# AI agent commissioning brief

A working template from PlatformBridge for comparing delivery proposals.
Complete permissions and acceptance criteria with the people who own the work.
This template does not establish that an agent is the right solution.

Companion guide: https://www.platformbridge.ai/guides/ai-agent-development-services

## 1. The job and its boundaries

- Process owner and users:
- Trigger, input, and completed business outcome:
- Decisions the model needs to make about the next step:
- Why a fixed workflow or simpler software is insufficient:
- Allowed tools, records, and data sources:
- Excluded actions and data:
- Maximum attempts, elapsed time, and cost before stopping or escalating:

## 2. Action permissions

| Action | May prepare? | May execute? | Required approval and evidence | Downstream enforcement | Revocation / expiry |
| ------ | ------------ | ------------ | ------------------------------ | ---------------------- | ------------------- |
|        |              |              |                                |                        |                     |

- How approval binds to the exact action, content, and recipient:
- What invalidates an existing approval:
- How duplicate or uncertain writes are handled:
- Who may expand or revoke permissions:

## 3. Acceptance cases

| Case                                           | Expected outcome and permitted actions | Escalation behavior | Measured outcome / cost / time | Release decision |
| ---------------------------------------------- | -------------------------------------- | ------------------- | ------------------------------ | ---------------- |
| Representative normal work                     |                                        |                     |                                |                  |
| Missing or ambiguous evidence                  |                                        |                     |                                |                  |
| Request for an unauthorized action             |                                        |                     |                                |                  |
| Malicious instructions in a retrieved document |                                        |                     |                                |                  |
| Expired approval or changed input              |                                        |                     |                                |                  |
| Timeout after a write                          |                                        |                     |                                |                  |
| Tool unavailable or budget exhausted           |                                        |                     |                                |                  |

- Evaluation cases held apart from development examples:
- Acceptance thresholds by consequence, not just one aggregate score:
- Known limitations, manual route, and accountable risk owner:

## 4. Cost of a completed case

- Model and tool cost:
- Additional infrastructure and licenses:
- Review minutes and loaded labor rate:
- Expected correction work, with basis for the estimate:
- Shared support costs, allocated separately:
- Baseline and measurement window for comparing the whole process:

Avoid counting review effort twice if it is already included in a net
time-reduction assumption. Label all estimated inputs.

## 5. Handover and change control

- Account, data, and credential ownership:
- Source or usage rights and deployment instructions:
- Versioned prompts, tools, models, and test cases:
- Monitoring owner and incident escalation:
- Demonstrated pause and recovery procedure:
- Support scope, hours, and exclusions:
- Changes that require new evaluation and approval:
- Release approver, date, evidence version, and unresolved limitations: