Every conference session about AI in learning and development this year has covered the same ground, and it isn’t the ground practitioners need.
The talks describe capability. Teams need a map of where that capability holds up.
The gap between “AI can write a course” and “AI improved our training” is where budgets get spent and quietly written off. Parts of this work now and are badly under-used. Parts are oversold in a way that will be obvious by 2028. Two jobs shouldn’t be delegated at all.
Here’s that map, task by task, with the reasoning rather than a verdict.
Try Mini Course Generator free – test the parts that work on a document you already have.
Key Takeaways
– The strongest, most under-used capability is generating training from material you already own. It’s unglamorous, it’s reliable, and it’s where the largest backlog sits in most companies.
– The most oversold is adaptive personalisation. It demos beautifully, needs data volumes most corporate programmes never reach, and rarely beats a well-sequenced linear course.
– Post-publish maintenance is the biggest unclaimed opportunity: everyone automated authoring and nobody automated the three years afterwards.
– Two jobs shouldn’t be delegated: deciding what to teach, and judging whether training was the right intervention at all.
– The organisations getting real value aren’t the ones with the best tools. They’re the ones that had good documentation before they started.
The Maturity Map
Sorted by how confidently you can rely on it today.
| Task | Status | The honest read |
|---|---|---|
| Generating courses from your own documents | Works well | Reliable, under-used, biggest backlog |
| Translation and localisation | Works well | Mature, cheap, mostly solved |
| Building interactive activities and simulations | Works with a rubric | Excellent output; useless if nothing is scored |
| Post-publish maintenance and content health | Works, rarely deployed | The largest unclaimed opportunity |
| Needs analysis from support and ticket data | Works, rarely deployed | Evidence is sitting there uncollected |
| Assessment writing | Partial | Defaults to recall unless explicitly instructed |
| Generating courses from a topic prompt | Partial | Fluent, generic, and learners notice |
| Adaptive personalisation | Oversold | Needs data volumes corporate L&D doesn’t have |
| Predicting skills gaps | Oversold | Confident output, weak evidence base |
| Measuring business impact | Doesn’t work | The attribution problem is not a data problem |
| Deciding what to teach | Don’t delegate | The judgement that determines everything downstream |
| Deciding whether to train at all | Don’t delegate | Often the answer is a process fix |
The rest of this article is the reasoning behind each row that matters.
What Works, and Is Under-Used
Turning documentation into training
The least exciting item on the list and the one with the highest expected value.
Most companies hold a large, growing pile of documented knowledge and a small, ageing pile of training. The gap isn’t a knowledge problem.
It exists because converting documentation into a sequenced, assessed course is slow, unrewarding work. It loses every scheduling argument it enters.
That conversion is now cheap. An assistant connected to your documentation and a course platform does it in one request.
The output is a real course, not a draft to assemble. The recipes differ by source – Confluence, Notion, Drive, a support channel – and the shape is identical.
Why this beats the flashier options: the input already exists and is specific to your company. A course built from your runbook is about how your business works. A course built from a topic prompt is about the internet’s average view of that topic. Learners spot the difference by lesson two.
Interactive activities, when something is scored
Role-play simulations, branching scenarios and decision drills are genuinely good now, and they used to be the most expensive thing an instructional designer built.
The condition is the rubric. Without explicit scoring criteria, an AI role-play is conversation practice. It’s pleasant, and it rehearses whatever the learner already does.
Add 4 to 6 observable criteria and it becomes assessment. The mechanics are in building a rubric-graded roleplay.
Maintenance, which almost nobody has deployed
The largest unclaimed opportunity in this list.
Every AI feature in this market points at authoring. A published course then lives for 3 years, and its decay is invisible.
The quiz question 97% of learners pass isn’t testing anything. The lesson where completion drops 30 points is losing people. The policy that changed in March 2026 is still described the old way in 4 courses.
All of it is detectable from data the platform already holds. Almost nobody looks, because looking was manual and carried no deadline.
An agent with read access finds all four in a 15-minute monthly conversation. We made that argument in full in what an agent does after publish.
If you do one new thing this year, this is the one with the best ratio of effort to result.
What’s Oversold
Adaptive personalisation
The demo is compelling: the system detects a learner struggling and adjusts the path. In corporate L&D it under-delivers, for two structural reasons.
The data volume isn’t there. Adaptive systems need thousands of assessment events to separate signal from noise. A compliance course taken once a year by 340 employees produces 340. The personalisation is real; the confidence behind it isn’t.
Corporate content is short. Adaptivity pays off across a long curriculum with branching paths. Most workplace training is a 20-minute module with one correct sequence. There’s nothing to adapt.
Where it does work: large-scale customer education, sales certification with 500+ participants, curricula with genuinely optional depth. Real cases, and a small share of L&D spend.
For most teams a well-sequenced linear course with an honest assessment beats an adaptive one built on thin data. It also costs a fraction of the attention.
Predicting skills gaps
Skills-gap prediction produces confident, well-formatted output from evidence that can’t support it.
The inputs are job titles, self-assessments and completion records. Self-assessed skill data is unreliable. Completion records measure attendance, not capability.
Feed weak evidence into a fluent model and you get a persuasive artefact. That artefact then enters planning conversations, where its provenance is invisible.
