The AI training market filled in quickly. A leadership team that wants to upskill an engineering function or a knowledge-work team has dozens of options inside an hour of searching, most of them self-paced, most of them inexpensive on a per-seat basis. The cost case is straightforward enough that the training decision often gets made by procurement before anyone asks whether the training will change how the team works.
Sixty days after the certificates are issued, the honest answer is usually no. The team is producing the same work in the same way. Worth understanding why that happens, because the failure mode is consistent across organisations.
Generic curricula teach tools the team cannot deploy
Most off-the-shelf curricula are built around the tools the curriculum author can demo cleanly, not the tools the participants can actually use at work. That gap is wider than it sounds. A lesson on prompting a public model is not directly transferable when the company's data cannot leave the network. A lesson on a vendor's flagship feature is not transferable when the company uses a different vendor or no vendor at all.
Participants finish the course knowing things they cannot apply on Monday. The knowledge fades, because nothing reinforces it.
Nobody built anything on their own work
Hands-on work in a generic course is hands-on with the course's data, not the participant's data. The exercises are clean, the prompts converge, the outputs look good. None of that is how the participant's real work behaves. Real work is messy in specific ways, and the technique for working with it has to be developed against the mess, not against a textbook example.
If the participants do not build anything on their own work during the training, the post-training transfer is left as an exercise for them to do alone, on top of their day job. Almost nobody does.
There is no follow-up to reinforce
Skills that are not used start fading inside two weeks. A team that comes out of an intensive training and then returns to the same workflows, the same review processes, and the same managers will revert to the pre-training pattern almost immediately. A single follow-up session six to eight weeks after the training, focused on what the team has actually tried to apply, is one of the highest-leverage interventions available, and it is missing from almost every training package on the market.
What does work, in our experience
Training that changes behaviour shares a few traits. The curriculum is built around the team's actual stack and constraints, not a generic one. The hands-on work happens on the team's real data and real workflows, during the session, not as homework afterwards. The team leaves with something they built, something they own, something running somewhere they can return to. There is a scheduled follow-up four to eight weeks later, focused on what got tried, what worked, and what got stuck.
Done this way, training looks more like a short engagement than a course. It costs more per seat than a self-paced subscription, and that gap is real. The honest comparison is not between the per-seat costs, though. It is between the cost of training that changes behaviour and the cost of training that does not.
When the commodity option is the right one
None of this is an argument against self-paced courses across the board. For an individual contributor who is motivated, has time, and wants to develop a personal capability outside of any specific organisational context, a good self-paced course is excellent value.
The argument is narrower: if the goal is to change how a team operates, the commodity option is almost always the wrong purchase. The team-capability problem is not a content-delivery problem. It is a behavioural one. Behavioural problems do not yield to content, no matter how good the content is.