Most small business AI pilots end without a decision. Nobody wrote down what success would look like before it started, so at day 30 the conversation is about whether people liked the tool. Liking a tool is not a result.

A pilot worth running takes four weeks, costs a few hundred dollars, and produces a number you can act on. Here is how to set one up in a company of 5 to 100 staff.

Pick one workflow, not a department

Choose a single task that happens at least five times a week, takes someone more than 20 minutes each time, and ends in text or a document. Writing up site visits. Answering repeat customer questions. Turning meeting notes into action lists. Drafting quotes from a standard scope.

Leave out anything where a wrong answer reaches a client without a human reading it first, and leave out regulated data entirely for these 30 days. Write the workflow on one line before you spend any money. If it does not fit in one sentence, it is too big to pilot.

Choose three to five people, and none of them by volunteer call

Open sign ups get you the staff who are already enthusiasts, and enthusiasts produce results you cannot repeat with everyone else. Pick people who do the chosen task every week: one who is fast at it, one who is average, one who is skeptical. Three is enough in a 10 person company. Five is enough in a 60 person one.

Tell them plainly that the pilot judges the tool, not them, and that reporting it as useless is a valid outcome. Name one person as the owner who collects the numbers. Without an owner the logging stops in week two.

Measure the current state for one week before you start

Spend week zero recording how the task is done today. Each participant logs three things per instance: the date, minutes spent, and whether the output needed rework. That is about 10 seconds of logging per task, in a shared spreadsheet.

Without that week every claim afterwards is a feeling. Twenty to thirty logged instances is enough to decide on.

Know what the 30 days actually costs

Business tier seats for the main assistants run roughly $25 to $35 per user per month, billed monthly if you ask for monthly. Five seats is around $150 for the pilot month. Meeting notetakers sit near $20 to $30 per user. Microsoft 365 Copilot is usually sold on an annual commitment, which makes it a poor fit for a 30 day test unless you already have the licences.

The larger cost is internal time: roughly 90 minutes of setup, two hours of training spread over the first week, and 15 minutes per weekly check in. Budget six to ten hours of someone's attention across the month and pay for business accounts rather than free ones, so your data is not used for training and you can see usage.

Run 15 minute check ins every week

One short call each Friday. Three questions: what did you use it for, where did it waste your time, and what did you stop trying. The third question finds the real ceiling of the tool faster than any feature list.

Fix prompt and setup problems in week one and week two. After that, freeze the configuration so the last fortnight measures a stable thing. Keep a running note of anything that broke, including wrong facts in output.

Judge it on four numbers at day 30

Compare against the baseline week and look at four things.

  • Time per instance. Median minutes now versus median minutes before.
  • Rework rate. What share of outputs needed a meaningful correction. If this went up, time savings are not real.
  • Adoption. How many working days each person actually used it. Under half is a signal the workflow was wrong, not the tool.
  • Failures. The count of wrong or unusable outputs, with examples kept.

Then do the arithmetic. A task taking 30 minutes, done 10 times a week, cut by 20 percent, returns an hour a week per person. Across five people at a loaded rate of $40 an hour that is about $800 a month against $150 of subscriptions. If the saved time does not clear the subscription cost at your own loaded rate, stop, and do not let a rounding argument keep it alive.

Write the decision down on day 31

Three outcomes are allowed. Expand it to everyone who does that task, with a one page usage policy and a named owner. Extend it 30 more days with exactly one change, stated in advance. Or cancel the seats and record why, so the same tool does not get proposed again in six months without new information.

Cancelling is a good outcome when the numbers say so. A company that has run and killed one pilot is in a better position than one that has run none, because it now knows what its own baseline looks like.

If you want a second pair of eyes on the design, the first call is free and runs about 30 minutes. We pick the workflow, agree the measurements, and set the threshold you will judge it against before anything is bought. Get in touch and pick a time.