Mentivis OS OpenMentivis OS EntrepriseMentivis OS Edu
Contact usDemo
All articles
M · AI & TrainingGeneral

Published 9 September 2026 · Updated 10 October 2026 · By Marie Castelli, Partner, Mentivis

How an AI agent supports a learner without replacing the trainer

An AI learning support agent is a conversational system, available at all times, that answers the learner's questions, rephrases explanations and gives immediate feedback on their answers. It does not replace the trainer: it provides a presence between human interventions. Its effectiveness depends on its educational design, the validated corpus it draws on and regular review by experts.

Pillar: Adaptive learning and personalised learning paths: the guide

What does an AI learning agent do well?

Answer questions when they arise. Without an agent, a comprehension question often waits until the next exchange with the trainer. This wait interrupts the learning effort.

Rephrase differently. The agent can explain the same concept in several ways, with other examples, until the learner understands.

Give immediate feedback. Feedback on an answer or a piece of work helps the learner correct their understanding without waiting for the next assessment.

Guide rather than answer for the learner. This is the main condition. In a randomised controlled trial with university students, Kestin et al. (Scientific Reports, 2025) found that an AI tutor designed on the same educational principles as an active learning class enabled students to learn more in less time, with higher reported engagement. Conversely, Bastani et al. (PNAS, 2025) showed that unrestricted access to a general-purpose assistant later lowered the results of students working alone, an effect largely mitigated by a version designed to guide without giving answers.

What does the agent not do?

It does not diagnose deep learning difficulties. It does not handle complex emotional situations. It does not assess relational or behavioural skills. It does not replace the human connection that sustains perseverance.

The agent can spot signs of difficulty from answers and progress (repeated errors, an abandoned exercise) and suggest another approach. It must not try to infer emotions: since 2 February 2025, the AI Act has prohibited AI systems intended to infer the emotions of a person in the workplace and in education institutions, except for medical or safety reasons.

What are the conditions for a reliable agent?

An agent that answers from a general-purpose language model, without configuration, produces fluent and sometimes wrong answers. Reliability rests on three elements:

  1. A validated document base: training content, competency frameworks, relevant regulatory texts.
  2. Usage rules: what the agent answers and what it must refer to a trainer.
  3. Regular quality review by subject-matter experts, who check the accuracy and relevance of answers.

Two obligations apply in addition. Since 2 August 2026, Article 50 of the AI Act has required providers to design systems that interact with people so that those people know they are interacting with an AI, unless this is obvious. And learners' exchanges are personal data, subject to the GDPR.

How to share roles between the agent and the trainer?

The agent handles The trainer handles
Repetitive comprehension questions Complex and situational learning
Requests for examples and rephrasing Emotionally charged situations
Procedure reminders and revision Assessment of skills requiring human judgement
Immediate feedback on exercises Relationship and long-term motivation

Frequently asked questions

Can an AI agent give wrong information?

Yes, like any system based on a language model. Configuration on a validated corpus and regular quality review reduce this risk without eliminating it.

Is an AI tutor more effective than a class?

It depends on its design. The trial by Kestin et al. (2025) measured better learning with a tutor designed on specific educational principles; the study by Bastani et al. (2025) measured the opposite effect with a general-purpose assistant without guardrails.

Must the learner know they are talking to an AI?

Yes. Article 50 of the AI Act, applicable since 2 August 2026, requires people to be informed that they are interacting with an AI system, unless this is clear from the context.

Does the agent learn from its interactions?

It depends on the system. If exchanges are used to improve the model, this purpose must be covered in the data processing agreement concluded under the GDPR and disclosed to learners.

Sources

Further reading

Related products: MentivisOS Entreprise, MentivisOS Open

Adaptive learning: definition, mechanisms and evidence
How AI personalises and adapts a training path