Stop buying an AI academy brochure: the three levels that actually move adoption (literacy, workflow, governance)

Someone in procurement forwards you a short email. A platform has put a new AI academy on special, the seats get cheaper if you sign this quarter, and the completion chart will look healthy by Christmas. You can sign it from your desk. The person who has to change Monday's workflow never sees that email.

Then the review arrives. The academy reports a high completion rate. You ask the business what changed and they tell you the same people still paste a brief into a chat window, take the first draft and call it a pilot. Nobody is lying. You measured who showed up. You have yet to change the work.

What the brochure can sell

A brochure sells you a wide menu, a badge and a usage chart, with a story on the cover. If you run a mid-market learning team of two, you cannot write a hundred modules yourself, and a badge gives a nervous executive something to show the board. Buy the catalogue and you have bought distribution of content. Monday's job still looks like Monday's job.

I treat adoption as a change in how a named person does a piece of work, under rules the organisation can defend. A course can support that change. A login and a completion rate only tell you that someone opened the window.

Literacy

Start with a shared language. People need to know what a model handles well, where it invents, what they must keep out of a tool, and how they tell a fluent paragraph from a useful one. Everyone who will touch the tools needs that language, from the graduate to the general manager. A brochure can carry this part, because it does not depend on one team's process.

Run it, and then keep going once people can talk about the tools in a town hall. I have sat with teams who can explain a model and still process Monday's inbox the way they did last year. They share a language. The job looks the same. Keep the session short. A half day covers most knowledge workers if you pull the examples from their own documents. They should leave having broken a model on purpose, so they have felt a confident wrong answer, and having written down what they will keep out of a tool. A long academy that never leaves the vendor's sample data is a longer tour of the same window.

Workflow

Next, pick a named workflow. The weekly customer summary. The first pass on a policy exception. The research brief before a pitch. The triage of a ticket queue. Give it one owner, a before and an after, and a rule for what a person still checks. A slogan about using the tools more often leaves you with a slogan.

A use case library can point you here and then it has to leave the work with you. You cannot copy the steps from a bank to a council to a manufacturer without the steps going bland. Sit with the operator. Pick one workflow that happens often and carries little risk. Write the new steps so that person can run them next week without a facilitator in the room.

Name the workflow in the operator's words, including how long it takes today and what a wrong output costs. Mark the steps a model can draft, the steps it must leave alone and the step where a person decides. Run it on real work for two weeks, with the same person, and keep the outputs. Write the failure you actually saw. Then keep the workflow, change it or stop it, in a sentence a sponsor can read.

Who may let it act

If you want that day built around a team's own documents, the AI training workshop is the format we use. People leave with the workflow, and with the check they agreed to keep.

Governance

After that come the decisions that let the first two spread without a nasty surprise. Who may use which tool. What data stays in bounds. When an output can reach a customer. Who carries it when the output is wrong. What an agent may do while a person is out of the loop.

Give people the language and skip the rules, and you get shadow use. A local workflow with no rules is the success legal hears about from an auditor. A policy with no shared language and no named workflow sits in a folder nobody can apply, which is how you end up with a ban and a dozen unsanctioned accounts.

Product and operations people can hold this if you give them the questions in AI governance for non-lawyers. What data does it read? Who sees the output? Can the system act, or only draft? What would we do if it were wrong in public? Those questions catch most of the trouble a team is about to ship. The NIST AI Risk Management Framework, ISO 42001 or the EU AI Act is how you file the answers once you have them. I would not start a Tuesday afternoon there.

How to buy

Before you sign, ask the vendor to sell you the levels as separate offers. If they cannot split them, buy the literacy layer and build the other two yourself. Pay for a decent common language. Keep the redesign of the work with the people who do it.

Test any proposal, internal or external, on one page. Which level is this, literacy or workflow or governance? What will a named person do differently within 30 days, and how will you see it? What data is in play, and who has already allowed it? What will you stop if the two week trial leaves the workflow untouched? Who owns the habit after the completion email goes out?

If they tell you it covers everything, that people will feel more confident, and that you will review it next quarter, you are looking at a brochure. Confidence is worth noticing. You can point at a changed workflow in the next planning round.

This month

Pick one cohort. Give them the language in a half day, using their documents. Choose one workflow with the manager who will feel it if the new steps fail. Run it for two weeks. Write the governance answers on a single page before anyone lets a tool send the output on its own.

Then look at the academy again, if you still want it. You will know which modules are worth the seats, because you will have met the work. Keep the completion chart as a record of who showed up. Judge the purchase by whether the named workflow looks different two weeks later.