01 / Decision
Proceed with an incremental review layer
Build a bounded intake and review layer around the existing system of record. Complete one request type from receipt through reviewed handoff, preserving the current application's business rules. Evaluate AI extraction against a representative baseline after the workflow and validation rules are explicit.
The situation
An operations team receives requests by email, checks details against an existing application, and tracks exceptions in a spreadsheet. Leadership wants to introduce AI extraction. Before choosing a model, the team needs a consistent way to capture requests, validate fields, assign review, and understand what happened to each record.
The alternatives and their tradeoffs
Extend the spreadsheet workflow
Useful for establishing a baseline
Small initial change; ownership, history, and reconciliation still depend on manual coordination.
Replace the operational platform
Defer until a wider replacement case is established
Broader redesign; requires migration of business rules and integrations before the first workflow is proven.
Add an intake and review layer
Recommended first increment
An additional component to operate; creates a bounded way to test the workflow while preserving the current platform.
A complete path through the system
Each step carries the request forward with an owner, a status, and a source reference.
01
Capture
Retain the request and its source reference.
02
Validate
Check required fields, identity, and duplicates.
03
Review
Assign exceptions to an authorized reviewer.
04
Handoff
Record the reviewed outcome in the existing system.
02 / Delivery
One request type, one review queue, one verified handoff
- 01
Establish the evidence
Walk through real user tasks, map identifiers and permissions, inspect a representative request sample, and measure handling time and rework. Record the baseline and acceptance criteria.
- 02
Build the complete path
Implement intake, validation, review ownership, durable status, and the agreed integration. Keep the scope to one request type so the full workflow can be demonstrated.
- 03
Demonstrate and decide
Exercise normal and failure paths with users, compare behavior with the baseline, and document open issues. Choose whether to expand intake coverage or evaluate AI extraction next.
What must be demonstrated
- A valid request reaches the agreed destination with a traceable source and review history.
- Missing or conflicting information enters a visible exception state with an assigned owner.
- Repeated delivery does not create an additional operational record.
- Access checks cover both the interface and the underlying record operations.
- A failed handoff can be investigated and retried without losing the request or its status.
Where AI fits
Use AI to assist code investigation, implementation, and test development under engineer review. Evaluate an AI extraction feature separately using representative requests, field-level error analysis, and reviewer effort. Release decisions depend on observed behavior and agreed criteria.
What the team takes forward
A working increment, recorded architecture decisions, validation evidence, known issues, operating instructions, and a prioritized next milestone.
Expand only after the first workflow can be operated and evaluated. Use measured handling time, error patterns, and reviewer effort to choose the next investment.