Why reskilling programmes mostly fail

Because they measure completion and hope it means capability, and the sector has known these are different things for at least thirty years. A completion rate tells you someone finished a module. It tells you nothing about whether they can now do anything they could not do before, and organisations keep buying the first because it is the only one that arrives as a number.

This is a disagreement page. Reskilling is the near-universal institutional answer to AI, endorsed by every major report, and the argument here is that most of it does not work and that the reasons are structural rather than a matter of trying harder.

Five structural reasons#

1 · Completion is the metric because capability is hard to measure. Not because anyone believes completion matters. It survives because it is available, reportable and defensible, and because nobody is asked for the other number. Everything downstream follows from this one substitution.

2 · The training removes the difficulty that would have produced the learning. Well-designed courses feel smooth, and smoothness is the enemy of retention. Bjork's work established that conditions raising performance during study frequently lower long-term learning, and corporate learning is optimised almost entirely for the study experience, because that is what gets rated.

3 · There is no practice afterwards. A skill taught and not used decays, and reskilling is almost always followed by a return to the same work. Ericsson's deliberate practice requires effortful activity at the edge of ability, with feedback, sustained over time. A two-day course followed by nothing is not that, and nothing about calling it reskilling changes the mechanism.

4 · It is aimed at tools, which depreciate fastest. Most AI reskilling teaches interfaces and prompting. Prompting is not scarce, not durable and not the constraint, and interface knowledge has a shelf life measured in months. See why "learn to prompt" is weak career advice.

5 · The organisation is simultaneously removing the work the skill applies to. This is the contradiction nobody says out loud. Firms automate the tasks that build judgement and run a programme to build judgement, in the same quarter, funded from different budgets, with neither party talking to the other.

Rahim Hirji put the same point in one paragraph on an earlier version of this site: "Faced with a talent pipeline problem, the typical corporate response is ‘we will train our people more’. Training alone cannot fix this. Workshops, courses and bootcamps transfer knowledge; they cannot replicate the learning gained through real work over time. Most professional development comes from doing: stretch assignments, first drafts, mistakes corrected by someone senior. If those opportunities dwindle, no amount of classroom instruction fills the gap."

The evidence, stated honestly#

The strongest support for this argument is indirect, and saying so is the point of grading evidence.

Bastani and colleagues showed that when the interface did the work, performance rose while capability fell, and that a guardrailed design largely removed the harm. Same content, opposite outcome, decided by whether the learner had to do the effortful step. That is a direct demonstration of reason two, in education rather than corporate learning.

The Vaccaro meta-analysis and the deliberate-practice literature support reason three. The PwC data showing skills requirements changing 66 per cent faster in the most AI-exposed jobs supports reason four, since a curriculum built for last year's tools is already behind.

What does not exist: a body of evidence measuring whether corporate reskilling programmes change unaided capability. Completion rates are published constantly; capability change almost never is. That absence is itself the finding. It is the reason this page is an argument rather than a report.

The strongest objection#

Some reskilling clearly works, and the good version is recognisable: it is embedded in real work, spaced over months, involves doing rather than watching, and is assessed by whether the person can now do something. Apprenticeship has worked this way for centuries and remains the most reliable capability-building technology anyone has built.

So the claim here is narrower than "reskilling fails". It is that the dominant form, procured as content, delivered as modules and measured by completion, mostly fails, and that it dominates because it is cheap, fast and reportable rather than because anyone believes in it.

A second fair objection: sometimes the programme is about signalling that the organisation is doing something, or about legal and regulatory cover, and capability barely enters into it. Article 4 of the EU AI Act will accelerate exactly this, since it creates an obligation that a certificate appears to satisfy. That is a legitimate purpose, and it should be named rather than dressed as capability building. See what AI literacy means for leaders.

What would change my mind#

Published data from a large employer showing measured capability change, not completion or satisfaction, sustained six months after a programme, in work the organisation was simultaneously automating. Nobody has published it. If someone does, this page changes and the change gets dated in the corrections ledger.

Measure whether they can do it unaided#

• Measure one thing: can they do it unaided? Before and after, on real work. Everything else is proxy.
• Protect the practice before buying the training. If the work that would exercise the skill is being automated, the programme is refilling a bucket with a hole in it.
• Teach judgement, not interfaces. What survives is knowing when the output is wrong, which is domain expertise applied, not tool familiarity.
• Space it and embed it. Months of real tasks with feedback beats days of content, and usually costs less.
• Name the purpose honestly. If this is compliance cover, call it that. Compliance cover measured as capability is how organisations end up believing something that is not true about their own people.

Related SuperSkills research#

On the mechanism, desirable difficulty and the missed reps. On measuring properly, usage theatre. On what erodes underneath, capability debt. On the workforce plan, AI workforce strategy and the CHRO guide.

Key sources

• Bastani, H. et al. (2025). Generative AI can harm learning. PNAS.
• Ericsson, K. A. et al. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3).
• Macnamara, B. N. and Maitra, M. (2019). The role of deliberate practice in expert performance: revisiting Ericsson. Royal Society Open Science, 6(8).
• PwC (2025). Global AI Jobs Barometer.
• Bjork, R. A. and Bjork, E. L. Desirable difficulties in theory and practice.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. This is a position page arguing against the near-universal institutional answer. The supporting evidence is indirect, which the page states rather than obscures, and what would change the position is set out above. He also sells advisory work in this territory, which is a direct commercial interest in the argument and is disclosed in how this research works.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.