AI tutors in corporate training: what the research actually says

Why look at the evidence

Claims about AI tutors range from “a personal tutor for everyone” to “it just gives students the answers.” Both can be true, depending on how the tutor is built and used. A few recent, well-designed studies help separate the two. None of them is about corporate training specifically, so read them as signals about design, not as promises about your results.

Harvard: a scaffolded tutor beat active learning

Kestin, Miller and colleagues ran a randomised controlled trial in an introductory physics course at Harvard, published in Scientific Reports in 2025. Each of the 194 analysed students had one lesson with an AI tutor and one in a well-run active-learning class, in a crossover design. Median post-test scores were 4.5 with the AI tutor against 3.5 in class, from a pre-test baseline of 2.75: more than double the learning gain. Students in the AI condition also reported higher engagement and motivation, and most spent less time than the class period.

The design matters more than the headline. The tutor was not a general chatbot. Its system prompt followed research on active engagement, cognitive load and growth mindset; a platform walked students through each part of each problem in order, because the authors found a prompt alone didn’t keep multi-part problems on track; and prompts included step-by-step expert answers, so the model wasn’t solving physics on its own. The authors list the limits plainly: one course at one institution, lessons aimed at understanding and applying rather than higher-order synthesis, and some authors also designed the comparison class.

Khanmigo: access alone moved little

A two-year cluster-randomised trial in 18 Tennessee middle schools, published as NBER Working Paper 35620 by Oreopoulos and Low in 2026, tested Khan Academy’s AI tutor, Khanmigo, during existing remedial maths sessions. Assignment to the tool raised scores by roughly 0.06 to 0.08 standard deviations over a school year, similar to Khan Academy practice without the AI assistant. The bottleneck was use: nearly every student tried it, but students messaged it in only about 17% of the exercise sessions where they made a mistake, and most messages were bare answers or clicks on suggested prompts rather than real discussion.

Unguarded AI can hurt learning

A field experiment with nearly 1,000 high-school maths students in Turkey, by Bastani and colleagues and published in PNAS in 2025, adds a warning. Students with an unrestricted GPT-4 chat interface did much better during practice, but on the later exam without AI they scored about 17% lower than students who never had it. A version with guardrails, which gave hints instead of answers and used teacher input, improved practice even more and largely avoided that harm, though it didn’t raise exam scores either.

What this means for corporate training

Read together, the studies point the same way: the effect comes from design and use, not from switching a chatbot on. For a training team, that translates into a few practical rules.

• Ground the tutor in your course material and approved answers, as the Harvard tutor was, so it isn’t improvising policy or procedure
• Scaffold: guide learners step by step, and in practice give hints and questions rather than answers
• Assign it as an activity inside the lesson, such as a scenario or a set of questions, instead of leaving a chat box for people to discover
• Measure what learners can do without the tutor, not just how they perform with it
• Expect it to help most with understanding and applying; keep people for judgement, coaching and higher-order work

How to pilot an AI tutor

You don’t need a research team to learn whether a tutor helps your learners. Pick one course with a clear skill and an existing assessment. Run two comparable cohorts, one with the tutor assigned as practice inside lessons and one without, and keep everything else the same. Compare results on the same assessment taken without the tutor, then look at how the tutor was used: how many learners used it when they got something wrong, and whether their messages were real questions or requests for the answer. Read a sample of conversations for accuracy. A few weeks of this will tell you more than any vendor’s claim, including ours.

How Classbase helps

Classbase’s AI tutor is designed along these lines: it answers from your course and shows where the answer came from, it coaches rather than giving answers away during practice, and you can add it to a lesson as a practice card so it’s an assigned activity, not an optional extra. Whether it improves results in your setting is something to measure, and we’re happy to plan that with you in a demo.