Why Your AI Training Isn't Sticking (And What to Do)

Your team completed the training. They can define a prompt, name three use cases, and pass the quiz. Six weeks later almost nobody is using the tool. The reason is that you ran an education program against a workflow problem.

AI training doesn't stick because it's sequenced wrong and aimed wrong. Aimed wrong: it teaches people to operate a tool inside a job whose handoffs, approvals, and definition of "done" haven't changed. That makes using the tool extra work, done in addition to the process rather than instead of it. Sequenced wrong: it arrives after the licence purchase and before anyone has redrawn how the work moves, so the curriculum is written against workflows that are about to be obsolete.

Nobody abandons the tool out of laziness. They abandon it because you gave them a faster route and left every tollbooth standing.

The fix isn't more curriculum. It's a different order: People before Process before Tool. And a first hour that's genuinely painful.

Your training probably worked. The behavior change is what failed.

Start by separating two things your adoption dashboard has merged.

The first is whether people learned anything. The second is whether the organization now works differently. Almost every enablement program in this category succeeds at the first and is measured as if it had attempted the second.

The size of the gap is documented. Microsoft and LinkedIn's 2024 Work Trend Index, 31,000 knowledge workers across 31 markets, found that only 39% of people who use AI at work had received any AI training from their company, and only 25% of companies planned to offer training on generative AI that year. Slack's Fall 2024 Workforce Index, surveying 17,372 desk workers, found 61% had spent fewer than five hours learning how to use AI at all, and 30% had received no AI training whatsoever, including no self-directed experimentation.

Read those numbers the obvious way and you get the obvious conclusion: there isn't enough training. Build a program, fill the gap, adoption follows.

Now read the other half of the evidence.

BCG's AI at Work 2025 study surveyed more than 10,600 people across 11 countries and found that only 36% of employees felt they had been adequately trained, in organizations where regular AI use was already common. The training gap and the adoption gap are not the same gap. Companies with heavy enablement investment still report shallow, sporadic use. And Slack's earlier June 2024 index found that only 15% of desk workers strongly agreed they had the education and training necessary to use AI effectively, in the same population where 37% said their company had no AI policy at all. No rules about when AI use is appropriate, no statement of what's sanctioned.

That second number is the interesting one. It says the constraint isn't always knowledge. Sometimes it's permission. And no curriculum has ever granted permission.

Why doesn't AI training stick?

Three failures, in the order they compound.

It's aimed at the tool instead of the workflow

A standard AI enablement program teaches capability: here's what the model does, here's how to write a good prompt, here's a library of prompts for your function. All true. All useless the moment the trainee returns to a workflow that still requires the same three approvals, the same status update, the same handoff to the same person on the same cadence.

You've made them faster at producing a draft. The draft still waits four days for review, because review is where the process actually spends its time. The individual got faster. The company didn't move.

McKinsey put a number on the consequence. In its 2025 State of AI survey, out of 25 attributes tested, the redesign of workflows had the biggest effect on an organization's ability to see EBIT impact from its use of generative AI. The single strongest correlate of financial return, ahead of every governance, talent, and tooling variable in the set.

And in the same survey: only 21% of organizations using generative AI said they had fundamentally redesigned even some workflows.

Four out of five are layering AI on top of processes designed for a world without it, then running training to explain why nobody's using it.

BCG's long-standing 10-20-70 rule describes the same allocation error from the budget side: roughly 10% of AI value comes from the algorithms, 20% from the technology and data around them, and 70% from rethinking people and processes. Most enablement programs are a line item inside the 30%.

It's dosed as an event instead of a floor

The second failure is quantity, and it's more boring than the first but easier to fix.

BCG's 2025 study found regular AI usage was sharply higher among employees who received at least five hours of training and had access to in-person sessions and coaching. That combination, not either half alone. Set that against Slack's finding that 61% of workers have spent under five hours total, and the shape of the problem is clear: the median AI enablement intervention is a ninety-minute session and a link.

Ninety minutes is enough to produce recognition. It is not enough to produce a floor. A floor is the level of capability below which nobody on your team operates... the shared, unspoken baseline of what everyone assumes AI can do today. Floors are built by hours of contact, not by slides.

And the floor matters more than any individual's skill, because it's what your team's plans get written against. A team whose floor was set by a demo in early 2024 designs 2026 workflows for a tool that no longer exists.

It's sequenced after the tool and before the redesign

The third failure is the one that makes the other two inevitable.

The default order in a 20–200 person company is: buy the Tool (someone sponsors licences), then run the training (someone has to justify the licences), then eventually, maybe, if adoption is bad enough, look at the Process. People come last, or arrive only as a compliance metric.

