Why AI adoption stalls outside engineering, and what fixes it

Most companies that rolled out AI broadly in the last year tell the same story. Engineering adopted it inside a week. Finance, legal, operations and sales tried it, found it useful for the occasional email, and drifted back to how they worked before. The licences are paid for. The dashboard says half the company logged in last month. Nobody can point to a process that changed.

The tool was designed for the wrong user, and that is fixable.

The blank box is the problem

A chat window asks the person using it to design the workflow. What goes in, what steps to take, what a good answer looks like, how to check it. Engineers already think that way, because writing software is that. So they turn a blank box into a tool in an afternoon.

A credit controller or a paralegal does not think in inputs and outputs. They think in the job. They open the box, ask it something reasonable, get an answer that is nearly right, and quite sensibly decide it is faster to do the work themselves. After three of those, they stop opening the box.

So the people closest to the code get most of the value, and the ninety percent who are not engineers get a slightly better email drafter. The adoption numbers flatten at that line.

Three moves that unstick it

Start with the workflow, not the tool

Pick one team that is buried in repeat work. Sit with the people who do it and build three to five workflows end to end, rather than handing over a prompt library. A contract review that checks the clauses your legal team cares about. A ticket triage that routes the way your support lead routes. An invoice match that knows your suppliers.

Each one takes a day or two to get right, and each one is something a non-technical person can run without designing anything.

Package what works so it travels

The person who figured out a great workflow is usually one person on one team. If it lives in their chat history, it dies when they go on holiday. It needs to become a shared skill, reviewed by someone who owns that process, and available to everyone who does the same job.

Most rollouts leak here, because the best work stays private when there is no place to put it. Skill sharing turns one person’s discovery into the team’s default.

Measure use per team, not seats

A seat count tells you what you paid for, not what happened. Track weekly active use by team, against the workflows you built. If a team has the skills and is not using them, the skills are wrong. If a team has no skills yet, that is the next place to build.

This also tells finance something they can act on. Teams with real adoption justify frontier models. Teams still at the email-drafting stage do fine on fast, cheap ones. Routing by team is how AI spend stays predictable while adoption grows.

What this looks like in practice

What we see work is a short, hands-on build with one team, followed by that team teaching the next one. The first workflow into production is the hard part. After that, people ask for the second. That is the shape of the implementation we run.

If your rollout has plateaued, non-technical teams were most likely handed a box and asked to be engineers. Book a call and we will map the first three workflows with you.

Common questions

Why does AI adoption stall outside technical teams?

Because a general-purpose chat window asks the user to invent the workflow. Engineers already think in inputs, steps and outputs, so they do. Everyone else opens a blank box, tries something, gets a mediocre answer and concludes AI is not for their job. Adoption stalls at the people who were never shown a task that works.

How do you get non-technical teams to use AI?

Start with three to five real workflows per team, built end to end with the people who do the work, then packaged as reusable skills anyone can run. Once a team has one workflow that saves them an hour a week, they find the next one themselves. Training sessions and prompt libraries rarely get there on their own.

What is a good AI adoption rate for a company?

Measure weekly active use per team rather than a company-wide seat count. A seat that logs in is not adoption. A team where most people run at least one AI-assisted workflow every week is. The gap between the two is usually where the budget is being wasted.

Should we buy more AI training?

Training helps people who already have a workflow to practise. It does not create the workflow. The teams that plateau after training are the ones that were taught how to prompt but never given a task worth prompting for. Build the task first, then train around it.