# Production AI agent scope worksheet

By Revenue Arc Development | Version 1.0 | September 7, 2026

Guide: https://revenuearc.com/en/resources/production-ai-agent-scope

Use, edit, and share this original template for your own business or client work.
Blank inputs require your decisions and evidence. This is a scoping tool, not a
certification of reliability or a promise of savings.

## 1. One bounded job

- Business / team:
- Workflow owner:
- Workflow name:
- Trigger:
- Inputs:
- Finished result:
- Recipient / reviewer:
- Current steps:
- Current volume (source and date):
- Current handling time (observed or estimated):
- Current error / rework baseline (source and date):
- Business value of a correctly completed task:
- Explicitly out of scope:

## 2. Architecture choice

- Candidate: fixed automation / model-assisted workflow / copilot / agent
- Why this choice fits the decisions the system must make:
- Simplest option compared:
- What improvement justifies added complexity:
- Ordinary path:
- Ambiguous path:
- Failure / stop path:
- If multiple agents are proposed, why separate roles help:

## 3. Systems and permissions

| System / data | Authoritative record | Permission owner | Read allowed | Draft allowed | Write allowed | Approval required | Retention / permitted use |
| --- | --- | --- | --- | --- | --- | --- | --- |
| | | | | | | | |
| | | | | | | | |
| | | | | | | | |

- Actions never allowed:
- Exact proposed change shown at approval:
- How permissions are enforced outside model instructions:
- Treatment of untrusted text and conflicting instructions:
- Treatment of stale or missing records:
- Verification that an external action succeeded:
- Duplicate action protection when a task retries:
- Timeout, attempt, and spending limits:
- Pause / stop control and owner:

## 4. Representative evaluation cases

Use ai-agent-evaluation-cases.csv as a starting structure. Its example rows are
illustrative, not results from a Revenue Arc deployment. Replace them with cases
that represent your workflow and that you have permission to use.

- Case collection source and permission:
- Routine cases:
- Ambiguous cases:
- Missing-data cases:
- Access-denied cases:
- Tool timeout / failure cases:
- Duplicate-trigger cases:
- Cases where the correct behavior is to stop:
- Expected result for each case:
- Unacceptable action for each case:
- Automated checks:
- Human review rubric and owner:
- Repeated trial plan where behavior varies:
- Evaluation set held apart for release decisions:

## 5. Acceptance criteria

| Measure | Baseline | Release threshold | Measurement method | Owner |
| --- | --- | --- | --- | --- |
| Accepted task completion | | | | |
| Unacceptable errors / actions | | | | |
| Human rework | | | | |
| Time to accepted result | | | | |
| Cost per accepted task | | | | |
| Recovery after a failed tool call | | | | |

- System version evaluated:
- Evaluation dates:
- Resulting system state verified, beyond the generated explanation:
- Remaining known limitations:

## 6. Build and operating costs

- Discovery and workflow mapping:
- Build and integrations:
- Evaluation work:
- Model use:
- External tool / data charges:
- Hosting:
- Human review:
- Monitoring and maintenance:
- Support arrangements:
- Low / expected / high volume assumptions:
- Retry / rework cost assumptions:
- Estimated cost per accepted task:
- How recovered staff time will be used:

Keep currencies separate. Distinguish estimates from observed costs and capacity
created from proven cash savings.

## 7. Staged rollout

| Stage | Scope and permissions | Evidence required to continue | Owner | Decision date |
| --- | --- | --- | --- | --- |
| Controlled prototype | | | | |
| Shadow / draft-only use, if appropriate | | | | |
| Limited live scope | | | | |
| Wider operation | | | | |

- Launch decision: proceed / revise / stop
- Person accepting the release evidence:
- Rollback procedure:
- Failure alert and support owner:
- Code, data, and account ownership:
- Documentation and operating handoff:
- Next review date:
- Conditions requiring re-evaluation:

## Source context

Workflow and agent design distinctions:
https://www.anthropic.com/engineering/building-effective-agents

Evaluation design and checking outcomes:
https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents

Revenue Arc service: https://revenuearc.com/en/solutions/custom-ai-agent-development