are limited to 100 MB per file. A scanned policy that exceeds that has to be split.

---

Two routes, and they behave differently

Before the job exists — create from the policy. From Jobs → Create, choose to create from a document. Estimatics reads the policy and opens the creation form pre-filled with what it found. You review the form and save. This route is a head start on typing, and the form is your review step.

On a job that already exists — upload and confirm. Upload the document to the job, set its type, and analysis dispatches immediately. Here nothing touches the job record until you explicitly confirm, field by field.

Re-running an analysis while one is already in flight returns the one that's running rather than starting a second.

---

The five states

The distinction between Extracted and Confirmed is the whole design. Extraction proposes; you decide.

---

What every extracted field carries

Not just a value:

The source snippet is the point. You can take any extracted number back to the page it came from without opening the PDF and hunting.

---

What it pulls from a policy

Identity and dates — carrier, policy number, named insured, property address, effective and expiration dates, policy type, assigned adjuster, independent adjuster.

Four separate deductibles — all other perils, hurricane, roof, and sinkhole. They are extracted independently, because a single "deductible" field is how a hurricane deductible gets missed.

Coverages and limits — Dwelling (A), Other Structures (B), Personal Property (C), Loss of Use (D), water damage sub-limit, fungi/mold sub-limit, EMS mitigation limit, ordinance or law limit, roof surface payment limitation, and the loss settlement type (ACV vs RCV).

Coverage flags that decide claims — whether accidental discharge water damage is excluded, whether seepage or leakage over time is excluded, whether windstorm is excluded, and whether cosmetic roof damage is excluded.

Appraisal and dispute terms — whether appraisal is available, whether either party can invoke it or it takes mutual agreement, whether mediation is required first, and whether arbitration is mandatory.

Mitigation and repair conditions — whether a preferred contractor is required, whether there's a managed repair program, and the mitigation reporting deadline in hours.

Property detail, when the policy states it — year built, roof year, roof material, construction type. Only when it's explicitly in the document, typically the rating information page.

Endorsements — every endorsement with its form ID, title, effective date, relevance, and pages. Plus the controlling endorsement for water, wind, mold, and appraisal/mediation, so you know which form actually governs.

---

Contradictions

Policies contradict themselves. An endorsement modifies a limit stated on the declarations page; two endorsements overlap; a form and a schedule disagree.

Estimatics resolves these by one rule — endorsements override the base form — and where that isn't enough, it doesn't pick. It records the contradiction with its topic, both conflicting passages with their pages, and flags the field for review. You decide which one controls.

That's the behavior you want in a document that may be read back to you in appraisal.

---

Scanned documents

If text extraction looks incomplete, the analysis marks the document as scanned, sets ocr_required, and raises the review flags rather than filling gaps with plausible values. A scanned policy still extracts — it just tells you it's working from a harder source.

For best results: complete document, all pages, endorsements and schedules included. A declarations page on its own extracts a declarations page's worth of data.

---

Confirming — how values reach the job

Review the extraction, then confirm the specific fields you want applied. You send a list; only those fields write to the job. The response tells you exactly what was:

• Applied — written to the job record
• Skipped — you didn't select it, or there was nothing to write
• Unresolved — a value that couldn't be matched to a job field, usually a carrier name that isn't in your organizations yet

Confirming also runs a verification pass and returns any alerts. And a confirmed extraction can be re-confirmed — if you got a field wrong the first time, correct it and apply again.

---

Mold assessments

A mold assessment is handled differently, because its output isn't a set of policy fields — it's a condition per area.

Estimatics classifies the document (assessment, clearance, or other; only assessments are processed downstream) and extracts the IICRC S520 / NORMI condition for each area:

• Condition 3 — actual mold growth: visible, confirmed by sampling, or actively colonized material
• Condition 2 — settled spores or contamination without growth, including cross-contamination from a Condition 3 area
• Condition 1 — normal fungal ecology

Where the document states a condition number, that number is used. Where it only describes findings, the description is mapped to a condition and flagged for review.

You then confirm area by area, mapping each assessed area name to a room on the job. Every area you submit has to match one the document actually names — a fabricated area rejects the whole request rather than writing a single row. Confirmed conditions land on the room's loss context and project into the estimate.

---

Tips

• Upload before you analyze photos. Coverage context on the job makes the AI's coverage notes worth reading.
• Read the contradictions first. They're where the money is.
• Check the four deductibles separately. A hurricane deductible read as an AOP deductible is a five-figure error.
• Verify anything flagged for review. The flag exists because the document was ambiguous, not because the model was lazy.
• Confirm deliberately. Nothing lands on the job until you say so — which means nothing lands by accident, and nothing lands unless you looked.

---

---

Last updated: September 2026 · Questions? Email support@aiestimatics.com