# AI Agent PRD Template

A one-page product requirements document for teams moving from a chatbot demo to
an agent that runs inside real operations. Replace bracketed text.

## 1. Summary

- **Agent name:** [name]
- **Owner:** [business owner]
- **Technical owner:** [engineer]
- **Status:** Draft / In review / Approved
- **Target launch:** [date]

## 2. Problem

Describe the operational bottleneck in numbers.

- Current volume: [e.g. 1,200 tickets per week]
- Current cycle time: [e.g. 4.2 hours per case]
- Current error or rework rate: [e.g. 11%]
- Cost of the bottleneck: [e.g. $18,000 per month]

## 3. Job to be done

> When [trigger], the agent must [action] using [sources] so that [outcome],
> while never [hard constraint].

## 4. Scope

### In scope

- [Task 1]
- [Task 2]

### Out of scope

- [Explicit exclusion]
- [Explicit exclusion]

## 5. Tools and data

| Tool / source | Access method | Read or write | Failure behaviour |
| --- | --- | --- | --- |
| [CRM] | [API] | Read | [Retry, then escalate] |
| [Policy library] | [Vector retrieval] | Read | [Answer with citation only] |

## 6. Decision policy

- Confidence threshold for auto-approval: [e.g. 0.92]
- Escalation rules: [what forces a human]
- Forbidden actions: [payments, deletions, external commitments]
- Approval authority: [role]

## 7. Output contract

Define the exact shape the agent must return.

```json
{
  "case_id": "string",
  "decision": "approve | reject | escalate",
  "confidence": 0.0,
  "evidence": ["source-id"],
  "notes": "string"
}
```

## 8. Evaluation

- Golden dataset size: [e.g. 250 labelled cases]
- Quality bar: [e.g. 96% field accuracy, 0 critical policy violations]
- Latency bar: [e.g. p95 under 20 seconds]
- Cost bar: [e.g. under $0.12 per case]
- Human review sample: [e.g. 10% weekly]

## 9. Observability and audit

- Every step is logged with inputs, tool calls, and outputs.
- Traces are retained for [duration].
- Overrides are attributed to a named reviewer.
- A replay path exists for any decision.

## 10. Rollout

1. Shadow mode for [duration]; no writes.
2. Assisted mode behind human approval.
3. Limited auto-approval for [segment].
4. Full rollout after [exit criteria].

## 11. Risks

| Risk | Likelihood | Impact | Mitigation |
| --- | --- | --- | --- |
| [Hallucinated field] | [M] | [H] | [Cite source, threshold, review] |

## 12. Sign-off

- Business owner: ____________________
- Security: ____________________
- Engineering: ____________________

---

GrowBig IT · growbigit.com · hello@growbigit.com