# Is this agent making the work better?

A working template by Tommy Magee. Adapt it to one real workflow. This is an evaluation template, not a report of measured outcomes.

## 1. Name the work

- Workflow and accountable owner:
- Business outcome:
- Who receives the result:
- Current bottleneck:
- Can this step be removed or simplified without AI?

## 2. Establish a baseline

Use the same definition of a completed, acceptable task before and after the change.

- Observation period and number of representative tasks:
- Task volume per month:
- Active human minutes per task, including checking, corrections, and handoffs:
- Elapsed completion time, measured separately from active effort:
- Acceptance criteria and current error/rework rate:
- Data source for each estimate or measurement:

## 3. Bound the agent's responsibility

- Sources it may read and actions it may take:
- Decisions requiring human approval:
- Conditions that require stopping or escalating:
- Examples of acceptable and unacceptable outputs:
- Evaluation set, including missing data and unusual cases:
- Who can stop the system and how to return to the previous process:

## 4. Compare under real conditions

- Human minutes per task after the change, including review and correction:
- Failed tasks requiring fallback, and the effort they consume:
- Output quality against the same acceptance criteria:
- Elapsed completion time after the change:
- Actual usage and the fraction of eligible work covered:
- New work created for other teams:

## 5. Test the economics

Monthly capacity released (hours) = actual tasks handled × (baseline human minutes − human minutes after the change) ÷ 60 − recurring support hours.

Include failed and fallback tasks in the average. If the result is negative, the workflow consumes more human capacity. Report elapsed-time improvements separately so waiting time is not counted as labor savings.

If useful, value that capacity using an explicit hourly cost assumption. Subtract incremental recurring costs such as model usage, hosting, and software. Do not count support labor twice if it is already included above.

Track one-time design, integration, training, and rollout costs separately. A payback estimate only makes sense when recurring net benefit is positive.

**Capacity is not automatically cash savings.** Say what the released time enables. Only claim cash savings when spending is actually avoided or reduced, and avoid double-counting capacity and headcount benefits.

- Estimated benefit before testing:
- Observed result and observation period:
- One-time costs:
- Recurring costs:
- Uncertainty and sensitivity to task volume, error rate, and review time:

## 6. Make the next decision

Choose: stop, revise, continue the bounded pilot, or expand.

- Evidence supporting that decision:
- Quality or safety conditions that must hold regardless of savings:
- Owner and support plan:
- Next review date:
- What would cause us to reverse the decision?