The version that works is narrower: look at where people currently fail. Support tickets, error rates, escalations, the questions your team answers 40 times a quarter.
That’s observed behaviour rather than predicted capability. It’s sitting in systems nobody is mining.
What Doesn’t Work
Measuring business impact
Worth being blunt, because a lot of money is aimed at this.
The attribution problem in L&D isn’t a data problem, so more data won’t fix it.
Sales rose 12% in Q3. The team also took the negotiation course, got a new comp plan, and shipped a better product. No model separates those. The information needed to separate them was never captured.
AI makes the reporting faster and prettier. It can’t establish causation from observational data with a sample of one team and no control group.
Treat any tool claiming to prove training ROI as claiming more than its evidence allows.
What’s honest instead: measure what you can actually observe. Did the questions in the support channel about this topic decrease after the course shipped? Did the assessment scores move? Did the specific error rate drop? Those are narrow, real, and defensible, and they’re worth more in a budget conversation than a modelled ROI figure that anybody can challenge.
The Two Things Not to Delegate
Deciding what to teach
An agent builds what you ask for very fast. A vague brief produces a polished course about nothing, and now you own a professionally formatted artefact that looks like work.
Knowing that new hires stall in week two on one specific escalation path is a human judgement built from talking to people. It’s the input that determines whether everything downstream was worth doing, and it’s the step that gets skipped fastest when generation is cheap.
The failure mode to watch for is volume mistaken for progress. A team that ships 40 courses a quarter and can’t say which three mattered has automated the wrong thing.
Deciding whether to train at all
The one most likely to be quietly dropped, and the most valuable thing an L&D team does.
A large share of requests arriving as “we need training on X” aren’t training problems.
They’re a confusing interface, a process with no owner, a form that asks for the wrong thing first, an incentive pointing the wrong way. Training papers over those and has to be repeated forever. It’s why a compliance module reaches its 9th annual edition.
When building a course costs a week, that friction forces the conversation – somebody asks whether it’s worth it. When it costs ten minutes, nobody asks. Cheap production removes the natural check on whether production was the right response.
Keep the check deliberately. It’s now the most valuable thing in the workflow, precisely because it’s the only part that got harder rather than easier.
What Separates the Teams Getting Value
A pattern worth naming, because it’s not about tooling.
The organisations getting real results from AI in learning and development had one thing first: good documentation. Current runbooks, processes written down, support content maintained.
That isn’t a coincidence, and it isn’t fair. AI in L&D is a conversion technology. It turns knowledge you already captured into training.
If nobody captured it, there’s nothing to convert. Generating from general knowledge produces exactly the generic output nobody wanted.
That suggests an uncomfortable order of operations. If your documentation is thin, this year’s highest-value AI project isn’t a training project. It’s writing 6 processes down properly, then converting them in an afternoon.
The second pattern: teams treating courses as things they run, not things they ship. That distinction is what the term agentic LMS names. It’s also where a fast content factory and a functioning training programme stop looking alike.
The Short Version
AI in learning and development works best on the least glamorous task: turning documentation you already own into sequenced, assessed courses. That’s where the backlog is and where the output is specific enough to be worth taking.
Interactive activities work when something is scored against explicit criteria. Without that, they’re conversation practice.
Post-publish maintenance is the biggest unclaimed opportunity. Everybody automated authoring. Nobody automated the 3 years afterwards.
Adaptive personalisation and skills-gap prediction are oversold at typical corporate data volumes. Proving business impact isn’t a data problem and won’t be solved by better models – measure what you can observe instead.
Two jobs stay human: deciding what to teach, and deciding whether training is the right intervention at all. The second matters more than it did in 2024, because cheap production removed the friction that used to force the question.
And the teams getting the most out of this had good documentation first. If yours is thin, that’s the project.
Frequently Asked Questions
What does AI actually do well in learning and development?
Generating courses from documentation you already own, translation and localisation, building interactive activities where a rubric defines what good looks like, and detecting content decay in published courses. The first and last are the most under-used relative to their value.
Is AI-powered adaptive learning worth it for corporate training?
For most teams, no. Adaptive systems need large learner volumes and long curricula to work; typical corporate training is a short module taken once a year by a few hundred people. A well-sequenced linear course with honest assessment usually outperforms an adaptive one built on thin data.
Can AI prove the ROI of training?
No, and treat claims otherwise carefully. The attribution problem isn’t a data problem – too many things change at once, and no model separates them from observational data. Measure observable proxies instead: support questions on the topic, assessment scores, specific error rates.
What should L&D teams not delegate to AI?
Deciding what to teach, and deciding whether training is the right response at all. A significant share of training requests are really process or interface problems, and cheap course production removes the friction that used to force somebody to ask.
Where should a team start with AI in L&D?
With the documentation-to-course conversion, on one real process where the documentation is current and the training doesn’t exist. It’s the highest-confidence use, the output is specific to your company, and it tells you within an afternoon whether the rest is worth pursuing.
Why do some organisations get more out of AI in L&D than others?
Almost always because their documentation was already good. AI in this space is largely a conversion technology, so the quality of what you already captured sets the ceiling. Teams with thin documentation get generic output and conclude the technology doesn’t work.
