Why Most Corporate Training Can't Show That It Worked

Companies spend a fortune on training and most of them cannot say whether it worked. They count who attended, who finished and who enjoyed it, and none of those numbers shows whether anyone does their job better.

The fix uses data most companies already hold: find the skill gap in the work itself, train against that gap, then measure the work again in the same way. Done at the level of teams and skills rather than names, it shows where training money should go without turning learning into surveillance.

The sums are large. US organisations spent $102.8 billion on training in 2025. UK employers invested £53 billion in 2024, and across the EU, training courses alone cost companies about 0.7% of their total labour costs.

Most of that spending is judged by attendance. The metric US training teams use most is the number of employees trained, and the industry’s own association says it “will not tell leaders how effective a training program was”.

In the UK, only half of learning and development teams have any process for assessing the impact of what they deliver (CIPD, 2023).

When 96 Fortune 500 chief executives were asked around 2010, 96% said they wanted to see the business impact of training, and 8% said they did.

The feedback form at the end of a course measures something else. A meta-analysis found that how much trainees enjoyed a course had almost no relationship with how much they learned (a correlation of .02) or with whether they used it at work (.07) (Alliger and colleagues, 1997).

When people rate how much they learned, the rating follows how satisfied they felt more closely than what they know (Sitzmann and colleagues, 2010). A course can score well and change nothing.

What happens after the course decides whether it sticks. A meta-analysis of what makes training transfer to the job found that motivation and a supportive workplace mattered, alongside the training itself (Blume and colleagues, 2010). Training aimed at a skill people need now, with a manager who expects to see it used, has a far better chance than a course chosen from a catalogue.

Finding the right skill is where the company’s own data comes in. A meta-analysis of leadership training found programmes built on a needs analysis improved both learning and use on the job (Lacerenza and colleagues, 2017).

Most companies already hold the evidence for that analysis in their daily work: client emails, call reviews, quality audits, error logs, complaints sorted by type. Read across a team, it shows which mistakes keep coming up. That is a better guide to spending than an annual appraisal or a manager’s impression.

When I’ve trained staff for companies as an outside trainer, I was rarely told what the team was actually getting wrong. I had to guess, and the course ended up feeling like an afterthought. If a company only wants a box ticked, that works. The trouble starts when it also expects results.

The same reading, repeated after the training, is the measure. If a sales team’s call reviews showed reps pitching before asking about the client’s needs, the next round of reviews shows whether that has changed. Using the same measure before and after turns “we think it worked” into a number the business recognises, and if nothing changed, the company finds out early and stops paying for it.

Where a company can, it should train one team first and read a similar team’s work alongside it, so the change can be put down to the training and nothing else.

The problem is that data about people’s work has to be handled with care.

What is read is a sample of work already reviewed for quality. It is read for the mistake and not the person, and reported only as a pattern across a team. Staff are told what is read, and why, before it starts. Data at Work sets out the reasons.

None of this needs names. Monitoring individuals raises stress slightly and does nothing for performance (Siegel and colleagues, 2022).

People who fear judgement also stop owning up to the mistakes training is meant to fix (Edmondson, 1999).

So the data is read by skill and by team, and every group in a report has to be large enough that nobody can be picked out. Survey platforms set a default minimum of five, and the UK’s data regulator warns that a colleague can often work out who a small-group figure is about.

Learners see their own progress. Managers see patterns. Nothing goes into a performance review.

A report built this way might read: across the customer service team, replies missing a clear next step fell from four in ten to one in ten over a quarter, and the remaining cases cluster in refund requests. It tells the company what improved, where to put the next round of training, and nothing about any one person.