Ask an agency owner whether the team uses AI and the answer is almost always yes. Ask how, and it gets vaguer. Someone has a personal ChatGPT account. A couple of people pay for Claude themselves. Copilot arrived with the Microsoft licence and gets opened now and then. Nobody is doing anything wrong. Equally, nobody has been shown what good looks like for their own job.

A piece published by CIO.com on 2 October puts a useful frame around this. Writer Shree Amujala describes four stages of AI maturity in small businesses. The first is shadow AI, where people use personal accounts with no oversight. The second is licensed but unstructured: seats have been bought and guides handed round, but there are no repeatable workflows. The third is structured enablement, where AI is tied to specific daily tasks with clear boundaries. The fourth is governed and embedded, where it connects to existing systems with permissions and audit trails.

Her observation is the part worth keeping: "Most demand I see today is the move from stage two to stage three. That transition is a services problem. A license can be purchased in minutes."

Two cautions. It is an American piece and the statistics it cites are US or vendor figures, so the argument is worth taking and the numbers are not. And the four stages are one writer's frame, not a research finding. They still describe what most agency owners will recognise.

Stage two is where agencies tend to stall, and the reason is not reluctance. Client work is deadline-shaped. Nobody has a spare Tuesday afternoon to rework how briefs get summarised or how first-draft research gets done, so the licence sits there and a few keen people carry on alone. Harvard Business School's Joseph Fuller told the Harvard Gazette on 29 September: "Companies are not training people formally, so employees are basically using it like glorified Google." He adds that only about a third of firms' AI experiments succeed, partly through poor implementation and thin training. That is economist commentary on firms in general, not a study of agencies, but it fits.

What does stage three look like in practice? It is smaller than most people expect. Take an account manager who writes weekly client updates. Using Claude for emails is not a way of working. A way of working is four lines on a shared page: the notes and tracker extract that go in, the brief given to Claude, what the account manager checks before anything goes out, and what never goes in at all, such as a client's confidential figures. Do that for a handful of recurring jobs and an agency has structure without a programme or a new platform.

It tends to move faster when someone outside does the first stretch. An owner who is also running the agency rarely has the room to find out where each person actually is, pick the right jobs and stay with the first few attempts. A short, honest outside view helps with all three.

That is how the work went with one agency I worked with, a global events and film business of about 36 people. Go Wisely began with a readiness diagnostic across six dimensions, so the sessions were designed around where the team really was, not a generic syllabus. Then came six bespoke sessions built on the team's own work, using Microsoft Copilot because that was their environment. Satisfaction averaged 4.6 out of 5 across 131 survey responses, and not one response was dissatisfied. Satisfaction is not the same as impact, and it would be wrong to claim more. But people who feel a programme was built for them are the ones who keep using what they learned.

So the useful question this week is a quiet one: which stage are we honestly in? If the answer is two, that is the most common place to be and it is not a failing. Pick one recurring job, write down how it should go with AI, and see what that tells you. The tools will keep changing. Deciding how work should get done, and making sure each person has been shown, will not.

Go wisely.