Training your people will do more for you than buying another AI tool

Last week Bloomberg reported that the crypto exchange OKX is making AI skills part of how it evaluates its workers (Bloomberg). It isn’t the first company to do this. KPMG said last fall that from 2026 its staff would be rated in annual reviews on how well they meet the firm’s AI objectives (International Accounting Bulletin).

I understand the instinct. You’ve paid for the tools and you want people to use them. But grading people on a skill you haven’t taught them gets the order backwards. My read is that most AI budgets have gone to licenses first and to teaching people second.

Half the workforce is teaching itself

A survey released two weeks ago puts numbers on the gap. Cornerstone, which sells workplace learning software, surveyed 2,000 employed adults in the US and UK in April (Cornerstone). Read it knowing the company has a stake in the answer. The numbers are still worth your attention:

• 46% of respondents use AI tools at work with no formal training from their employer.
• 47% learn AI through trial and error.
• 36% limit how much they use AI because they’re afraid of making mistakes.
• 17% pretend to use it.

That last number is the one that should bother a CEO. If you’ve told your team AI is a priority and given them no help, some of them will fake it. Your usage dashboard will look fine. The work won’t change.

DataWhat workers say about learning AI

• Use AI with no formal training46%
• Learn by trial and error47%
• Limit use to avoid mistakes36%
• Pretend to use it17%

The same survey found 75% of workers believe leadership has identified the AI skills the company needs, but only 33% say those skills have made it into training programs. Leaders know what they want. It isn’t reaching the people who do the work. The plans exist on paper, too: 65% said their employer has an AI upskilling plan, but only 36% said it was well communicated, and 56% said they have no clear path to build the skills.

A plan nobody has heard of doesn’t train anyone.

Another tool won’t fix this

When AI isn’t paying off, the easy move is to buy something else. A new assistant, a new agent platform, a new add-on to the software you already have. Every one of them comes with a good demo.

The problem is rarely the tool. A person who doesn’t know how to break their work into steps, write clear instructions or check output for errors will get mediocre results from any tool. A person who does know those things will get good results from the one you already pay for.

Buying more tools also adds to the pile people have to learn. Every new product means another login, another interface and another set of habits. If your team hasn’t mastered the first tool, a second one splits their attention further.

There’s a cost argument too. A license is a recurring line item that pays off only if someone uses it well. Training is mostly a one-time cost in hours, and it pays off across every tool the person touches afterward, including the ones you haven’t bought yet.

Train on their own work

Too much AI training is generic: a webinar on prompting, a library of video courses, a slide deck of use cases from other industries. People sit through it, nod and go back to doing their jobs the old way.

Training that changes behavior uses the person’s own tasks. Here’s the shape of it.

LoopTraining on real work

1. 1Pick two tasks from your week
2. 2Do them with AI
3. 3Compare time and quality
4. 4Manager reviews the output
5. 5Share what worked

A customer support rep practices on last week’s tickets. An accountant practices on this month’s reconciliation. A recruiter practices on the job descriptions open right now. They do the task with AI, compare it to how they’d normally do it, and a manager checks the result.

The manager’s check matters as much as the practice. They should look at three things: whether the output is correct, how long it took compared with the usual way, and whether the person can explain what the AI did and what they changed. That last one separates someone who uses AI well from someone who pastes whatever comes back.

Say you run a 30-person agency. For two hours a week over a month, each person brings two tasks from their own week and works through them with AI, with a lead checking the output. By the end, everyone has a handful of tasks they now do faster, and the lead has seen with their own eyes who can actually do it. That’s a better basis for a performance conversation than a usage count.

Give people the time

The second thing companies skip is time. If learning AI happens on top of a full workload, it happens at 9 p.m. or not at all. Set aside paid hours for it, on the calendar, the way you would for any other required training.

Do the arithmetic before you decide it’s too expensive. Thirty people at two hours a week for four weeks is 240 hours. That’s six weeks of one full-time person’s time. If each person comes out of it saving even an hour a week on their own tasks, you’ve earned it back within two months.

Then decide what good looks like. Write down, for each team, what a well-done AI-assisted task looks like and how it gets checked. People who limit their AI use out of fear of mistakes need to know where the line is. A clear standard gives them permission to try.

Build the same thing into onboarding. A new hire’s first month is when habits form. If part of that month is spent learning to do the core tasks of the role with AI, under a manager who checks the output, you never have to retrain them later. The people you hire this year will set the standard for the people you hire next year.

Only after that should AI skills show up in a review. Grade people on something you’ve taught them, given them time to practice and defined clearly. Otherwise you’re measuring who was willing to teach themselves at night.

Grade people on AI skills only after you've taught them, given them time to practice and told them what good looks like.

What to do this week

• Count your AI licenses and ask each team lead how many people on their team could show you one task they now do better with AI.
• Put a freeze on new AI tool purchases for 60 days unless they replace something you already pay for.
• Block two paid hours a week for the next four weeks for each team to practice AI on their own tasks.
• Ask each person to bring two real tasks from their week to the first session.
• Have each team lead write a one-paragraph standard for what good AI-assisted work looks like on their team.

Sources

• Crypto exchange OKX makes AI skills a part of worker evaluations (Bloomberg, May 26, 2026)
• KPMG to include AI performance reviews (International Accounting Bulletin, November 3, 2025)
• Mind the gap: New Cornerstone AI skills study reveals half the workforce is “winging it” to learn AI due to low employer readiness (Cornerstone, May 20, 2026)