That order guarantees the curriculum is written against a workflow map that is already stale, and it guarantees the training carries a burden it cannot hold: it's being asked to produce a behavior change that the surrounding structure actively punishes.

Which is why the second program does no better than the first.

The reframe: adoption is a structure problem wearing a training costume

Here's the diagnosis in one line.

You cannot train someone into a behavior your process penalizes.

Stalled rollouts tend to have this shape. The training was fine. The tool was fine. What nobody changed was the definition of done: the handoff, the approval, the artifact the job is actually measured on. So the rational move for every employee was to complete the work the old way and use AI on the side, in the gaps, when they remembered.

The tell is easy to spot in your own org. Ask what would happen if someone used AI to finish a task in a fifth of the time. If the honest answer is "they'd get more of the same work" or "their manager would ask them to double-check it manually anyway," you don't have a training problem. You have a structure that converts speed into nothing.

This is the same failure I traced through five different post-mortems in Why AI Rollouts Fail. Training that doesn't stick is one symptom of it, not a separate condition.

More on the underlying shape at AI-augmented teams.

So what do you do instead?

The correct order is the one almost nobody runs: People before Process before Tool. Not as a values statement. As a literal sequence, with the tool last and the training distributed across all three rather than bolted to the front of the third.

That principle is the spine of the ORBIT Framework, the redesign method behind this work, and the reason it starts where it does rather than at the licence purchase.

ORBIT is a function map before it's anything else. Five functions every AI-augmented team has to cover: Orchestrate (direct the agents toward an outcome), Run (manage the workflows the agents are inside), Build (create the systems, tools and prompts the agents operate within), Influence (drive adoption and culture change) and Translate (bridge agent outputs and human decisions).

Training lives inside Influence. Influence is the function that closes the gap between the stack existing and the stack being used: training peers on agent workflows, designing the rituals that make AI use the default rather than the special case, giving internal champions air cover, surfacing resistance before it goes underground. A ninety-minute workshop is one slice of that function, bought as if it were the whole thing. Most AI teams at the twelve-month mark are roughly 80% Build and 20% everything else, and the result has a name: the stack with no users. The agents work. Nobody uses them.

So the sequence below isn't a training plan. It's what covering the People functions looks like before a tool decision.

Here's what it means concretely, in order.

1. Replace the workshop with the first painful hour

Before any curriculum, put every person in direct contact with a genuinely capable tool on their own real work. Not a demo. Not a sandbox exercise with a fake dataset. Their actual backlog, for one uninterrupted hour, with the best tool you can put in front of them.

Call it frontier-token exposure. The mechanism is that the first hour is the only hour anyone resists. It's uncomfortable, the person feels incompetent, and everything in them wants to go back to the way that works. Past that hour, the resistance is gone and something more valuable has happened: the team's understanding of what AI can do today has become the new floor. Every subsequent conversation about process, staffing, and sequencing is now conducted against a current mental model instead of a 2023 one.

This does two things no workshop does. It raises the floor for everyone at once, and it shows you, immediately and without a survey, who your strongest and weakest operators are once the tool is in play. Both of those are inputs to the redesign you're about to do. Neither is available from a curriculum.

Enforce the hour. Don't negotiate it. The negotiation is the failure mode.

2. Sort the work before you automate any of it

Now, with a team that has a current floor, map every significant workflow into exactly one of three zones. This is the 3-Zone Co-Pilot map:

• Replace. Work where the human contribution is coordination, synthesis of known information, or routine generation. The AI does it end to end; the human defines inputs and accepts output. The common mistake is keeping a human in the loop "to be safe," which reinstalls the bottleneck you were removing.
• Augment. Work requiring judgment under ambiguity, novel synthesis, or stakeholder management. AI prepares; the human decides. The common mistake is trying to automate it outright.
• Refuse. Work where AI involvement creates legal, ethical, or relationship risk that outweighs the gain. Name these explicitly, in writing, so nobody drifts into them.

Notice what this does to your training problem. Zone 3 needs no training. It needs a written policy, which is the thing 37% of companies in Slack's survey didn't have. Zone 1 needs no training either, in the conventional sense; it needs someone to build the loop and a named person accountable for its output. Only Zone 2 needs anything resembling training, and what it needs is not prompt technique. It's judgment about when to trust the machine and when to override it.

Most enablement budgets are spread evenly across work that needed three different interventions. That's why the average is disappointing everywhere.

3. Change the definition of done before you change the curriculum

For every workflow you moved into Replace or Augment, rewrite the artifact and the approval chain to match. If the AI drafts it, the review step is different. If the AI does it end to end, the status update that existed to coordinate it should be deleted, not automated.

