The AI Literacy Gap Isn't About Training. It's About Trust.
The Training Was a Success. Nothing Changed.
A financial services firm we know rolled out AI training to roughly six hundred employees last autumn. Completion rate: 94%. Average satisfaction: 4.2 out of 5. The L&D team rightly called it one of their smoothest programmes of the year.
Six months later, someone thought to check usage of the company's sanctioned AI assistant. Fewer than one in five employees had opened it in the previous month. Of those, a handful of power users accounted for most of the activity. The rest had completed the training, nodded along, and gone back to working exactly as before.
This pattern is so common it barely registers as a finding anymore. Training completion and actual adoption have almost nothing to do with each other, and yet the standard response to low adoption is, reliably, more training.
Knowing How Was Never the Problem
The skills-gap diagnosis assumes that if people understood how to use AI, they would. It is a comfortable theory, because it implies a purchasable fix. It is also, in most organisations, wrong.
Typing a request into a chat window is not a difficult skill. Most of your employees have already done it, on their own time, with their own accounts. Microsoft's research on UK workers put unsanctioned AI use at 71% back in 2025, and the figure has not gone down since. The capability walked in the door years ago. We covered this at length in our piece on Shadow AI: the workforce is already fluent enough to be dangerous.
So the interesting question is not why employees can't use AI at work. It is why people who demonstrably use AI at home hesitate to use it at the office, or use it and keep quiet about it.
The Questions People Are Actually Asking
Sit in on enough adoption conversations and you hear the same hesitations, almost never voiced in the training feedback form.
"Will this make me look lazy?" In plenty of teams, admitting that AI drafted your report still carries a quiet penalty. The work gets a second, more sceptical read. Colleagues wonder what you actually contributed. Survey after survey finds a large share of employees concealing their AI use from managers, and they are not hiding it because they think it's forbidden. They are hiding it because they suspect it will discount their effort.
"Am I actually allowed?" Many AI policies are written to protect the company in the event of an incident, not to tell an employee on a Tuesday afternoon what they can and cannot paste into a model. The result is ambiguity, and ambiguity always resolves to one of two behaviours: abstinence or concealment. Both look like a literacy gap from the outside. Neither is.
"If the output is wrong, who carries it?" An employee who uses AI for real work is accepting a new kind of risk: errors that read fluently. Without a clear norm about verification and accountability, the rational move for anyone risk-averse is to skip the tool entirely. The people most diligent about quality are often your slowest adopters, for exactly this reason. That should worry you, because they are also the people you most want using it well.
Adoption doesn't stall because people don't know how to use AI. It stalls because nobody has told them what happens to them when they do.
Notice that every one of those questions is about trust, and that the trust runs in both directions. Employees are gauging whether the organisation will treat their AI use as initiative or as corner-cutting, and whether efficiency gains will be rewarded or quietly absorbed into higher targets. Organisations, meanwhile, signal their own distrust through surveillance-flavoured policies and approval queues. Training does not touch any of this.
The Regulation Already Assumes More Than a Slide Deck
Since February 2025, Article 4 of the EU AI Act has required organisations that provide or deploy AI systems to ensure a sufficient level of AI literacy among the staff operating them, taking into account their role and context. It is one of the Act's quieter obligations, with none of the drama of the high-risk timeline, and many companies have answered it with exactly the kind of generic e-learning module described above.
That is a thin reading of the obligation, and probably a short-lived one. "Sufficient literacy" for a recruiter using an AI screening tool means understanding what the tool does to candidates at the margins, not knowing what a token is. Role-specific competence is the standard the text points at. A 94% completion rate on a generic course demonstrates compliance activity, not capability, and the distinction will matter the first time a regulator or a litigant asks how the person operating the system was prepared for it.
What Literate Actually Looks Like
Strip away the vendor framing and AI literacy in a working organisation comes down to a small set of observable behaviours.
Calibration. Literate users know what the tool is good at for their tasks, and where it fails. They treat a model's confidence as decoration. They can say "it's strong on first drafts and useless on our pricing rules" because they have tested both.
Judgment about when not to use it. The most expensive AI mistakes we see are not bad outputs; they are good-looking outputs applied to the wrong category of problem. Knowing that the contract clause, the disciplinary letter, or the regulatory filing needs a different process is the core competence.
Verification as a habit. Not "always check everything", which nobody does, but a proportionate habit: light checks on low-stakes work, structured review where the output feeds a decision.
Open use. Literate organisations talk about how work gets done. The phrase to listen for in meetings is "I ran it through the assistant first". When that sentence is spoken comfortably in front of a manager, you have literacy. When it is whispered, you have a training certificate.
Building Trust Is Cheaper Than Buying Training
If the barrier is trust rather than knowledge, the interventions look different, and most of them cost less than another course.
Managers go first, visibly. Nothing legitimises AI use faster than a respected senior person showing their own prompts, including the failures. One director walking a team through how she drafted a board paper with AI assistance moves behaviour more than any module.
Make disclosure free. State, in writing, that AI-assisted work is judged on its quality like any other work, and that declaring tool use will never count against anyone. Then honour it. A single instance of someone being penalised after disclosure will travel through the organisation faster than the policy did.
Replace ambiguity with examples. A one-page guide with ten concrete situations ("client PII: never; anonymised meeting notes: fine; this quarter's unpublished numbers: ask first") beats a fifteen-page policy. People don't consult policies mid-task. They remember examples.
Teach by role, not by tool. A recruiter, a controller, and a developer face different failure modes and different obligations. Generic prompt training is popular because it scales; it is also why your completion rates and your usage rates live on different planets.
Measure behaviour, not attendance. Track weekly active use, the spread of users across departments, and the rate at which people disclose AI involvement in their work. Those numbers tell you whether the organisation trusts the tools and itself. Completion rates tell you whether people can click "next".
The Certificate Is Not the Capability
There is a version of AI literacy that fits neatly into an LMS, produces a dashboard, and satisfies an auditor. And there is the version where a cautious, capable employee feels safe enough to change how they work in front of their boss. The first is being purchased at scale across Europe right now. The second is what actually moves the numbers, and it is built out of management behaviour, clear norms, and a few honest conversations about what happens when the tool gets something wrong.
Train people, by all means. Just stop expecting training to do trust's job.
Building a literacy programme that goes beyond the slide deck? Our AI Literacy Curriculum Builder generates a role-specific curriculum for your organisation, mapped to how your teams actually work.