My, how time flies! Vendors have been selling AI in corporate learning for three years. The Fosway AI Market Assessment for Digital Learning 2026, published in April, is the closest thing to an existing record of what vendors have actually built with it. It surveyed digital learning vendors (custom content developers, learning platform providers, managed learning services firms), counted every AI capability they offer, and separated what’s running with customers from what’s still on roadmaps.
Vendors, it seems, have put most of that AI-focused-effort into one thing: making content cheaper to produce.
Content takes 64% of AI roadmap investment across the market. A year ago, four AI capabilities counted as standard, meaning more than half of vendors had them running with customers. There are now fourteen, and nearly all of them generate material or convert it: captions, transcripts, synthetic video, presenter avatars, quizzes, and full e-learning courses drafted from a folder of source documents.
For a buyer, that’s good news about price. Production is deflating, and Fosway’s advice is not to pay a premium for capabilities most vendors already offer.
The numbers do overstate how much of this is genuinely in use. Fosway counts a capability as live once a vendor has switched it on, which is not the same as customers working with it at scale. It’s worth asking a provider to name customers using the thing they’re demoing!
The more important point for buyers is that cheaper production only changes how fast content gets made. It doesn’t make the content correct, or current, or the thing that was worth building in the first place. Aim fast generation at documents nobody has checked, organized or brought up to date, and you get the wrong content faster, in greater volume, and with a more polished finish on it in all likelihood.
Which changes what a buyer should be looking at.
The source material decides the quality
When any provider can draft a course from a set of documents, what separates one output from another is the source material: your processes, your products, your policies, and the record of how the work actually gets done in your business.
Preparing that material is where the effort now sits, and it’s the part the industry has invested in least. AI tagging tools exist and are spreading – roughly three-quarters of vendors have them on a roadmap, about a third have them working – but they produce metadata a person still has to check, and the money has gone overwhelmingly into generation rather than preparation.
We see what that costs on the ground in our work at Sify Digital Learning. On one engagement we processed close to six million assets. Text was the straightforward part. Images and video needed vision models running as the files arrived, generating metadata describing what each one showed, with a person checking that metadata before anything entered the repository. Once material is described and validated that way, a generation tool can find the right piece and produce something accurate. Without it, the same tool is really just guessing.
The source material decides the quality
If a provider produced twice as much for you this year, that reflects their tooling rather than their value to your business. Volume has stopped saying much about anyone.
The hard work is in deciding what to build. That means, for example, judging whether a request from a business unit is a training problem or a process problem, working out which material already in circulation is out of date and should be pulled, and setting what gets checked before content reaches people, against which standard, and who signs it off. In regulated work that last one is also your audit trail, because an auditor asks where content came from and who validated it.
Falling production costs don’t touch any of that. It’s the work we think providers should be adding!
What to do about it
Your source material sets the ceiling on anything AI produces for you, so its condition is worth establishing before you buy more capacity to point at it. And when you next review a content partner, ask about what happens around generation: how they decide what should be built, what gets checked before it reaches people, and how material stays current once it’s out there.
Then look at the next content line in your budget. If what you’re paying for is capacity – hours, modules, output — you’re buying the part the market has spent two years making cheap, and not the part that decides whether any of it works.
Turning what your company knows into material that holds up when people use it is what we do at Sify. If you’re weighing what that looks like across your teams, let’s talk
