# Sample AI workflow audit — synthetic support triage

This is an anonymized demonstration only. It uses invented data and does not describe a
client's production systems, credentials, or guaranteed savings.

## 1. Current flow

`support inbox → classify request → retrieve approved answer → human review → reply + log`

### Observed risk boundaries

- Customer identity and payment information remain outside model prompts.
- Refunds, legal threats, safety issues, and low-confidence classifications route to a human.
- Every proposed reply stores the source snippet and reviewer decision.

## 2. Recommended bounded design

1. Receive a redacted message through the existing support tool.
2. Classify into `billing`, `technical`, `account`, or `escalate`.
3. Retrieve only approved help-center passages.
4. Draft a reply with citations and a confidence score.
5. Require human approval for `escalate` or confidence below `0.85`.
6. Log the input hash, retrieved source IDs, reviewer, and final action.

## 3. Acceptance checklist

- [ ] No secret, payment number, or unredacted PII enters the model step.
- [ ] Unsupported claims are blocked rather than improvised.
- [ ] Human escalation is testable with fixtures.
- [ ] Retry and duplicate-event behavior is documented.
- [ ] A reviewer can reconstruct why each reply was approved.

## 4. Handoff

The paid audit supplies this level of map, risk boundary, human-review design, test cases, and
next-step implementation notes for one real workflow. It is delivered asynchronously within 48
hours after a complete intake and includes one written revision.