How to Run an AI Training Program That Changes Behaviour

Most internal AI training teaches tool features and changes nothing. The programmes that work are built around the tasks people actually do, delivered with a sanctioned tool they can use the same afternoon, and measured by whether work changed. This playbook covers the sequence, drawing on FISTA Solutions' AI enablement work.

When is this worth doing?

When people are already using AI unofficially, when a sanctioned tool exists or is about to, or when a capability programme has an outcome attached to it.

It is not worth doing before a sanctioned tool exists, because the immediate result is enthusiasm with nowhere to go. It is also not worth doing as a general awareness exercise: awareness is high already, and what is missing is applied skill on specific work.

What does the sequence look like?

Step	Purpose
1. Segment by task	Group people by what they actually do
2. Gather real material	Collect the documents and questions they work on
3. Provide the tool	Sanctioned access before the first session
4. Teach on their work	Sessions built from the gathered material
5. Establish champions	Someone in each team who continues it
6. Measure and iterate	Behaviour change, not attendance

Step 1 — Segment by task, not by department

Group people by the work they do rather than where they sit. A contracts lawyer, a procurement analyst, and a compliance reviewer all spend their days reading long documents and extracting specific facts, and they need the same session even though they report to different functions.

Department-based segmentation produces sessions where half the room has no use for the example, which is the fastest way to lose an audience. Task-based segmentation produces sessions where everyone recognises the problem.

The practical method is asking a sample of people to describe what they spend most time on, then clustering the answers. That exercise takes a few days and it improves everything downstream.

Step 2 — Gather the real material

Collect actual documents, tickets, questions, and outputs from the audience's work — anonymised where necessary — before designing anything.

This is the step most programmes skip, and it is the one that makes the difference. Sessions built on generic examples teach generic skills; sessions built on a contract the audience actually reviewed teach them something they can use that afternoon.

It also surfaces where AI will not help, which is valuable. Some of what people spend time on is judgement that should stay human, and saying so in the session builds credibility for everything else.

Step 3 — Provide the sanctioned tool first

Access has to exist before the first session. Training people on capability they cannot use is worse than not training them, because it drives them to consumer tools on personal accounts — exactly the shadow usage the programme should reduce.

That means the procurement, security review, and access provisioning have to complete first. If they will not, delay the training rather than running it and hoping.

The tool also has to be good enough to prefer. People route around sanctioned tools that are slower or more restricted than the alternative, and no amount of training changes that calculation. See what is shadow ai.

Step 4 — Teach on their work, with time to try

Run sessions where people work on their own material with support available. Demonstration followed by guided practice beats demonstration alone by a wide margin, because the difficulty is never understanding what the tool does — it is getting a useful result on a real input.

Budget more time than feels necessary for the practice portion. The moment that changes behaviour is when someone's own difficult document produces a useful result, and that moment needs someone nearby when it does not work first time.

Cover the failure modes explicitly: what the system gets wrong, how to check it, and when not to use it. People who know the limits use the tool more, not less, because they trust their own judgement about when to rely on it.

Step 5 — Establish champions in each team

Identify someone in each team who is genuinely interested and give them a bit of structure: a channel to ask questions, occasional time with the programme team, and permission to help colleagues.

Champions sustain what training starts. A single session decays within weeks without someone nearby who can answer the next question, and a central team cannot be nearby to everyone.

Choose them by interest rather than by seniority. The person who has already been experimenting is a better champion than the manager who was assigned the role.

Step 6 — Measure behaviour, then iterate

Measure what changed in the work: time per task, volume handled, output quality, and whether people are still using the tool six weeks later.

Attendance and satisfaction measure the session. They are worth collecting and they are not evidence that anything changed, and reporting them as if they were is how programmes lose credibility with the people funding them.

Then iterate on what the measurement shows. Sessions that produced no change need redesigning rather than repeating, and the reason is usually that the material was not close enough to real work.

What about review capacity?

Training raises output volume, which moves the constraint to review. A team that can now draft three times as much still has the same capacity to check it, and the result is either a backlog or unreviewed work going out.

Address that explicitly in the programme: teach people to check outputs efficiently, establish what needs review and what does not, and be honest that some tasks should not be accelerated. Programmes that raise generation without addressing review produce a quality problem and get blamed for it.

What about people who are resistant?

Some resistance is well founded. People who have seen a tool produce confident errors in their domain are right to be cautious, and treating that as a change management problem to overcome damages credibility.

Engage with the specific concern. If the system is bad at their task, say so and focus on something else. If the concern is about job security, acknowledge it honestly rather than deflecting — evasion is more corrosive than a difficult answer.

The people who convert are usually persuaded by a colleague's result rather than by a trainer's argument, which is another reason champions matter.

Who needs to be involved?

A programme owner with authority to prioritise, subject matter people from each task group, whoever owns the sanctioned tool, and champions in each team.

It does not need a large central team. It does need someone senior enough to say that a task group's material is worth gathering and that managers should release people for the sessions.

How long does it take?

Segmentation and material gathering take one to two weeks. Sessions run over a few weeks depending on audience size. Measurement needs six to eight weeks after delivery before the numbers mean anything.

The mistake is measuring immediately, when enthusiasm is high and behaviour has not settled.

What are the common failure modes?

Running before a sanctioned tool exists. Teaching features rather than tasks. Generic examples. Measuring attendance. No champions, so the effect decays. And ignoring the review capacity constraint the programme creates.

How do you know it worked?

Sustained tool usage weeks later, measurable change in time or volume for the targeted tasks, champions answering questions without central involvement, and a reduction in unsanctioned tool usage.

What does it cost?

Mostly people's time rather than tooling. The expensive version is the one that stalls halfway and leaves the organisation with neither the old state nor the new one, which is why a narrow first pass beats a comprehensive plan nobody finishes.

Budget the work as an operated change rather than a project with an end date, because most of these need a maintenance tail. See AI total cost of ownership.

What should you do first?

Ask twenty people across different functions what they spend most of their time on. Cluster the answers. That list is your segmentation and your curriculum, and it takes a few days to produce.

How FISTA Solutions helps

FISTA Solutions runs this work alongside client teams rather than around them: curriculum built from your own documents and tickets rather than generic examples, sanctioned tooling in place before the first session, evidence produced as the work proceeds, and handover that leaves your people able to continue without us. Delivery runs through AI agents, AI enablement, and forward deployed engineers. The record is 150+ projects for 50+ companies across 12+ countries, with 47% average efficiency gains where measured.

To run this with support, message FISTA on WhatsApp, or read AI training program cost.