Training beats tools for AI adoption

Training gets more staff using AI than handing out a tool does, and it works best when staff know their employer expects them to use it. In a study of about 25,000 Danish workers, 57% had used AI chatbots at work where the employer only supplied a company tool, 69% where it only trained people and 93% where it encouraged use, supplied a tool and trained staff¹. BCG's 2025 survey found regular use far more common among workers with five or more hours of training². No randomized trial has yet isolated what training adds on its own.

What the evidence says about AI adoption training

Economists Anders Humlum and Emilie Vestergaard surveyed about 25,000 Danish workers in 11 AI-exposed occupations, from accountants to software developers, and linked the answers to government records¹³. They grouped workers by what their employer had done about chatbots: encouraged use, paid for a company version, trained people, some mix of those, or nothing¹.

About 41% of workers had used a chatbot at work even where the employer did nothing, so plenty of staff try these tools unasked¹. Encouragement on its own did more than a tool or training on its own, and among the pairs, encouragement plus training came closest to the full set¹.

Daily use moved less. It went from 7% with no steps to 18% with encouragement alone and 28% with all three¹. Training alone also came with smaller reported time savings, which the authors say fits programs built to prevent misuse, since trained workers often reported new ethics and compliance tasks¹. The authors call these patterns descriptive¹.

Why AI rollout adoption stalls with access alone

The cleanest test of access on its own is a randomized trial across 66 firms and 7,137 knowledge workers, run by three Microsoft researchers and a Harvard Business School economist⁴. Workers picked at random got Microsoft 365 Copilot inside apps they already used, and over 90% tried it at least once in six months⁴.

Whether they kept using it depended on the firm. Over the trial's last three months, average weekly use ran from 6.3% of the workers given Copilot at one firm to 70% at another⁴. The firm explained nearly twice as much of the variation as people's own past work habits, and the authors suspect management practices or training they couldn't measure⁴.

BCG's 2026 survey of regular AI users at large companies points the same way. Of users with a clear AI strategy but limited tools, 80% reported measurable impact, against 60% of users with strong tool access but little clarity⁵. For what happens to licences after the first few weeks, see why AI seats go quiet.

Employee AI adoption rises when managers back it

Gallup sees the same pull in the U.S. Where organizations have adopted AI, employees who strongly agree their manager actively supports their team's use are 1.7 times as likely to use it a few times a week or more⁶. In BCG's 2025 survey, 82% of frontline employees with clear leadership support were regular users, against 41% without it².

The UK Department for Business and Trade saw the reverse in its 1,000-licence Copilot pilot: staff used the tool less when colleagues and line managers were hesitant, though the evaluators record that as anecdotal⁷.

AI workshop vs demo: what learning research says

Researchers have compared watching with doing for decades. None of these studies involved AI tools, so applying them to AI training is an inference.

A generic demo looks a lot like the lecture and the re-reading. In the UK Copilot pilot, staff saw the departmental and Microsoft sessions as introductory and found their own practice more useful⁷. People who spent at least two hours on self-led training were more satisfied with Copilot, though the usage data couldn't identify individuals⁷.

What hands-on AI training looks like

BCG's 2025 survey of 10,635 workers in 11 countries and regions compared regular use, meaning AI use daily or several days a week, with how much training people had².

Microsoft and LinkedIn's 2024 survey of 31,000 knowledge workers found frequent experimenting was the top predictor of becoming a power user, and power users were 35% more likely to have had training for their own role¹².

Practice should build understanding as well as output. In an Anthropic trial with 52 mostly junior engineers learning an unfamiliar code library, those who asked the AI for explanations scored 65% or more on a quiz afterwards, and those who handed it the coding scored under 40%¹³. Those were small groups, and Anthropic sells AI assistants.

What the evidence doesn't show yet

No randomized trial has yet measured what an AI training program does to productivity or pay. The figures above are associations, and firms that train staff may also manage change better. In Denmark, none of the employer steps, all three included, produced a detectable change in earnings or hours through 2024¹³.

The samples also lean large. BCG's 2026 respondents work at companies with US$100 million or more in revenue and include no Canadians⁵, and the Copilot trial drew large firms keen to adopt early⁴. Many staff teach themselves: in the Microsoft and LinkedIn survey, only 39% of AI users had training from their company¹². Why staff training decides the result covers that gap.

How to get employees to use AI at a 10 to 500 person firm

The evidence favours doing all three things the Danish employers did, with training built around each team's real work. Check where your team sits first.

1. Week 1: the owner or each team lead tells staff in writing which tasks to use AI for, which tool to use and what client data stays out. Encouragement alone went with 76% use, against about 41% with no steps¹.
2. Also in week 1, train managers in a session of their own. Workplace training carried over best when trainees' superiors were trained too¹⁰.
3. Each team lists two or three tasks it repeats every week, like the Monday report or meeting notes. Practice on trainees' own scenarios carried over best¹⁰.
4. Weeks 2 to 4: give each person at least five hours of practice on the approved tool, in small in-person groups, with a named coach². Open with one worked example on a real document, then have everyone do the next one¹¹.
5. Start later sessions by having people redo last week's task from memory, and teach them to ask the tool to explain its work⁹¹³.
6. Before you start, count weekly active users per team and time two repeat tasks. Check both at 12 weeks, since weekly use in the Copilot trial took about that long to settle⁴.

The readiness check takes about three minutes and shows where your team stands on goals, people, workflows, tools and rules before you plan the sessions.