Your AI cannot read a claim file.
We make it AI-ready.
Submissions and claim files into clean, checked, page-traced data — and with local-only models, no outside AI provider sees a page.
The documents we read
Start with the one that costs you most today. That is where a pilot starts.
Broker submissions
Cover emails, applications, supplementals, expiring policies (insurance_submissions).
Statements of values
TIV by location, construction, occupancy, protection class.
Loss runs
Carrier PDFs, each laid out differently.
Not built
Inspection and engineering reports, regulator notices, bordereaux and treaty schedules: no pack reads these yet.
Policy wordings & endorsements
Declarations pages, policy wordings and endorsements, read for coverage checks (claims).
Life applications (draft)
Applications, paramedical exams, lab reports, APS, MIB and Rx reports (us_life_applications, a draft: every field goes to a person).
FNOL packs
Portal forms, broker emails and adjuster notes.
Police reports & proofs of loss
Police reports and sworn proofs of loss, read into the claim file (claims).
Estimates & invoices
Line items, totals and duplicates.
Medical records
Medical records read into the claim file (claims). CMS-1500 and UB-04 health claims are in the draft us_health_claims pack.
Demand letters deadline risk
Time-limit demands, with their dates extracted on arrival (claims).
Life, health, workers’ comp & surety claims (draft)
Claimant statements, death certificates, health claim forms, first reports of injury and payment bond claim notices (draft packs: every field goes to a person).
Documents in. Checked records out.
One pipeline, inside your environment. Usable data at step three, trusted data at step four.
Money and coverage wait for a person
Reading a document is not the same as trusting the data in it.
- A named adjuster or underwriter signs off on anything that moves money or coverage
- Low-confidence fields go to a person, page and box open
- Thresholds are set by your claims and risk leaders, not by us
- Every adjuster correction retrains the agent
An examiner will ask how your AI decided
The evidence is produced as a by-product of doing the work.
What they ask for
- A written AI Systems Program
- Sample case files
- Testing and bias evidence
- Oversight of vendor models
- Documents on request
How the layer answers
- Lineage on every published value
- Thresholds as versioned configuration
- A named person on every gate
- Ground truth your experts signed
- With local-only models, no page reaches an outside AI provider
What you get
One source of clean data every team, model and agent can use.
What changes
not yet measured- Risks scored before they are opened
- Claims arrive decision-ready
- Medical records and demand letters read into the claim file
- An audit trail for the examiner
How much runs without a person
not yet measured — generate your own numberYour straight-through rate is measured on your own documents during the pilot; we publish no figure until it is measured. None of these packs is yet measured on a customer’s labelled documents, and a draft pack sends every field to a person.
Start with one document type
Three gates: Prove (days 0–30), Expand (31–60), Industrialize (61–90), stopping at day 90 if targets are missed.
Choose the documents
A month of auto claim packs, a quarter of submissions, or one line's medical files.
Map and agree
Fields you need, confidence thresholds, today's manual effort measured.
See the result
A working pipeline, scored against answers your experts approved.
See what Gothink reads from your claim files.
Send a sample. Get back scored, page-cited fields — before you buy anything.