AI pilot in an SME: five steps from problem to a measurable test
Most AI attempts in an SME fail not on the technology, but because in the end no one can say whether they helped. A cleanly set-up pilot answers exactly that question.
Using AI well · 16 October 2026 · 8 min read
An AI pilot is a small, time-boxed test on a real task. It should not impress. It should let you decide. That needs a clearly defined problem, a measured baseline, fixed limits with human approval, and a date on which you evaluate.
Do not start with the tool. Start with the question of where your team loses time every week. A pilot that is measurable from the start tells you honestly, at the end, whether the effort pays off.
Step 1: Choose a problem, not a tool
Describe a task that happens often, follows a similar pattern, and clearly costs time today. Tasks whose result a person can check in seconds work well: triaging requests, summarising documents, or preparing replies. A plumbing and heating company, for example, could start with its daily quote requests, because every request follows roughly the same pattern.
State the goal in one sentence and decide who owns the pilot. Not “we want to use AI”, but “we want to halve the time to a first reply without lowering quality”. The person who already does the task today is usually the best owner.
Step 2: Measure the baseline
Without a baseline, any improvement is a guess. Over two or three typical weeks, count the cases, the handling time, the follow-up questions, and the corrections. For the plumbing company, that means timing every quote request from arrival to the finished reply.
Note which mistakes happen today and which ones would be costly, such as a transposed number in a quote. These figures become your yardstick and protect you from a gut feeling that is easy to fool.
Step 3: Set clear limits and a human approval
Define which data may go in, which parts of the task the AI takes on, and where a person checks. Customer, staff, and financial data should not go into an arbitrary tool unchecked. For the quote request, that means the AI drafts a reply and a staff member checks prices and measurements before anything goes out.
Data protection law, such as the GDPR, applies to AI-supported processing too. Before you start, clarify which service is used, where the data goes, and who approves the result. Prices, payments, and binding commitments stay in human hands.
Step 4: Test small with real cases
Take the owner from step one and a fixed, manageable number of real cases, say twenty to thirty requests. A small scope quickly shows where rules are missing and how often you still need to correct something.
During the test, record the same figures as for the baseline: handling time, corrections, open questions. That creates a fair comparison instead of a polished demo.
Step 5: Evaluate and decide honestly
Compare handling time, corrections, and open questions with the baseline. Include the review and rework effort. If checking the plumbing company's quotes ends up taking longer than writing the reply by hand did before, the pilot has not paid off, however fast it looks.
The result is one of three decisions: expand, adjust, or stop. All three are good outcomes when they rest on numbers, not on a feeling.
Set the stop criteria in advance
A pilot is allowed to fail. Decide before you start when you stop:
- Rework is greater than the time saved.
- Mistakes do not stay reliably visible before the next action.
- The needed data is incomplete or unclear.
- The team bypasses the new workflow and works as before.
Sources and further reading
FAQ
Long enough for enough real cases, short enough for a quick decision. For a frequent task like quote requests, two to four weeks is often enough. What matters is the number of cases, not the calendar.
Then the pilot starts with measuring. Estimate roughly, measure cleanly from today onward, and accept that the first round mainly brings clarity about the workflow, not yet the comparison.
No. The biggest input is attention, not money. A responsible owner, a clearly bounded workflow, and honest numbers matter more than an expensive tool.
Mygenzy works with small and mid-sized businesses on processes, marketing, and AI. This article was created with the help of AI and reviewed by people.
Find the workflow that deserves a pilot
In the Growth Check we choose a suitable task together, set the baseline, and define limits and stop criteria. You start with a test that lets you decide.