A claim decided step by step, on the page behind every fact.
Six decision engines take a claim from first notice to settlement and recovery, each one citing the page it relied on and sealing a signed receipt. A claim is paid without a person only where that has been earned — and one override takes it away.
Seven checks, and when each one stops for a person.
Documents are read by Gothink AIR with the claims pack and assembled into one claim, with conflicts between documents kept and cited. Then each step decides.
| Step | Outcomes | Holds for a person when |
|---|---|---|
| File check | FILE_READY · AWAITING_DOCUMENTS | A required field is missing or held |
| Coverage | COVERED · QUESTION · REVIEW | Anything but COVERED |
| Fraud | REFER_SIU · CLEAR | REFER_SIU — and an SIU referral is opened |
| Triage | STP · DESK · FIELD · SIU | SIU |
| Severity | RESERVE · REVIEW | No calibrated interval, or one too wide to reserve on |
| Settlement | STP_ELIGIBLE · REVIEW | REVIEW, or the autonomy gate says no |
| Subrogation | PURSUE · REVIEW · NONE | Never holds — it opens a recovery referral |
Paid without a person only after the reviews say so.
Autonomy is granted per claim type, from reviewed outcomes — the same autonomy and gates model the rest of the platform uses.
You choose the claim types
Only claim types you list, up to an amount limit you set per type, can ever be paid unattended. With no configuration, nothing is.
598 clean reviews first
At the default error budget of 0.5%, a claim type needs 598 clean human reviews before a single payment goes out unattended. The console shows how many are still needed.
One override revokes it
Unattended payments are sampled for QA. An override revokes autonomy for that claim type at once and opens a payment-correction referral.
The claim file as it arrives.
The claims pack (v1.0.0) covers FNOL packs, declarations, policy wording, endorsements, police reports, repair estimates, invoices, medical records, demand letters, proofs of loss and adjuster notes.
Reads the clauses themselves
An exclusion that mentions the cause of loss becomes a QUESTION citing the clause, its page and the sentence; an endorsement that may change the cover is shown with it. A cause named only in an exception (“however, we will pay for ensuing loss…”) does not exclude.
A cited chronology
Encounters, codes as written, charges, gaps in treatment and treatment before the loss — feeding severity for bodily injury and a soft fraud signal. The summary is treated as PHI and shown only to authorised users.
Documents find their claim
A new claim number opens a claim; late documents are attached and flagged, not re-decided; documents without a claim number wait on the policy number until the claim exists.
One console for the desk, one trail for the examiner.
What each claim is waiting for
The decision trail with the page behind every value, conflicts between documents, and a review form: accepted, modified, overridden or escalated, by a named reviewer.
Per claim type
Reviews, errors, the error bound against its budget, clean reviews still needed and the unattended amount limit.
Stored before anything leaves
Every payment and referral is recorded as pending before it is sent, and marked sent or failed afterwards.
Claims in the examiner’s sample
Claims share the Witness receipt ledger: a Claims tab, case files in the evidence pack, and sampled claims an examiner can open with their full trail. About Witness →
Drafted from your own records
Witness drafts the Annex IV technical documentation and the Article 27 fundamental-rights impact assessment, each split into what the records show and what you must still write. They are drafts, not a conformity assessment.
NAIC readiness
Claims decisions carry the case-file evidence the NAIC AI report asks vendors for. The evidence map →
Connected carefully, measured before it is trusted.
| Piece | What it does | Status |
|---|---|---|
| Guidewire ClaimCenter | Payment, referral as an activity, and note — with a connector check that prints a dry run of every request and sends nothing | Template — paths are placeholders until confirmed with the carrier; refused until verified |
| Duck Creek Claims | Payment, task and note, checked the same way | Template — as above |
| Subrogation pilot kit | Maps your closed-claims export, runs the recovery screen and writes a report and an opportunities list. Nothing leaves the machine. | Built |
| Claims console | Queue, autonomy, outbox, open a claim, subrogation backtest | On the cloud service — with demo data |
What is not built yet.
A verified core-system tenant
ClaimCenter and Duck Creek connections are verified against a carrier’s own tenant, not ours. Until then they are templates.
Litigation propensity
A triage model for litigation risk needs a carrier’s closed claims with outcomes. We will not ship one trained on anything else.
An accuracy figure
The demo uses synthetic claims. It shows how the engines behave, not how accurate they are on your book.
Start with your closed claims.
Run the subrogation screen on a closed-claims export inside your environment, then decide whether to go further. We are looking for carriers to verify the core-system connection with us.