How to try AI without betting the company

The worst way to adopt AI is all at once. The second worst is not at all. There is a sensible middle: one small, well-chosen project, run carefully, that tells you whether there is more worth doing. Here is how to set one up.

Step 1. Choose a job, not a technology

Do not start by asking which tool to buy. Start by asking which task wastes the most time. A good first candidate is frequent, repetitive, and low-risk if it goes slightly wrong. Following up enquiries, drafting standard replies and producing a weekly report all qualify. Signing contracts does not.

Step 2. Write down how it works today

Sit with the person who does the job and record every step, including the ones nobody would think to mention. This is the most skipped stage and the most valuable. You will often find the task can be simplified before any new tool is involved.

One job. Four weeks. Then decide.

Step 3. Decide what a person must still check

Before building anything, agree where a human stays in the loop. A draft that someone reviews is a very different thing from a message that sends itself. Begin with review on everything. Remove a check only when you have seen it be right many times.

Step 4. Agree what the tool may see

Be specific about which information goes into which tool and where it is stored. Customer details, prices and staff records deserve a deliberate decision, in writing, not a default setting nobody read.

Step 5. Run old and new side by side

For the first weeks, keep doing the job the old way as well. Compare the results. This costs a little extra effort and buys you something important: you find the problems while they are still harmless.

Step 6. Measure one thing

Pick a single measure before you start: hours spent, time to first reply, errors caught. Check it at the beginning and after a month. If it has not moved, stop and say so. A small project that honestly failed is cheap. A large one that nobody dares to question is not.

Then decide

After a month you will know three things you could not have learned from a presentation: whether the tools suit your work, whether your team will use them, and whether the people you hired to help are any good. Only then is it worth talking about a bigger plan.

Next: What a $2,500 website should actually do