Why are we still learning this way?

We are getting used to AI agents working across our tools and carrying out increasingly complex tasks. Yet some mandatory e-learning would have looked much the same five or ten years ago.

I felt that gap during a recent governance training session. The money laundering module explained the regulation and presented individual situations, but gave me less help understanding how they connected. I found myself wanting a clearer sense of what the rules were trying to prevent. I was looking for a mental model or a decision tree. And if I am being very honest, I just wanted to make it go away as fast as possible too.

Those wishes are quite compatible. A useful explanation might have helped me make sense of the material sooner. Instead, I was working through the course in the order it had chosen for me, while trying to assemble the connections myself. I suspect many people recognise that experience, although my frustration with one module does not establish that all mandatory training is poorly designed.

There is a reasonable organisational purpose behind the familiar format. A large company can give everyone the same approved material and document who completed it. It knows what was presented and can update the course when the requirements change. An employee's conversation with an AI sidekick is less predictable. One person asks about the principle behind a rule, another wants to understand an exception, and a third is trying to obtain the answer that will let them finish. I would not exclude myself from that last group. Giving everyone a chatbot would not, on its own, improve the learning.

The question is whether we can preserve the required ground while allowing a more useful route through it. I would start by deciding what someone should be able to explain or do at the end. That gives the conversation a purpose against which it can be assessed. The approved material would remain the source for the rules, including their limits and the circumstances in which a person needs help. The sidekick could respond to questions and work through examples, but it would need to recognise when the material did not support an answer. A plausible explanation invented to fill a gap would be a particularly unfortunate contribution to governance training.

Consider a hypothetical course on a company's payment controls. Its fictional policy requires additional checks when a supplier requests a change of bank details. An employee can identify the expected answer in a multiple-choice question, yet still be unsure what problem the check addresses. A guide could ask them to explain what they would want to establish before proceeding. If they reply that the email looks genuine because it came from the usual address, the next part of the discussion could examine whether familiarity alone establishes that the change is authorised. The learner's answer would reveal where an explanation was needed. Someone who already understood the issue could move on to a less straightforward case.

I would then change the situation and ask the employee to reason through it without help from the tutor. Perhaps the request now arrives through a familiar colleague, or part of the information needed for the check is missing. The assessment would examine whether they recognise what remains uncertain and can explain the next step under the fictional policy. Reciting the earlier answer would offer less evidence of understanding than applying the principle when the details change.

That would let different conversations lead towards common expectations. It would also make the assessment less dependent on how helpful the tutor thought it had been. A system that leads someone through every step may produce an excellent answer while leaving the learner unable to reproduce the reasoning. I would want the unaided attempt assessed against criteria agreed by the people responsible for the subject, with examples of acceptable reasoning and mistakes that require further work. Some ambiguous answers would still need a human review.

There is evidence worth exploring. A randomised study published in 2025 found stronger immediate learning outcomes with a carefully designed AI tutor than with an active-learning classroom in an undergraduate physics setting. That supports investigating the approach; it does not establish that an unrestricted chatbot improves compliance training or that knowledge will survive until it is needed at work.

For that reason, I would try this on a bounded part of a course before replacing the familiar format. Alongside checking the quality of the conversations, I would compare how learners handled a changed scenario afterwards and again after some time had passed. A quicker completion time or a more pleasant experience would be welcome, but neither would tell us whether someone could use what they had learned. The organisation would also need to decide what assessment evidence to retain. It should have a reason for keeping personal questions, rather than collecting every conversation simply because it can.

I would have welcomed a way to ask why the situations in my module belonged together, work through the answer, and then find out whether I could apply it. The course would still have asked something of me. It might just have made that effort feel more connected to the decisions the training was meant to prepare me for.