I designed the Snapsheet AI layer that captures a claim's documents with OCR and vision, extracts the fields that matter, and turns them into suggestions an adjuster reviews in a tap. No black box.
Snapsheet is claims management software used across 170+ insurance customers, including 16 of the top 20 US property and casualty carriers. It already unifies a claim's documents, communications, notes and vendors into one view.
Snapsheet's stated goal for AI is to take it out of the black box. I lead design on the Claims and AI teams for that layer: the suggestions, summaries and document capture that sit on top of the claim, and the builder that lets the business shape them.
Every claim arrives as a pile of documents: the first notice of loss, police reports, repair estimates, photos. The platform gathers them in one place, but a person still had to read each one and key the facts into fields by hand.
Re-keying is slow, easy to get wrong, and it buries the work that actually needs a human: judging coverage, setting a fair reserve, deciding the next step. The documents were the bottleneck, and the adjuster was the OCR.
A single police report can hold the parties, citations, narrative and a dozen more facts an adjuster would otherwise transcribe by hand, one field at a time.
A representative claim, before this work:
All the data was in the file. None of it was structured.
The model was simple to say and hard to earn trust for. OCR and vision read the documents. Extraction turns them into structured fields. The system proposes those fields and the next action as a suggestion. A person reviews and confirms. Nothing reaches the claim on its own.
Extract a police report into 14 fields. Read damage photos for severity and a repair estimate. Compare a shop estimate line by line. Each agent is scoped to its parent document, so the adjuster always knows exactly what was read and where a value came from.
A summary, a fraud check, a liability estimate: each is a named, versioned, configurable AI Action. Designing the action as a first class object meant the same block could power the tray, a workflow step, or a button, without rebuilding it three times.
From inside a claim, the adjuster opens Snapsheet AI. They can run an agent, summarize the claim, draft an insured update, or ask a question in plain language. The summary and the suggestions arrive together, each carrying its sources and a single clear action.
The suggestions are where capture turns into closure. A coverage gap on the rental, a missing police report holding up reserve approval, a reserve sitting below the comparable median. Each one shows the policy section or claim data behind it, and offers one move: draft the heads-up, request the document, adjust the reserve. The adjuster stays in command of every decision.
A simple tray for the adjuster sits on a serious builder for the business. The suggestion is only as good as the system that produced it, so I designed that system to be inspectable, composable, and tunable.
A managed library of agents, Summarize Claim, Fraud Detector, Liability Estimator, Damage Vision, each with a status, a last modified date, and a model. Treating agents as governed objects is what lets a carrier trust them in production.
A claim is validated and triaged, then claim data, exposures, history and documents are extracted in parallel, analyzed, and merged into one reconciled summary. Building blocks for AI processing, flow control and outputs let a non engineer assemble this without writing code.
Every step chooses its foundational model (Claude, GPT-4o, Gemini), its own token, temperature and sampling settings, and its own system prompt. Designing this control surface let the team tune accuracy task by task, and stay model agnostic as the field keeps moving.
In claims, a wrong answer delivered confidently is worse than no answer. Three principles kept the AI accountable to the person responsible for the outcome.
Every suggestion cites its sources, a policy section, the FNOL, comparable claims. Every summary links back to the dates and documents behind it, so a surprising result invites a look instead of blind trust.
Suggestions stay pending until a person acts. Every response says it plainly at the bottom: AI generated, audit logged, verify before action. The human, not the model, moves the claim.
When a step times out or a dependency is missing, the run stops, names the failed steps, and returns partial results, rather than papering over a gap with a confident guess.
This was less about adding an AI feature and more about operationalizing AI responsibly inside work adjusters already do. The design made the model legible: every read is traceable to a document, and every action is a human's to confirm.
The goal was never to replace the adjuster's judgment. It was to hand them the facts already sorted, so judgment is all that's left to do.
Capture and extraction are only worth anything if the adjuster trusts them, and trust comes from showing sources and never acting alone. The most important pixels on the screen were not the answer. They were the citation and the confirm button.
The simple tray and the complex flow builder turned out to be the same idea at two altitudes: make the AI legible to the person who is accountable for the outcome. That is the through-line I carry into every AI surface I design now.