I would measure the current job before judging the AI version. Otherwise, a polished demo can become the baseline, and the awkward parts of the real process disappear from the comparison.
You do not need a claim about hours saved to get started. You need a clear task boundary and a record of what actually happens. Here is a worksheet I would use for a small pilot. It is a proposed method, not a report of measured client results.
Choose one repeatable task
Pick something with an identifiable beginning and end. “Make the office more productive” is too broad. “Turn an approved public product sheet into a draft answer for a general product question” is a task you can inspect.
Write the boundary before gathering observations. The clock starts when the required inputs are available. The task ends when the reviewer accepts the answer. If information is missing, record that explicitly instead of pretending it was a normal case.
Keep sending the answer outside the pilot unless that action has separate approval. Draft quality and permission to communicate are different questions.
Copy this baseline worksheet
Use your normal spreadsheet or a paper form. Record one row per attempt:
- Task reference: a non-sensitive ID that lets you recognize the same item later.
- Task type: ordinary, incomplete input or unusual request.
- Start and finish: the defined boundaries, recorded consistently.
- Hands-on work: time spent preparing, checking and correcting the draft.
- Waiting: time spent waiting for information or another person, recorded separately.
- Review result: accepted, corrected or rejected, using a definition agreed in advance.
- Rework: what had to be fixed and the effort needed to fix it.
- Missing inputs: what was unavailable and who supplied it.
- Notes: interruptions, unusual conditions or changes to the process.
Do not put unnecessary customer details into the worksheet. The task reference should be enough to connect a row to an authorized internal record when needed.
Define “good enough” before comparing drafts
For a fictional product-answer task, I would require the draft to match the approved product sheet, avoid unsupported delivery promises, identify missing information and leave the reviewer a clear next step. Another task may need different criteria.
Write those criteria in plain language. “Sounds professional” is not a substitute for “uses the correct product specification.” A draft that reads well but needs extensive correction has not passed the same test as an accepted answer.
Use fictional or otherwise authorized test material. A measurement exercise does not create permission to upload private information to a new tool.
Keep effort separate from elapsed time
An item can take a long time to finish because it sits in a queue, even when the hands-on work is short. If you combine those measures, you may credit AI for a scheduling change, or blame it for a delay it did not cause.
Record both when they matter. Include the review and cleanup work. If a colleague has to correct the draft later, that effort belongs in the observation, not outside the story.
You can begin with a modest pilot you can personally review. There is no universal sample size in this worksheet. The amount and variety of evidence you need depend on the task and the consequences of getting it wrong.
Read the baseline without overselling it
At the end, list what you observed and what is still unknown. If only routine requests were included, do not describe the result as covering every request. If an interruption affected an attempt, keep the note instead of quietly discarding the inconvenient row.
My first decision would be whether the task is stable enough to compare at all. If the inputs or acceptance rules keep changing, I would clarify the process before testing an AI-assisted version.
On Friday, I will use this baseline in a shadow-mode trial: the AI proposes an answer while the existing process remains in charge. If you want help choosing a narrow pilot and defining what evidence would count, visit White Rabbit Advisory Group at https://wrag.ai.