Claims Decision Desk: Help

1. Introduction

Claims Decision Desk is an AI-assisted decision-support tool for claims and remediation teams. It helps a handler move from a lodged claim to a defensible, evidence-based recommendation: it finds the right policy clauses, checks the evidence, calculates the amount exactly, and routes the decision to the person with the authority to approve it.

It is recommend-only. The tool never approves, denies or pays a claim. Authorised people make every consequential decision, and every step is recorded.

It is part of a connected set of tools: ARA assesses the risk of AI use cases, and CALIBRE Governance Fabric registers and governs them. Claims Decision Desk is registered as initiative GOV-CLM-02.

This is a portfolio prototype on synthetic data. Its rules are simplified and illustrative, not legal or financial advice. Do not enter real customer information.

2. Features

Feature	What it does
Policy retrieval	Finds relevant clauses (keyword plus meaning-based search), shows relevance and uses the policy version in force at the claim date
Evidence checks	Flags OK, WARN or BLOCK conditions such as late notification, quote above benchmark, missing identity verification or passcode disclosure
Exact calculation	Computes the amount in code (not by the AI), showing each step: excess, betterment, refund, interest, recoveries
Delegated authority	Works out who may approve the amount and escalates when it exceeds the user's limit
Guardrails	Each action returns Permit, Escalate or Deny. Denying a claim is always a human-only decision
Hallucination guard	AI-drafted text is rejected if it cites a clause that was not retrieved or states an amount that was not calculated
Decision brief	A structured summary: recommendation, clauses relied on, checks, routing and status
Audit trail	Each recorded decision is hash-chained so tampering is detectable
Swappable AI model	The language model can be changed (for example Anthropic, OpenAI, Hugging Face) without changing the rest of the tool
Governance link	Connected to CALIBRE Governance Fabric and ARA for registration, assessment and runtime control

Three example scenarios are included: an insurance motor claim, a bank fee remediation, and a bank unauthorised transfer.

3. How it adds value to banks and financial services

Faster, more consistent decisions. Handlers spend less time searching policy wording and doing arithmetic, and similar cases are treated alike.
Fewer errors and less leakage. Calculations are deterministic and tested, and quotes are checked against benchmarks.
Traceability for regulators and auditors. Every recommendation shows its clauses, figures, policy version, approver and timestamp.
Control over AI risk. The AI drafts but never decides; outputs are verified; actions are limited by an authority envelope; and the tool is governed through ARA and CALIBRE.
Better customer outcomes. Faster responses, clear reasons, early identification of vulnerability or urgency, and SLA tracking (for example a 15 business day remediation window).
Reusable across sectors. The same engine works for general insurance claims, bank remediation, fraud and scam disputes, and other regulated decisions. New products, jurisdictions or rules are added as packs.

4. How to use this tool

Choose a scenario from the tabs at the top (motor claim, fee remediation or unauthorised transfer).
Read the claim facts in panel 1, then edit any number (for example the repair quote or months affected). Everything recalculates immediately.
Review the policy clauses (panel 2). Higher relevance means a closer match to the case.
Check the evidence flags (panel 3). A red BLOCK means the payment must not proceed until the issue is resolved. WARN means check before approving.
Check the calculation (panel 4). Each line shows how the recommended amount was reached.
See who can approve (panel 5). Select the role you are acting as. The guardrails show Permit, Escalate or Deny for each action.
Record your decision with Record human decision. It is approved, escalated or held depending on the amount, your role and any blocking checks, and added to the audit trail.
Read the decision brief (panel 6). Press Verify with live backend to run the full pipeline on the server and see which model was used, whether the hallucination guard accepted the AI text and which clause versions were cited.

Using the API

POST /assess with JSON such as {"pack":"remediation","facts":"Fee charged in error.","inputs":{"fee":35,"months":26,"rate":8,"id_ok":1,"days":6}}. Packs are motor, remediation and unauthorised. Use /health to check the service. Public requests use the offline demo model; real models need an access code from the owner.

5. Limits

Synthetic data and simplified rules only.
The AI never approves or denies; always apply your own judgement and your organisation's policies.
Check any legal or regulatory reference with a qualified person.

Questions or access codes: contact Chirag Dani via chiragdani.com.