What problem is worth solving?
AIG’s 2025 annual report describes contextualized data, an ontology and workflow changes for underwriting, with capabilities extended to claims. Anthropic’s guide separately reports early improvements in review speed and data accuracy. The learning focus is separating evidence preparation from risk judgment.
Reduce evidence preparation time so experts can judge risk from complete, traceable facts.
- WORKING ROLES
- Underwriting experts, document operations, data engineers, security and FDEs
How does an FDE turn symptoms into a solvable problem?
The following is an FDE exercise inspired by the case, not a claim about undocumented project events. Reported team actions appear in the context above.
Replay a document journey
Trace intake, missing-document requests, extraction, cross-checks and expert review on a synthetic submission. Ask whether the bottleneck is interpretation or locating evidence.
Inspect field-level error cost
Inspect currency, amounts, dates and exclusions separately. Diagnose attachment conflicts, lost table headers, unit conversions and invented missing values.
How does the solution enter the workflow?
- 01
Document register
Track type, version, pages and missing material; reject unsupported formats.
- 02
Extract
Return value, unit, source location and validation status; preserve unknowns.
- 03
Reconcile
Expose contradictions between attachments and rules without silently picking a value.
- 04
Expert workspace
Show summary beside evidence, capture corrections and add them to evaluation sets.
Which decisions make the solution reliable?
Separate extraction from risk ownership
Define extractable fields and restricted decisions. Authorized specialists remain responsible for appetite, pricing and underwriting decisions.
Preserve fidelity before brevity
Do not drop negation, conditions or exclusions. Score retrieval, extraction and summarization separately to localize faults.
Conflicting amounts across attachments
Synthetic materials show 1,200,000 CNY in a main form and 1,200,000 USD in an attachment. Matching numbers do not establish matching facts.
- Extraction
- Keep both values, currencies, pages and versions.
- Resolution
- Set currency_conflict, request clarification and withhold the aggregate.
- Test fixture
- An undetected currency conflict fails regardless of speed.
What should you do and deliver?
Map documents and expert decision points
- OUTPUT
- Interview guide + field ownership
- PASS CONDITION
- Each critical field has a reviewer
Specify access and retention
- OUTPUT
- Data flow + minimal field set
- PASS CONDITION
- Unauthorized reads are denied
Test conflicts, missing pages and negation
- OUTPUT
- Layered eval set
- PASS CONDITION
- Extraction and summary failures are distinguishable
How do you prove completion and reuse the work?
Acceptance checks
Measure preparation, critical-field error, missing-document requests and expert rework. Include difficult attachments rather than only clean forms.
Reusable assets
Evidence location and conflict handling transfer to contracts, specifications and audit work; domain judgment must be redefined by experts.
Further reading and original sources
- Claude for the financial industry: A practical deployment guide (AIG case) ↗Anthropic · 2026
- AIG 2025 Annual Report · The Future of Insurance ↗AIG · 2025 年报 / Annual report