This is the step that converts training into behavior, and it's the step that is nobody's job on the current org chart. It's not IT's. Their mandate is systems. It's not L&D's. Their mandate is capability. It sits with whoever can change how work is measured, which in a 20–200 person company is the founder or the COO, and no one else.

Skip this step and everything upstream of it evaporates within a quarter.

4. Then train, five hours minimum, with a human in the room

Only now does curriculum make sense, because now there's something specific to teach: how the redesigned workflow runs, where the judgment calls are, what the quality gate checks.

Dose it properly. BCG's finding was five hours plus in-person sessions and coaching. Both halves. A recorded module nobody attends live produces recognition, not capability. And recognition is exactly what your current program is already producing.

Pair it with explicit permission. State in writing what's sanctioned, what requires disclosure, and what's off-limits. In Slack's data, workers at companies that had established clear guidelines were nearly six times more likely to have experimented with AI tools. Permission is cheaper than curriculum and does more.

5. Measure use, not completion

Completion rates measure whether people sat through your program. Nobody outside L&D has ever cared about that number, and rightly so.

The graduation criterion is behavioral: is the team using this as the starting point for real work? Not "did they approve of it," not "did it score well," not "did they finish it." Used. On a live workflow. This week.

Directionally, watch who is consuming the most tokens in your org. Don't turn it into a leaderboard with promotions attached. It will be gamed inside a month. Use it as a heat map that tells you who is exploring the frontier of what your tooling can do, and spend management time with them.

What should L&D own instead of adoption metrics?

Your situation is harder than it looks and more important than it's being treated.

You will be handed adoption metrics for a rollout that never mapped how work actually moves. And you will be blamed for numbers you were never given the authority to change. Enablement cannot fix a structural problem. It can only absorb the blame for one.

Two moves.

Refuse the metric, and say why in writing. Not as obstruction. As diagnosis. "Adoption of this tool requires a change to the approval chain in these four workflows; here they are; until they change, training will produce capability without behavior." That memo makes you the person who saw it first. Silence makes you the person who owned the failure.

Then take the question nobody else is equipped for. AI absorbs the junior work: the first drafts, the research passes, the deck assembly. The ladder people used to climb from junior to senior disappears with it. Nobody becomes senior by supervising a model that was already better than they were at the thing they were supposed to learn on. That's an L&D question in the deepest sense of the discipline, and almost nobody in the function is working on it. I've written up the shape of the problem in The Apprenticeship Crisis.

It's the more valuable thing to own, and it sits directly on top of the structural change described in The Hourglass Collapse. The coordination layer collapsing is also the layer most apprenticeship used to run through.

The binary

There are two versions of your next enablement cycle.

In the first, you build a better program. More modules, a prompt library, a champions network, office hours. The workflows stay exactly as they are. Six weeks after launch you're reading the same dashboard, wondering whether the content was the problem.

In the second, you accept that training was never the intervention. That what you were always being asked to do was change how the work moves, and the training was the only part of that job anyone had given you a budget for.

One of those is happening at your company this quarter. The only thing you choose is which.

Read this next: The Co-Pilot Zone... the zone sort from step 2, run in full, workflow by workflow.

---

FAQ

How do you train employees on AI to actually get them using it regularly?

Regular use is produced by three things, and curriculum is the third. First, direct exposure: every person spends an uninterrupted hour using a capable tool on their own real work, which sets a shared floor for what AI can do today. Second, a workflow change: rewrite the definition of done, the handoff, and the approval for the specific tasks you want AI in, so using it is the path of least resistance rather than extra work. Third, training, dosed properly. BCG's 2025 research found regular usage was sharply higher among employees who got at least five hours of training plus in-person sessions and coaching. Add explicit written permission: in Slack's data, workers at companies with clear AI guidelines were nearly six times more likely to have experimented with AI tools. If you do only the third thing, you get capability without behavior, which is what most programs currently produce.

How do you train your team to effectively use AI automation tools?

Sort the work before you train anyone on it. Every workflow goes into one of three zones: Replace (AI runs it end to end; the human defines inputs and accepts output), Augment (AI prepares, the human decides), or Refuse (AI involvement creates legal, ethical, or relationship risk). Each needs a different intervention. Replace work needs a built loop and a named owner, not a class. Refuse work needs a written policy. 37% of companies in Slack's June 2024 survey had none. Only Augment work needs training in the conventional sense, and what it needs isn't prompt technique; it's judgment about when to trust the output and when to override it. Most enablement budgets spread the same generic curriculum across all three, which is why the results are flat.

How do I train my team to use AI without slowing them down?

