AI can automate the task. Can it automate the learning?

You’re reading The AI Directive — practical AI intelligence for business leaders. Subscribe here if someone forwarded this.

A conversation in my kitchen has stayed with me. One of my children was working through a maths problem, and the answer was only a search away: AI could produce it instantly and explain each step.

We ended up talking not just about whether the answer was right, but whether getting it meant the principle had been understood. The homework could be completed, but what had my child actually learned? That question stayed with me, because something similar is now happening in business.

AI can assemble research, analyse data, write code, prepare a first draft and recommend the next step. The employee receives a strong-looking output faster than they could have produced it themselves, but have they built the understanding needed to explain, challenge or reproduce the answer?

So I became curious and looked at the emerging evidence. One theme kept returning: AI value depends not only on technology, but on an organisation’s ability to learn, adapt and embed new ways of working. That creates a tension: organisational learning may make AI successful, yet some uses of AI may remove the practice through which that learning is built. What might happen if AI improves today’s output while quietly weakening tomorrow’s capability?

Most AI measurement begins with the work: time saved, cases handled, quality improved or cost removed. Those are legitimate measures, but they describe the output of the system today, not necessarily what the human will be capable of doing tomorrow.

Consider a junior analyst investigating a fall in product margin. Gathering data, testing assumptions and revising an incorrect diagnosis can be slow, but it is also how they learn which data to trust, which questions reveal the issue and where exceptions hide. If AI produces the analysis in minutes, the organisation gains a faster answer while the analyst may lose the attempt, error and correction through which judgement develops.

Nobody should preserve hours of spreadsheet drudgery because suffering is character-building. But useless friction and useful practice are not the same thing.

The most direct evidence I have found comes from Anthropic itself. In a randomised controlled trial involving 52 mostly junior software developers, the AI-assisted group scored 17 percent lower on a test of the concepts they had just used. This was a small coding study measuring immediate mastery, not proof that AI inevitably weakens capability.

More usefully, participants who asked for explanations and combined assistance with their own coding learned more than those who simply delegated production. The important variable may not be whether somebody uses AI, but how the interaction is designed.

Microsoft Research surveyed 319 knowledge workers and found human thinking shifting from gathering and producing information to verifying, integrating and stewarding AI output. To verify a plausible answer, somebody needs enough knowledge to detect missing context, challenge a subtly wrong assumption and know when the evidence does not support the fluency of the conclusion.

Microsoft also found an association between higher confidence in AI and lower reported critical-thinking effort. The evidence was self-reported and correlational, but it raises a question: if people move from doing to checking, where will they acquire the judgement to check well?

Training can explain what an AI tool does, which prompts work, what policy allows and when a human must review the output. Learning happens when people use the capability, encounter reality, receive feedback, revise their assumptions and share what changed.

In McKinsey’s research on the learning organisation, nine in ten surveyed employees used generative AI at work, while only 13 percent considered their organisation an early adopter. Individual use was running ahead of organisational learning.

The awkward possibility is that organisations need more learning at precisely the moment AI makes it easiest to bypass the learning process. The answer does not have to be less AI; it may be better-designed AI.

Google’s Guided Learning uses questions, reasoning and testing rather than simply providing an answer. Google also reported that students using Learn Your Way scored 11 percentage points higher on a long-term recall test than those using a standard digital reader.

That is company-reported, education-specific evidence—not proof of workplace outcomes. But it shows that answering and learning require different designs.

The research on AI adoption points back to something familiar from earlier business transformations: access to the technology is only the beginning. A licence and a training session may help people start, but competence develops through use, feedback, discussion and repeated application to real work.

BCG’s research involving more than 10,600 employees found regular frontline use had stalled at around half, while more substantial training and visible leadership support were associated with stronger adoption. If rollout is measured mainly through licence activation, usage or hours saved, it can look healthy while understanding remains shallow. AI adoption is a people and change challenge, not a software deployment with some communications attached.

Conventional change management still matters: explain why, involve teams and create local ownership. AI makes the learning dimension harder to ignore, because people also need opportunities to practise, compare approaches, make mistakes, receive coaching and share what works. The Harvard Business Review work on learning organisations points to a supportive environment, practical learning processes and leadership that reinforces learning; without those conditions, organisations may collect polished outputs while missing the knowledge underneath them.

The quick answer is not the problem in itself. Used well, AI can explain concepts, expose alternatives and make expertise more accessible, but when speed substitutes for education and practice, an employee can become dependent without becoming capable, a team can lose shared understanding, and a business can exchange long-term judgement and resilience for short-term productivity. The better outcome is not to slow AI down, but to build learning into adoption as its use grows.

Choose one AI-enabled task that is spreading across your organisation and ask what people used to learn while doing it themselves, which of that understanding still matters, and how the new way of working will provide education, practice, feedback and shared learning.

If the answer is only “we trained people when we launched the tool”, the change plan is probably incomplete.

This is the first of three research-led articles exploring the learning gap behind AI automation.

Next: You cannot outsource experience to AI.

These are tools I use personally. Affiliate links marked - I earn a small commission if you sign up, at no extra cost to you.

Some links in this issue are affiliate links. I only recommend tools I actually use.