Corporate Training/L&D is Not Prepared for AI

Matt Beane has a long article pointing out that existing corporate training programs (or Learning and Development, i.e., L&D) are not really equipped to deal with the challenges of how AI is changing the workforce and how fast AI itself is changing.

The first part of this article is about why the usual way of doing training doesn’t work. There is a second part describing what actually works.

Corporate L&D doesn’t work:

Because the data had been in for almost a lifetime: we rarely measure the behaviors associated with training (it’s costly to do) and when we do training doesn’t move the needle.

and

Every L&D leader in every corporation I talk to knows this. Admits it freely. So does every leader with a P&L. About 20 years ago research settled this question, and it keeps relitigating it with each new learning paradigm. Same result. And yet: firms still spend their money on training as if it caused results.

and AI training is worse. Why?

So whatever formal training you went through on AI, if it was last week it’s out of date.

Companies have been spending on AI training, but some of them have begun seeing the problem with AI training:

What has all this sheep-dip AI training gotten us? Familiarization? Sure. Ask a CTO what effect that had: Logins? Nope. That took fiat. Daily use? Nah. Fiat. Token burn? Same deal. I bet you agree, don’t you? Ask around. Does anyone disagree? An increasing chorus of executives are with us, in fact. Wharton’s excellent 2025 AI report shows that confidence in training as the path to AI fluency fell 14 points and overall training spend went down 8 points. At the same time “we’ll just hire new talent instead” went up by 8 points. A clear and growing vote of no-confidence in training.

Ok. So what works?

To keep up, you have to regularly break your mind on this stuff by accomplishing impossible work in tiny amounts of time. That’s it. That’s the only way.

or

I’d guess I burn a minimum of 5% of my time a day keeping up. Then come two day intensives when a logjam of new capability collides with an irritation or an idea, and I burn 40% of my day building, testing a new thing while trying out the new stuff.

But this suffers from the 5% problem: most people do not have the drive and discipline to actually do this on a sustained basis.

So you need to do “training”, but it has to be structured differently.

In a recent post, Steve Yegge describes a training study/experiment at Netflix which found the following things:

First, among those learning AI, there are 3 kinds of people:

1. No agent: Those not using agents (like Claude Code or Codex CLI) at all for their regular work. (They might still be using the web based interface for some research, or writing reports or emails)

2. Single agent: Those using a single agent, synchronously, throughout their work day

3. Multiple agents: Those using 2-4 agents without watching

It takes a specific kind of training session to move someone from no agent to single agent or from single agent to multiple agent. What is needed is this:

The training must be done a team at a time, with 5 to 10 people, including their manager. The manager must opt the team in, during regular work hours, as “blessed” company time. The trainees must bring their actual work, and the instructor(s) will help them learn how to do it with agents.

The full team must be there, the manager must be there, they must bring their actual work, and it should be at least five hours, and the instructor must be someone who uses multiple agents for their own work on an ongoing basis.

Skip any one of these, and it doesn’t work:

They found that if they cut corners anywhere — shorter classes, larger audiences, individual opt-in classes — they didn’t get the same results. It didn’t “stick.”

But if you do it right, then it is really helpful because, here’s an important point: 5 hours of training is enough to switch a person from avoiding AI to actively learning by themselves:

Fortunately, Ezra has good news for you there: People can jump cohorts in 5 hours. That’s how long it takes people, in the right training setting, to graduate from AI illiteracy to AI savviness. And they stay there. It’s like flipping a switch. 96% of the trainees remained in the second cohort for six weeks after the course without showing signs of slowing down.

There’s a lot more detail to this in Steve’s post (note: Steve’s posts are long, so you might have to skim the earlier parts and get to the “AI Literacy, Beginner Cohorts” section).

I have myself done some sessions like these where people bring their own work and get them to a point where they start using an agent (Claude Code or Codex CLI) and accomplish something non-trivial in 3 to 5 hours (e.g., build an app that they want in their own work). And the results have been amazing: I regularly hear about something amazing they’ve done, often something they have wanted for years but could not.

If you’re not convinced that corporate training is the wrong way to do AI training, read Matt’s full post. And if you’re convinced, and would like to understand the right way to do this training, read Steve’s full post.