Two years ago a scenario video meant a script, a casting call, a location, a shoot day and an edit. Now we build one in an afternoon. The honest version of that sentence is less exciting than it sounds, because the afternoon is not where the project succeeds or fails.
What AI compressed was production — the part between a finished decision and a finished asset. What it left untouched is the part before that: working out what the learner actually needs to be able to do, and what has to be true in their head for them to do it. That part is still slow, still human, and still where almost every bad course goes wrong.
The bottleneck moved, and budgets have not caught up
When production was expensive, it made sense to spend most of a budget on it. Teams would settle the design quickly so there was money left to build.
Now the arithmetic has flipped and most budgets have not. We regularly see briefs that allocate weeks to media production and two days to deciding what the course is for. That was a reasonable ratio in 2019. Today it means paying a lot of attention to the cheap half of the job.
If a concept needs a diagram with fourteen boxes, it has not been understood yet. No rendering engine fixes that.
The teams getting real value from AI are the ones who took the time production gave back and spent it upstream — on needs analysis, on arguing with the subject expert, on cutting the course in half. The teams getting little value produced the same unclear course, faster.
Faster production makes unclear thinking more expensive, not less
This is the part that surprises people. When a video took three weeks, nobody shipped one they were unsure about. The cost imposed a kind of discipline.
When a video takes an afternoon, that discipline has to come from somewhere else, because the cost of being wrong has gone up, not down. You will now produce forty of the unclear thing instead of one. We have watched organisations build a whole library at speed and then discover, at rollout, that the library teaches the wrong model of the job.
The fix is not to slow production down. It is to put a real decision gate before it — someone whose job is to say we have not understood this yet and be listened to.
What this means for a brief
Three things we would push for on any AI-assisted programme:
- Spend the saving upstream. If production is now a third of what it was,
move that money to design rather than banking it. The course gets better and the total still falls.
- Approve meaning before medium. Sign off the explanation in plain text,
with no production value attached to seduce you. If it does not convince in a paragraph, it will not convince in 4K.
- Pilot before you scale. Build three, put them in front of real learners,
then build the other forty. Speed makes this cheap, which is the best argument for it.
Where AI genuinely earns its place
None of this is an argument against the technology. We built four AI products ourselves and use them on every project. The gains are real, and they are biggest in exactly the places people find least glamorous:
Re-versioning. A second language, an updated policy, a different audience — work that used to mean rebuilding now means regenerating. Consistency at volume. A hundred lectures that look like they came from one programme rather than eleven freelancers. Iteration. Changing your mind after seeing version one, which used to be unaffordable and is now routine.
Those are not small. They are just not the thing most people are sold. The pitch is usually "make courses faster"; the value is "stop paying a penalty for getting things right on the third attempt".
Technology makes learning faster to produce and cheaper to scale. It does not make it clearer. That is still a design problem — and it is the one worth hiring for.