How a small firm can start with AI safely

A small firm can start with AI safely by doing three things in order. Pick one small, boring job. Write down what the AI must never do before it touches that job. Then have a person check its work against something real, because the AI can't check itself. Everything else, including which tool to buy, matters less than those three.

Why most first attempts go wrong

The usual start looks like this. The owner buys ChatGPT or Copilot logins for the team, sends a message saying "use it wherever it helps", and waits. A month later two people use it every day, most opened it once, and nobody knows what has been pasted into it. One person has put in a customer list to "clean it up". Another asks it for contract wording and sends the answer out without reading it twice.

None of that is the tool's fault. It's what happens when a shop hands a new hire the keys on day one, with no list of what they're allowed to touch and nobody looking over their work for the first week. A person in that seat would at least feel unsure and ask. An AI helper won't, and that's the gap a safe start has to cover.

Step 1: pick one boring job

Choose a job the team repeats every week, that costs real hours, and where a mistake is cheap to catch. Sorting customer reviews for a small food business is a good example. Say it takes someone 3 hours a week to read every review and tag it as praise, a complaint about delivery, or a complaint about the product. That's small enough to watch closely, and a wrong tag costs nothing worse than a second look.

Don't start with anything that talks to customers on its own, moves money, or decides who gets hired. Those come much later, if at all.

Step 2: write the never list first

Before the AI does any of the job, write what it must never do. For the review sorter that list is short. It never replies to a customer. It never offers a refund, a discount or a free product. It never makes a health claim about the food. And when it isn't sure, it writes NOT SURE and says why, instead of guessing.

The same idea works for staff using a chat tool. Write down what never goes into it: customer phone numbers, salary sheets, signed contracts, anything a client shared in confidence. Put it on one page and talk it through once with the team.

I learned this the hard way. One of my own AI helpers followed its rules exactly, in a case I'd never pictured, and deleted 96 links between my notes without a warning. It took a day to repair them. Nothing on its list said "stop if you're about to remove more than a handful of links", so it didn't stop.

Step 3: a person checks the work against something real

The AI never marks its own homework. Before the firm trusts it, someone asks it ten questions the team already knows the answers to, and counts how many it gets wrong while sounding completely sure. That number shows how much checking it needs.

For the first few weeks, someone reads its output against the real thing: the actual reviews, the actual order, the actual complaint. The useful number to watch isn't how much the AI did. It's how often a person had to step in, and whether that number goes down as the firm adds rules.

Step 4: write one page after each try

After two or three weeks, write one page with what the team believed before, what actually happened, and the rule the firm is keeping. This is the step most firms skip. It's also the one that makes the second job safer than the first, because the rules from the first try are already written down when the second one starts.

Where to go from here

If one person runs this on one job, the firm learns a little. If the leadership team does it together, picks the job, writes the never list and agrees who checks, the firm has a way of working it can repeat on the next job. I run a hands-on workshop that takes a small leadership team through exactly this in one session, using the team's own work. You can ask about the workshop here.