AI-assisted FDE work: what changes and what does not
Updated
Drafting gets faster; judgment gets more valuable. Where assistants help inside customer work and where they never decide.
Assistants compress drafting time from hours to minutes. They do not compress trust, and trust is the FDE product.
The short answer
Use AI for drafts and summaries; keep humans on interviews, access decisions, metrics and sign-off. Every AI artifact ships with a named reviewer.
The decision table
| Task | AI drafts | Human decides |
|---|---|---|
| Interview notes summary | Yes | Operator confirms the number |
| Brief first draft | Yes | Owner signs the scope |
| Test scaffold | Yes | Engineer reviews edge cases |
| Error report triage | Yes | Owner approves quarantine rules |
| Access grants | Never alone | Named approver, logged |
| Readout numbers | Never invented | Measured on live data |
Worked example: the assisted slice
A fictional FDE (fictional) runs two interviews, dictates notes, and gets a brief draft in minutes. The draft invents a baseline number; the operator corrects it from the spreadsheet. The test scaffold covers the happy path; the engineer adds duplicate and expiry cases. The slice ships in three days instead of five, and every number traces to a human-confirmed source. The evaluation post shows how to score assistant output; the permissions post sets the access bar.
Checklist: assisted without reckless
- No AI number reaches a customer without a human-confirmed source.
- Access changes always carry a named approver.
- Drafts are labeled as drafts until review passes.
- The readout states which parts were assisted and who reviewed them.
Related reading
Straight answers
Frequently asked questions
Does AI replace discovery?
No. Interviews, trust and sign-off need a person in the room. Assistants summarize notes; they never sign the brief.
Where does AI help most?
First drafts: briefs, test scaffolds, error-report summaries, runbook formatting. Speed on drafts, humans on decisions.
What is the biggest risk?
Confident wrong output handed to a customer unchecked: invented data, fake citations, unreviewed access changes. Every AI draft gets a human check with a name.