Stop treating training as a separate event that competes with delivery. The hour of direct exposure happens on real backlog work, so it produces output rather than consuming time. The redesign work happens on the handful of workflows that consume the most attention in your org, not all of them. You can do ten workflows in days. And be honest about the temporary dip: the first hour with a capable tool is genuinely unpleasant, and people will be slower during it. That hour is the only moment anyone resists. Plan for it and enforce it rather than negotiating around it, because the alternative, spreading the discomfort thinly across a twelve-week program, extends the slowdown without ever getting anyone past it.

What steps should companies take to train employees on AI usage?

In this order. One: direct frontier-tool exposure for everyone, on their own work, before any curriculum. This sets the floor and shows you who your strongest operators actually are. Two: map every significant workflow to Replace, Augment, or Refuse. Three: rewrite the definition of done and the approval chain for the workflows you moved. This is the step that converts training into behavior, and it belongs to whoever can change how work is measured, usually the founder or COO. Four: now train, five hours minimum, in person, with coaching, against the redesigned workflow. Five: publish explicit permissions. Six: measure use on live work, not course completion. The order matters more than the content: McKinsey's 2025 survey found workflow redesign had the biggest effect on EBIT impact from generative AI out of 25 attributes tested, and that only 21% of organizations using it had fundamentally redesigned any workflows.

How do you introduce something like Claude Code to an existing team without blowing up how they already work?

Introduce it to people before you introduce it to processes. Pick a small group, including non-engineers, which is the part most teams get wrong. Have each person set it up and use it on one real task, start to finish, in a single sitting. The first hour is rough. That's expected and it's the whole cost; past it, people adapt quickly and the discipline transfers across functions. Don't mandate it org-wide on day one, don't build a curriculum first, and don't route it through IT as a tooling decision. Once the group has a working floor, let them tell you which of their workflows should move into Replace or Augment. They'll have better answers than any planning exercise you could run beforehand, because they'll be answering from contact with the tool instead of from a mental model of it. Then change those workflows formally, and expand.

---

Sources

1. Microsoft & LinkedIn: 2024 Work Trend Index Annual Report (8 May 2024). 31,000 knowledge workers across 31 markets, fielded 15 Feb – 28 Mar 2024 by Edelman Data & Intelligence. 39% of people who use AI at work have received AI training from their company; 25% of companies planning to offer generative-AI training that year; 66% of leaders wouldn't hire someone without AI skills. https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part
2. Slack: Workforce Index, Fall 2024 edition. 17,372 desk workers, fielded 2–30 Aug 2024. 61% have spent fewer than five hours learning how to use AI; 30% have had no AI training at all including self-directed learning; 76% feel urgency to become an AI expert; 7% consider themselves expert users; 45% lack explicit permission to use AI. https://slack.com/blog/news/the-fall-2024-workforce-index-shows-executives-and-employees-investing-in-ai-but-uncertainty-holding-back-adoption
3. Slack: Workforce Index, June 2024 edition. 10,045 desk workers, fielded 6–22 Mar 2024, administered by Qualtrics. 15% strongly agree they have the education and training necessary to use AI effectively; 37% say their company has no AI policy; workers at companies with established guidelines nearly 6× as likely to have experimented with AI tools. https://slack.com/blog/news/the-workforce-index-june-2024
4. BCG: AI at Work 2025: Momentum Builds, but Gaps Remain (26 June 2025). More than 10,600 respondents across 11 countries. Regular usage sharply higher among employees receiving at least five hours of training with access to in-person training and coaching; only ~36% of employees say they are adequately trained. https://www.bcg.com/publications/2025/ai-at-work-momentum-builds-but-gaps-remain
5. McKinsey: The State of AI: How Organizations Are Rewiring to Capture Value (March 2025). Of 25 attributes tested, redesign of workflows had the biggest effect on an organization's ability to see EBIT impact from its use of gen AI; 21% of respondents reporting gen-AI use say their organizations have fundamentally redesigned at least some workflows. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
6. BCG: the 10-20-70 rule (10% algorithms / 20% technology and data / 70% people and processes). See The Leader's Guide to Transforming with AI: https://www.bcg.com/featured-insights/the-leaders-guide-to-transforming-with-ai

Frameworks referenced are my own: the ORBIT Framework, the 3-Zone Co-Pilot (Replace / Augment / Refuse), and the Hourglass Collapse.

---

The sequence, in one page

The enablement diagnostic... the five questions worth asking before anyone writes a curriculum, and the three-zone sort that tells you which workflows need training at all. It goes out to the list.

Get the frameworks →

Running a program right now that isn't landing? Read what a Fractional CAIO does, or book directly: book a 30-minute scoping call

---