PL

How to plan your first AI pilot in a small business

, AI implementation
Published:

Start with one task that takes your team time and produces a result you can check. Measure how the team handles it today. In the pilot, compare that result with AI-assisted work, including time spent reviewing and correcting suggestions. Decide in advance what to do if the trial exposes a serious error.

Choose one task

For a first trial, I’d choose a task that comes up regularly. You need examples, someone who can recognise the right result and a way to correct a mistake quickly. If a simple inbox rule can identify some messages by sender, include that option in the comparison.

Hypothetical example. A team manually routes email from a shared inbox to sales, technical support, billing or complaints. In the pilot, AI suggests one queue or flags the message for manual review. A staff member checks the suggestion and then routes the message. Sending replies and updating the CRM are outside the trial’s scope.

Measure how the process works today

Choose someone in the business to own the test. Ask them to record the time it takes today to read a message, choose a recipient and correct misrouted messages. In the AI trial, measure the same activities on a comparable set of cases. Track urgent cases separately, because one delayed email can matter more to the decision than the average sorting time.

Prepare a sample and expected results

Choose typical messages from several categories and include harder cases. Someone who knows the process should label the expected result before seeing the AI suggestions. They can use past routing as a reference and should check whether the original assignment was correct. Keep a separate sample to use after any changes to the instructions.

Start with prepared examples and send only the text needed for routing. Remove unnecessary details from signatures, attachments and conversation history. Before using real mail, agree which data the tool will receive, who can see the results and how long to keep the test data. For the first trial you can work on a copy without connecting the model to the live inbox.

Check what happens when a case falls outside the rules

The messages below are made up. In a real test, set the expected outcome using your own business rules. Agree what staff should do if a suggestion is missing or wrong.

Example messageExpected handling
‘Please send a quote for 20 licences.’AI suggests sales. A staff member checks and approves.
‘Something isn’t working.’The message goes to manual review. A staff member clarifies what the problem is.
‘Urgent, the system has stopped working.’A staff member follows the existing urgent-case procedure. Check whether the AI suggestion distracts from the urgency.
‘We want to discuss the contract and report a billing error.’The message goes to manual review. A staff member decides the order and who owns each issue.

Also test what happens when the tool returns no answer. The message should stay in the normal queue so the team can continue working and investigate what happened.

Compare time and accuracy

Include reviewing suggestions, correcting them and handling exceptions manually in the AI-assisted time. To assess quality, compare both the AI suggestion and the final routing decision with the expected result. The number of approvals alone says little about accuracy. Count serious errors, such as a missed urgent request, separately.

  • Continue on a limited scale if the result meets your agreed criteria and the team can handle exceptions.
  • Make a change and repeat the test on a fresh sample if errors cluster around one category or an unclear instruction.
  • Stop the pilot if review takes too much time, errors are hard to catch or the trial exceeds the risk threshold you set.

First AI pilot worksheet

Copy these fields into a document your team uses. Fill them in before the trial and return to them when deciding whether to proceed.

FieldWhat to record
TaskWhat exactly does AI suggest and who approves the result?
Process ownerWho leads the trial and makes the decision?
Test sampleWhich typical messages and exceptions will you check?
Data sent to the toolWhich part of the message will you send to the tool?
BaselineHow long does manual handling take and what errors occur?
Success criteriaWhat time including corrections and what level of accuracy will be acceptable?
Stop conditionWhat error will stop the trial? How will the team return to manual handling?
Decision dateWhen will you review the results and decide what to do next?

Want to test one process in your business?

Tell me how your team handles the task today. On the call we can choose what to include in the pilot and how to judge the result.

Book a free callSee how I implement AI for businesses

Sources

Get in touchBook a free call