AI Agents in L&D Are Already the Platform
While L&D teams debate their AI strategy, AI agents in L&D have already been adopted — just not by L&D. Your people are being trained daily by a system nobody in your function specified, reviewed or measured.
Someone stuck on a process does not open the LMS. They ask the assistant embedded in the tool they are already in. They get an answer in seconds, it is usually good enough, and they act on it. That interaction taught them how to do the job. It is training by any honest definition, and it happened entirely outside your remit.
However, multiply that across a few thousand people and several hundred times a week, and the question is no longer whether AI agents belong in your learning strategy. They are your learning strategy. You just have no visibility of it.
What do AI agents in L&D actually change?
They move the moment of learning to the moment of need, and that breaks the assumption most curricula are built on.
Formal learning assumes you can predict what someone needs and deliver it in advance. That has always been imperfect, but it was the only option — the alternative was asking a colleague, which was slow and inconsistent.
An always-available assistant removes that constraint. Nobody needs to anticipate the need, because the answer arrives when the need does. Therefore the case for teaching things in advance now has to be made, not assumed. Some of it survives that test easily. A lot of it does not.
Which parts of your curriculum does this make redundant?
The parts that exist to make information retrievable, which is a larger share than most teams expect.
Procedural how-to content, system walkthroughs, reference material, policy detail — all of it was built because looking things up was hard. It is not hard any more. Continuing to build it is spending design effort competing with a tool that is faster, always available, and updated more often than your library.
In fact, what survives is everything that is not retrieval. Judgement under ambiguity. Skills that need practice with feedback. Anything requiring someone to have internalised a standard rather than looked one up. Conversations that only work between people.
The uncomfortable version: if an agent can answer it, you probably should not have built a course about it. The comfortable version is that this frees your team to work on the things only they can do.Why is this a governance problem before it is a design problem?
Because the agents are already answering, and nobody has checked what they say.
Ask your organisation's assistant a question about a policy with real consequences — expenses, data handling, escalation. You will get a confident answer. Nobody in L&D wrote it, nobody reviewed it, and it may be drawn from a document three versions out of date.
Specifically, that is a quality and risk exposure that would be unacceptable in any formal channel. If a module gave that answer it would be pulled. Because it comes from a tool rather than a course, it sits outside the governance that would otherwise apply.
This is the most useful place for L&D to intervene right now, and it does not require an AI strategy. It requires asking what the agents are being asked and whether their answers meet your standard.
What should an L&D team actually do about it?
Three things, and none involve building an agent.
Find out what people are asking. Most platforms expose query logs. That data is the clearest picture of real capability gaps your organisation has ever produced — far better than a training-needs analysis, because it is unprompted and behavioural rather than self-reported.
Fix the sources. Agents answer from documents. If the source is wrong, ambiguous or stale, the answer inherits that. Curating what the agent draws on is L&D work, and it has more effect on capability than another module.
Ultimately, design for what remains. AI agents in L&D handle retrieval; your team handles the rest. Stop building retrieval content and move that effort into practice, judgement and assessment — the things an agent cannot do for someone.
Does this shrink the L&D function?
It shrinks one part of it and enlarges another, which is not the same thing.
The production of informational content will shrink, and that is where a lot of L&D headcount currently sits. Pretending otherwise is not doing anyone a service.
But curating what the organisation's agents teach, defining what good looks like, and building the judgement that no agent can supply is more valuable work than producing modules, and considerably harder to outsource. The teams that come out of this stronger will be the ones that moved early, not the ones that defended the module count.
What does good look like in a year?
Not an L&D team that has built its own agent. That is the expensive answer and rarely the right one.
Good looks like an organisation where the answers people get are correct, because someone owns the sources they come from. Where the query data feeds capability planning instead of sitting unread in a platform admin panel. And where formal learning has contracted to the things that genuinely need it, and got better because the effort concentrated.
That is a smaller content operation and a more influential function. It is available to any team willing to give up production volume as a measure of its worth.
