The LMS is no longer enough: what an AI training system adds

What an LMS does well

An LMS centralises content, manages enrolments, delivers modules to the right populations and produces activity reports: logins, time spent, completion rates, scores (see What is an LMS?). For regulatory training that everyone must take, or a catalogue to make available, it remains the right tool.

The criticisms levelled at it rarely concern what it does, but what is expected of it without it having been designed to do so.

What are the three limits of an LMS?

Personalisation remains manual. An LMS delivers the same module to an entire population, or variants defined in advance by conditional rules. Producing a distinct path for each profile, matched to what the person already knows, requires design work that does not scale to hundreds of employees. An employee who already knows half the subject may follow the same path as someone starting from scratch.

Completion is not acquisition. The data an LMS records by default, with the SCORM standard, is essentially limited to status, score and time spent. It answers the question "did the training take place?", not "was the skill acquired?". The xAPI standard (IEEE 9274.1.1-2023) can record richer activities, but the programme still has to assess the skill. This is developed in Measuring skills acquisition beyond the completion rate.

Content ages. A module produced once becomes outdated at the pace of its subject. Updating it depends on internal or external producers, with lead times poorly suited to fast-moving fields such as AI use or regulation.

What does research say about learning that lasts?

Several robust findings explain why delivering a module once has little lasting effect.

• Forgetting is rapid after a one-off learning episode. The forgetting curve described by Ebbinghaus was replicated by Murre and Dros (2015): loss is steep in the hours and days that follow, then slows. The experiment used nonsense syllables and a single subject; it cannot be converted into a percentage for professional skills, but it shows the value of returning to content.
• Spacing reviews improves retention. The synthesis by Cepeda et al. (2006), covering 317 experiments, shows that spaced sessions produce better retention than massed sessions, and that the optimal interval increases with how long the material needs to be retained.
• Testing yourself helps you learn. Roediger and Karpicke (2006) show that after one week, participants who tested themselves retained substantially more than those who reread the same text, even though rereading made them more confident.
• Transfer also depends on the workplace. The meta-analysis by Blume et al. (2010), covering 89 studies, links applying training on the job to motivation, ability and a supportive work environment.

An LMS can host quizzes and reminders, but it does not organise them by itself according to what each person has retained.

What an AI training system adds

An AI training system starts from the target skills and each person's actual level. It generates the path that closes the gap, without making the learner go over what they already know. It assesses acquisition along the way, through questions spread over time and practical scenarios, and recalculates what comes next at each step. Finally, it produces a record of acquisition by skill, usable for management, a funding body or an audit.

The available evidence mainly concerns intelligent tutoring systems, which rest on the same principle of adjustment. Kulik and Fletcher's 2016 meta-analysis of 50 evaluations found a median effect of 0.66 standard deviations on test scores, weaker on standardised tests than on locally developed tests. These results support the approach, not the effectiveness of any given product: that has to be checked on your own data.

What AI does not replace

AI does not replace human support in complex situations, peer-to-peer learning and the role of line managers, or the working conditions that allow people to apply what they have learned. Research on transfer shows that the work environment weighs on how training is applied. A tool optimises the learning programme; it does not, on its own, create these conditions. See What AI cannot do in training.

Replace or complement the LMS?

The question is often asked too early. An AI training system can integrate with an existing LMS, via API or through the SCORM, xAPI and cmi5 standards, with the LMS continuing to provide access and traceability. It can also operate on its own, depending on the infrastructure. The right criterion is what the organisation needs to produce: if the need is measured acquisition and individualised paths at scale, delivery alone does not cover it. The families of tools are distinguished in LMS, LXP, adaptive learning, AI training engine: how to tell them apart, and market offerings in the overview of alternatives to the LMS.