I recently spent an evening arguing with a laptop. I was demonstrating a conversational AI tool to my new boyfriend who also works in learning and development: me, out loud, having a difficult conversation with something that flatly refused to make it easy for me.
He laughed for the first two minutes, watching me get politely taken apart. Then he went quiet and concentrated. At the end he asked why our industry is not talking about this constantly. That question has stayed with me, because it points at the most expensive gap in our profession.
Here is the awkward secret everyone in L&D knows and almost nobody budgets for. Practice is how people get better at difficult conversations. Not models. Not frameworks. Reps. Ask a room of managers to describe a good feedback conversation and most will do it fluently. They know the acronym. They can draw it on a flipchart. Ask when they last rehearsed one before having it for real, and the room studies its shoes.
The research on this is old enough to have grey hair. Baldwin and Ford set out the transfer problem in 1988: most of what gets learned in training fails to travel back to the job, and the conditions that decide whether it travels sit outside the classroom, in the opportunity to apply, repeat and refine. Decades of studies since have kept arriving at the same conclusion. Knowledge is the cheap part. Application is where development lives or dies. Ericsson’s work on deliberate practice makes the same point from the other direction: skill is built by practising at the edge of your ability, with honest feedback, in conditions where mistakes cost nothing.
Now measure the traditional training room against those conditions. Repetition? One go, if you volunteer. Honest feedback? The facilitator is watching twelve people at once. Realistic difficulty? Your colleague playing the difficult employee is far too polite to be genuinely difficult, so the resistance is never real enough to build any confidence against. Cheap failure? Fluffing it in front of your peers is the most expensive failure most managers can imagine.
There is solid psychology underneath that last point. We have known since the earliest social facilitation studies that being watched helps people perform simple, well-learned tasks and actively degrades performance on complex, unpracticed ones. A difficult conversation is about as complex and unpracticed as workplace tasks get. So the very format we chose for practising hard conversations, a circle of observing peers, is the format most likely to make people worse at them in the moment, and to make sure they never want to try again.
We have known role-play does not work socially for as long as we have known practice works technically. The facilitator asks for a volunteer and the energy leaves the room. So we colluded, all of us. One carefully rationed round of role-play, everyone agrees it was useful, nobody asks to repeat it, back to the slides. I have run those sessions. I have also been the delegate staring at my shoes, and if you are honest, so have you.
The cost of that collusion is brutal when you say it plainly. A manager’s first attempt at a genuinely hard conversation, the underperformance one, the restructure one, the one where somebody cries, is the real one. It happens with a real person whose confidence and career are on the line, at precisely the moment the relationship is most fragile. Aviation does not work this way; pilots meet their emergencies in simulators first. Medicine does not work this way; surgeons practise on models before they practise on you. Management works this way as standard. We hand people a laminated framework and our best wishes, and then we run engagement surveys to find out why trust in line managers is low.
This is why I think the most interesting thing AI offers HR has nothing to do with generating course content or summarising policy documents, which is where most of the current noise sits.
Conversational AI has quietly changed the economics of practice. A manager can now rehearse a difficult conversation out loud, against something that answers back and pushes back like a person, in private, as many times as they need, with structured feedback at the end of every attempt. Repetition, feedback, realistic difficulty and cheap failure, all four conditions of deliberate practice, available at a desk on a Tuesday afternoon.
And the private part matters more than the clever part. Remove the audience and you remove the cringe, and the cringe was the barrier all along. Role-play, without the cringe, turns out to be something people actually want to do. In our own workshops, delegates have started giving up their lunch break to retry a scenario they fluffed an hour earlier. Voluntarily. Anyone who has ever begged a room for a role-play volunteer will understand why we find that mildly unnerving to watch. The motivation to practise was never absent. The format was intolerable, and for forty years we mistook a format problem for a motivation problem.
Before anyone reaches for a chequebook, some honesty, because this market is filling fast and unevenly, and our industry has bought its share of shiny things that ended up as shelfware. These tools are only as good as the scenario design and the feedback model underneath them. An AI that flatters everyone builds false confidence, which is worse than no confidence at all. The behavioural criteria need to come from people who understand behaviour, not from whatever a general-purpose model defaults to. And practice belongs inside a designed programme, alongside teaching, coaching and real accountability, not as a substitute for any of them. A simulator did not replace flight school. It made flight school work.
If you are an HR director being pitched AI tools every week, and you are, three questions will cut through most of the noise. First: who designed the feedback, and against what behavioural criteria If the answer is vague, walk away. Second: does the simulation offer realistic resistance, or is it engineered to be agreeable? Ask for a live go, and try being difficult with it. Third, and sharpest: is there any evidence people use it when nobody is making them? Voluntary repetition is the difference between building a practice culture and renewing another licence nobody opens.
The workshop was never the problem. The three weeks after the workshop is where development budgets quietly die, and it has been dying there since before most of us joined the profession. Close that gap and the money you already spend starts working properly, because everything you teach finally gets practised. Leave it open, and the next framework will transfer exactly as well as the last one did.
