Copyable template
A before-and-after plan for one AI-assisted workflow
Use four weeks to see whether one task improved. Keep it small enough that you can still tell what happened.
1. Write down what you are testing
We want to improve [task] for [group].
We expect [gain] to change from [baseline] to [target].
We will not accept a decline in [quality, safety, experience, or service measure].
If the test works, we plan to use the gain for [specific choice].
Example: We want to improve meeting preparation for eight account managers. We expect median preparation time to fall from 45 to 30 minutes. We will not accept lower reviewer scores for factual accuracy. The team will use the time for customer follow-up.
2. Decide where the task starts and ends
Write down exactly where the task starts and ends. Meeting preparation might start when someone opens the customer record and end when the briefing is ready for review. Count searching, prompting, editing, checking, and fixing mistakes.
- Start event:
- End event:
- People involved:
- Typical output:
- Required checks:
- Reasons to stop or escalate:
3. Pick three to five measures
| Measure | How collected | Owner | Frequency |
|---|---|---|---|
| Median task time | Simple task log | Participants | Every sampled task |
| Factual corrections | Reviewer checklist | Team lead | Every sampled output |
| Task effort, 1 to 5 | One-question survey | Participants | After each task |
| Customer follow-up completed | Existing CRM record | Team analyst | Weekly |
Use measures your team already collects when possible. If the test creates an extra hour of admin for everyone, people will quite reasonably stop doing it.
4. Record the starting point in week 1
Collect 10 to 30 examples of the current method. Compare similar tasks, but keep the difficult cases. Record the median and range. Write down why an example was unusual instead of quietly deleting it.
Ask two short questions after each sampled task:
- How much effort did this task require, from 1 to 5?
- What caused delay or rework?
5. Try the new method in weeks 2 and 3
Agree on a simple method before the test. Say what AI can do, what the person must check, and what information stays out of the tool. Let people report failures without making them feel they have ruined the pilot.
Keep using the same measures. Do not change the rules halfway through because the first result looks disappointing.
6. Review what happened in week 4
What improved? ___
What became worse or harder? ___
Who benefited most and least? ___
Where did any saved time go? ___
What evidence is still missing? ___
Decision: keep, change, expand, or stop? ___
7. Write the result like a normal person
Eight account managers tested Copilot for 24 client briefings. Median preparation time fell from 43 to 31 minutes. The range remained wide. Reviewer corrections did not increase. Six participants found the method easier, while two said checking source material took too long. The team will improve the source-checking step before testing with another group.
This says what happened, where it happened, who took part, and what comes next. It also avoids pretending one small test has settled the entire AI debate.