How to measure AI productivity: a 30-day business pilot

Days 1–7: measure the work as it is

Pick one repeated task with enough volume to test. Preparing customer follow-ups is a better pilot than an annual strategy presentation. Name one person responsible for the result, and agree on what ‘done correctly’ means before building anything.

For a week, record the time spent preparing, checking, correcting, and recording each result. Include representative easy and difficult cases. Track the volume, the backlog, and mistakes that require rework. Without that baseline, a fast demo can look like a saving even when the overall process gets slower.

Days 8–14: build one narrow workflow

Give the agent a defined input and output. For example: take a new inquiry and prepare a customer brief plus a draft reply. Use reliable business records, display the sources, and keep a person approving the output while you learn where the system struggles.

Try old examples first, including incomplete requests and contradictory information. Check that the agent asks for help when it should. Before testing with live work, decide who deals with failures and how the team returns to the manual process if needed.

Days 15–21: compare complete tasks

Use the pilot on a limited set of real requests. Compare similar types of work, and include cases where the agent failed. Measure the time a person actually spends after introducing AI: preparing inputs, reviewing drafts, fixing errors, and handling exceptions.

Here is an illustrative calculation, not a client result. A task used to take 20 minutes. With AI, preparing the input takes 2 minutes, review takes 5, and corrections take 3. That is 10 minutes of human work, so the time saving is 50%. If review and corrections push the total to 18 minutes, the saving is 10%, even if the first draft appeared instantly.

Days 22–30: decide with the team

Review the numbers with the people doing the work. Look for hidden costs: another person checking outputs, duplicate records, slower customer responses, or maintenance that nobody has counted. Include software, model usage, and ongoing ownership in the cost estimate.

My 50% ambition is a useful target for suitable repetitive workflows. It is not a reason to discard a reliable 20% improvement, and it is not permission to accept worse work. Agree on the minimum saving and quality standard that make this particular process worth running.

• Expand if human time falls, quality holds, and the team can handle exceptions.
• Revise if useful drafts still take too much effort to check or the inputs are unreliable.
• Stop if the process creates more work than it removes or nobody can own it after launch.

Give the saved time a purpose

If the team saves five hours a week, decide what those hours should make possible: quicker follow-ups, fewer overdue requests, or more attention to complex customers. Otherwise the saved time can disappear into other admin work without changing the business.

Once one workflow works, use what you learned to choose the next one. That is how I would build a more efficient business: one measured improvement at a time, with software that fits the people using it.