The wrong promise
Most stories about AI and learning are about speed. A module in a day, a course in a week, a manual that writes itself. That is not untrue, but it is the least interesting half of the story, and it is the half most L&D managers distrust, rightly.
Anyone who has ever opened a generated e-learning module recognises it: correct, complete, and utterly soulless. Text that holds up without being about anything. Examples that come from no reality at all. An assessment that checks whether you read the previous paragraph.
That material was not made faster. It was delivered faster, which is a different thing. The hours you save on writing, you lose explaining why this is not what you meant.
What does change
The real change does not sit in the first version. It sits in what happens afterwards.
Training material for industrial environments has a problem that has little to do with writing: it ages. An installation gets extended, a procedure changes, different equipment arrives, regulation moves. With traditional production, every change means a new programme: quote, schedule, lead time, budget. The consequence is predictable. Small changes do not get made. Material falls behind. At some point nobody trusts it any more and everybody learns from a colleague again.
What AI changes in the production cycle is the cost of an amendment. Not the cost of the first design, which stays largely human, but the cost of version twelve. And with that, it changes which projects are feasible.
A few examples of what did not add up before:
- Five languages instead of two. Not because translating became free, but because the consistency check across five versions is no longer the bottleneck.
- Material per site instead of one generic version. The same didactic structure, but with the installation, the terminology and the case material of that specific site.
- Updating twice a year instead of a revision every three years.
- A clickable preview before the build decision, so stakeholders respond to something tangible rather than to a blueprint.
None of those is about writing faster. They are all about maintainability.
Where it goes wrong
Three patterns come back repeatedly at organisations that have tried it themselves.
The source material is the problem, not the model. AI given your outdated SOP as input produces an outdated module, faster, and with more conviction. When the procedures are wrong, automation amplifies that problem rather than solving it.
Content errors are invisible to whoever cannot see them. Generated text reads fluently, even when something is in it that an experienced operator recognises immediately as nonsense. A reviewer without subject knowledge notices nothing. This is the main reason the expert stays in the process: not to write, but to check.
Generic material confirms the sceptic. The most dangerous outcome is not a module that is wrong, but a module that is about nothing. Anybody who already doubted digital training now has their evidence. That costs you not one project but the next one too.
What people keep doing
In our own production the line sits roughly here.
What we do not automate: deciding what somebody has to be able to do afterwards, judging which mistakes are most expensive in practice, choosing which scenario is credible for this audience, and the content judgement of whether something is right. Those are didactic and subject matter choices, and they decide whether the material works.
What we do automate: producing variants on a fixed structure, keeping language versions consistent, enforcing terminology across a whole set, reshaping material for a different format, and delivering the first version an expert then corrects.
That second list is not the creative part. It is the part that used to eat the budget, leaving no room for the first.
What this means for your planning
If you have a training question now, this mostly changes your starting assumptions.
Stop counting on months of production as a reason to keep the scope small. The limiting factor is the availability of your own specialists for input and review, not build capacity. Plan around that.
Treat your source material as the real project. If your SOPs are not current, start there; everything you build on top inherits the problem.
And with every quote, ask not only what it costs to make, but what it costs to update in two years. That answer says more about the approach than the price of the first delivery.
That question belongs in a broader trade-off about how you set up your training offering. It is in the guide on setting up or renewing a corporate academy. What we do with it ourselves in production is under AI-assisted development.