You don't need an AI plan for the whole company.

The trial needs someone responsible who can set aside time and review the result. Set a ceiling on time and money for the trial. A hypothetical allocation: one person responsible and two employees each get a couple of hours a week. This is a suggested allocation of time that you will need to adapt to the task.

How the necessary material may be used needs to be settled before testing starts. If access or setup takes longer, testing will need to move. Taking more than 30 days is not a failure.

Week 1: Describe the work and requirements

The work as it is done today gives you a basis for comparison. Record working time, corrections and any waiting time separately. The examples need to cover both routine and difficult tasks.

A short description of a usable result makes the responses easier to assess later. It also needs to state which mistakes require correction and which mean the trial must stop.

Before testing, it needs to be clear what material the AI tool may use, who will have access and who will check the results. The usual way of handling the task also needs to be ready if the trial has to pause.

By the end of the week, you will then have examples to compare against and shared quality requirements.

Week 2: Test without sending anything out

In the first test, an employee gives the AI tool the agreed material, such as a previous customer enquiry that may be used in the trial, and assesses the response. Nothing is sent to customers and nothing changes in your systems while you test.

The instruction is the text that tells AI what to do. Use some of the examples to make that text clearer. Keep some examples you have not used so you can see whether it also works on new tasks.

A specific mistake gives you more to work with than the verdict “the answer is bad”. Note which detail is missing or wrong and what the employee had to do to correct it.

Week 3: Let employees try it

Now the people who normally do the task try the solution on the saved examples. They should be able to use the instruction and assess the result themselves.

The time needed to reach a usable result includes preparation, checking and corrections. Also record tasks where using AI has to be abandoned and the work completed in the usual way. That time counts too.

Employees’ experiences matter here: where did the work become easier, and where did it create extra work? If the trial moves checking to another colleague, include that time too.

Week 4: Decide what happens next

The work from the first week gives you something to compare against. Assess quality and total time separately. The cost of the AI tool and the time needed to keep instructions and material up to date also belong in the decision.

A hypothetical example of a benefit: an order takes fewer minutes to prepare, including checking and corrections, without introducing more mistakes. That is the kind of difference the trial should examine.

• Continue within the same limited scope if the result meets the requirements and the benefit is worth the work.
• Adjust and test again if a specific mistake can be corrected. Describe what the next test needs to establish.
• Stop if checking takes too much work or the result cannot be used.

If the task only occurred a few times during the period, you may have too little evidence to judge. The decision can wait until you have the missing examples.

The decision and examples belong in the same place. If you continue, record who is responsible, who checks the results and when you will review the solution again.