← Back to prototypes
Case study

AI Claim Review Workspace

AI output placed inside a controlled operational workflow where a human remains responsible for the decision.

The problem

Claims assessors need to review multiple sources of information — notes, uploaded documents, and policy context — before deciding what information is missing and what should happen next. The current process is slow and inconsistent, and a raw AI summary alone does not give assessors enough to act on responsibly.

Target user

Claims assessor. Key decision: approve the AI-suggested next step or redirect the case (request documents, specialist review, or escalate).

Product goal

Help the assessor review an AI-drafted case summary, verify it against evidence, correct it, and record a decision with an auditable trail — without treating the AI output as a black box.

User flow

  1. 1Open the claims queue and find a claim (search, filter by status, sorted by priority).
  2. 2Select a claim to open the case workspace.
  3. 3Read the AI-generated case summary, missing information, and risk flags.
  4. 4Open the evidence drawer to see which sources support each point.
  5. 5Edit the draft summary and add case notes where needed.
  6. 6Choose one action: request documents, specialist review, approve next step, or escalate.
  7. 7Confirm the decision and see it recorded on the activity timeline.

Design decisions

  • Evidence is always reachable from the AI summary — every claim the AI makes has an associated source excerpt, so the user can verify rather than trust.
  • The summary is editable in place; an "edited by assessor" badge makes the correction visible — this is the human-review contract.
  • Actions are explicit and mutually exclusive — one clear primary decision per review reduces ambiguity and keeps the audit trail clean.
  • Confidence is labelled conservatively — "Needs Human Review" signals that the AI output requires a person to decide.

AI interaction design

Input is a mock service call to ClaimReviewService.reviewClaim(id). Output is a structured AiClaimReview object — summary, missing information, risk flags, suggested actions, and evidence references. Uncertainty is handled by the "Needs Human Review" label. The human review flow includes editing, evidence inspection, decision selection, and a confirmation dialog. Every review outcome is appended to the activity timeline with the acting user and timestamp.

Technical approach

Next.js App Router with React client components for the interactive workspace. TypeScript types for Claim and AiClaimReview. Components are organised in an atomic structure (atoms, molecules, organisms, templates) so they are reusable and named by business purpose. The mock API layer is a ClaimReviewService adapter with simulated latency and an explicit error path so loading, empty, and retry states are real. Styling uses Tailwind CSS v4 with a small design-token system.

Tradeoffs

  • The AI service is mocked and deterministic so the full product flow can be validated before choosing or integrating a model provider.
  • There is no persistence: the queue and timeline live in component state for the session.
  • Only a small, representative set of claims is included to keep the demo fast to load and easy to narrate.

Success measures

  • Review completion time per claim
  • Rate of edited AI outputs (how often the human corrects the draft)
  • Action completion rate and reviewer confidence
  • Number of escalations and reasons

Next steps

Replace the mock adapter with a real, server-side AI service with privacy and compliance controls; add claim-level history persistence; add role-based views for team lead and operations manager; and validate the workflow with real assessors.