On this page
Key takeaways
- The how-to layer of L&D content, the everyday skills and enablement material teams build for people to look up when stuck, is moving to conversational agents. Formal learning, credentials and compliance, is not.
- Format is no longer the constraint. An agent can pull the right answer out of a slide deck, a PDF, a recorded call, or a five-year-old course without anyone rebuilding it first.
- Bad or duplicate source material is the real risk now. An agent will answer from the wrong version of a process with the same confidence it would give the right one.
- Completions measure attendance. The next measure is evidence pulled from real work, a sales call, an onboarding plan, a support thread, that shows whether something actually improved.
- L&D's job shifts from producing content to deciding which source is true and keeping it current.
Agentic learning is a simple idea with large consequences. Instead of sending someone to a course, you let them learn by talking to an AI agent, working through a problem in a real conversation. It’s already how most people reach for an answer at work. That’s why every education team is now facing the same question: if people already learn this way on their own, what happens to everything we built for them to sit through instead?
I have spent more than 25 years working with the world's leading enterprise learning programs. Based on that experience, here are my thoughts on how AI is transforming education: Formal learning is not going anywhere. But the everyday how-to layer, the skills and enablement content L&D teams have spent the last decade building, is about to get gobbled up by conversational agents. Most of it will stop mattering to how people actually learn at work.
At Intellum, we built an AI group and were running simulated AI interactions years before ChatGPT gave the rest of the market a reason to care. None of this is a sudden reaction to the latest tool.
The old course model made sense for the technology we had. For most of this field's history, learning software could deliver content and count who finished it. That was about it. It could not answer a question, so we built courses, loaded them into a system, and hoped people would find them when they got stuck. The library was the best way we had to organize and serve content, not the point of the work. Getting people the answer was the point, and the library was the closest the software could get us to it.
What Stays Formal Is What You Have to Prove
Some learning exists to prove something, and that part stays formal, because a machine can’t prove it for you. Think about credentials. A diploma, a degree, a certification. We chase them because we are wired to, and because they carry weight only if they mean the same thing for every person who holds one. An employer trusts a certification because it was earned against a fixed standard, on the record. An agent explaining the same material in a conversation, differently every time, can’t produce that. There is nothing to audit and nothing to compare.
Compliance survives for the same reason. The point was never that someone learned the rule. The point is that you can prove, on the record, that they met the requirement. That is a document, not a conversation.
So the dividing line is simple. If a person has to prove they did it, it stays formal. Everything else, the how-to layer built to answer questions, is where agentic learning takes over.
The Hard Part Is Accuracy, Not Volume
More content isn't the fix. The hard part was never volume, it's whether the source you're feeding an agent is the right one.
Format isn't the obstacle anymore. Give an agent a slide deck, a PDF, a recorded call, or a course from five years ago, and it can usually find the part that matters. You can throw it in the blender and it works. The blender is the agent itself. It does not care what shape the content comes in, a video, a doc, a transcript, an old module. For years we rebuilt the same content in one format after another so it would fit wherever people looked for it. The agent does not need that. It reads what you have and pulls out the piece someone actually asked for.
But a blender only works with what you put in it. It does not fix bad source material. If the same process lives in an old course, a help doc, and a buried slide, and two of them are wrong, the agent has a source problem. It may still give an answer, with the same confidence it would give a right one. That is the dangerous part.
So the work changes. Decide which version is true. Clear out the duplicates. Make the current source reachable. The content audit replaces the content calendar.
It is not glamorous work. It is still the work that decides whether any of this holds together. Tools can handle much of the technical lifting now. They can't decide which version your company stands behind.
That judgment stays with the people running the program. L&D does not disappear. It stops acting like a content factory and starts governing what the organization knows.
You Cannot Wait For Certainty
The harder problem isn't the technology. It's inside the company, sitting with legal and security review.
Most teams still use AI one way: prompt in, answer out, repeat. Useful, but limited. The shift that matters is agents that take several steps, use other tools, and work toward a goal without being told what to do at each one. That work is still uneven. Give an agent a long job and it may do something impressive, then stall halfway through. The direction is not in doubt, but the reliability is still being built.
None of that is what stops most companies. What stops them is their own security and legal review. The tool gets approved, the contract gets locked down, and by the end nobody can use it for anything real. A lot of that fear comes from one confusion, and it is worth clearing up, because it’s a problem. Using an AI model is not the same as training one. When your team uses a trained model, your data is not being poured back into it to teach it. But legal and security often treat those as the same risk, so they wall off the whole thing to prevent a danger that was never there.
Here is the part people get backward. The uneven reliability is not a reason to wait. It is the reason to start. These tools get better the more your people use them and learn where they hold up and where they do not. The company that waits until every answer is guaranteed does not arrive later with the same tool everyone else has. It arrives years behind, having learned none of it. Governing these tools well means using them now, carefully, not freezing until every risk is gone.
A company that will not let its people use these tools is choosing to fall behind the ones that will.
The Deeper Shift Is From Answers to Evidence
The deeper change starts when the agent can see the work itself. A sales call, an onboarding plan, a support thread: real output tells you far more than a quiz score ever could. An agent can read that work, find where someone is strong or stuck, and return the evidence while a person still handles the coaching.
We are doing this with early customers now. It is not an industry-wide switch someone flips tomorrow, and I would rather say that plainly than dress it up. But this is where the value moves.
A completion tells you the content was consumed. Evidence from the work tells you whether anything improved. One keeps a record. The other gives the business something it can use.
That's the shift underneath everything above. Formal learning stays formal because someone has to be able to prove it happened, on the record. The how-to layer moves to agents because proof was never the point there, getting the answer was. And once the source underneath it is one you can trust, the system stops asking whether someone showed up and starts showing whether the work got better.
