Why AI Chatbots Fail to Engage (And What Gets 70%+ Voluntary Engagement Instead)
Vincent Mateljan
In June 2026, The Atlantic ran a piece on Khanmigo, Khan Academy's AI tutor. The headline number: access grew from roughly 40,000 students in 2023 to nearly a million today. The number that mattered more: usage didn't grow with it. Giving learners an AI tutor turned out to be a very different thing from getting learners to use one.
We hear the same thing from L&D teams evaluating AI tutoring for corporate training: strong access numbers, thin engagement. It's tempting to blame the model, the interface, or the rollout. We think the real cause is more structural, and the fix is already working in one of the least “chatbot-friendly” audiences imaginable: time-poor compliance lawyers.
The pattern: answer engines require the learner to go first
Most AI tutor deployments, Khanmigo included, are built as chatbots. The AI sits there, capable and patient, available any time. But it's a static resource, not an active teacher, until the learner opens the chat and asks something. That single design choice puts all the initiating work on the learner: they have to recognize they're confused, know how to phrase what they don't understand, and choose to type it (every time!), before any learning happens.
Add that to a training culture where “click next” is the norm because employees are protecting billable hours or clearing a checklist before day's end, and it's obvious why utilization stalls even when access doesn't. This isn't a Khanmigo problem or acorporate learning problem. It's what happens whenever an AI's value depends on the learner initiating the interaction.
The same shift is happening across all knowledge work
Ethan Mollick made a related observation in his June 30 newsletter, “The Twilight of the Chatbots.” His argument: as AI capability compounds, usage is shifting away from chatbots, where a human prompts, checks, and prompts again, toward agents that run on their own and don't need constant intervention to produce valuable work. Even inside organizations racing to deploy AI, the chatbot pattern is already the less valuable one; the systems doing the most useful work are the ones that don't wait to be asked at every step.
Training has been slower to make that shift than the rest of the business. Most “AI-powered” learning tools are still chatbots wearing a tutor costume: patient, capable, and entirely dependent on the learner walking in and asking the right question. It's the equivalent of hiring a brilliant tutor and leaving them in a room, waiting to be asked something, instead of having them lead the session.
The fix: flip who does the asking
The alternative isn't a smarter chatbot. It's an AI that leads: one that asks the learner questions, waits for a real answer, and responds directly to what that person actually said, rather than sitting passively until prompted.
We got to test this directly with LawCPD, who deployed AI-led “Instructors” (not answerengine chatbots) inside compliance courses for lawyers. It's hard to imagine a tougher audience: lawyers are famously time-poor, compliance training is famously the thing people click through as fast as possible, and (critically) the AI activities were entirely voluntary. No one had to engage with them to get compliance credit.
Over 70% did anyway.
Learners told us it felt like the AI was actually responding to their specific answers, not routing them down a generic script. They said it meaningfully improved how much they retained compared to a standard click-through module. None of that came from the AI being smarter. It came from the AI going first: asking the question, requiring an actual answer, and reacting to that answer in a way a static module or a wait-to-be-asked chatbot never can.
Why this matters beyond one case study
This maps to something learning science has argued for a long time: being asked to retrieve and produce an answer builds stronger retention than passively reading or receiving information (the “generation effect,” retrieval practice). A chatbot tutor only delivers that if the learner already has enough time, self-awareness and motivation to generate their own question, precisely the thing corporate training audiences are shortest on. An AI that asks first removes that barrier entirely. It doesn't require the learner to know what they don't know.
For instructional design teams, the implication is concrete: the AI-tutoring question isn't “which model” or “which chatbot interface”, it's “who initiates.” An AI-led activity that requires a real response, and reacts to it, will outperform a passive answer engine on the metric that actually matters: whether anyone used it.
For the people who own the training budget, the implication is about risk and ROI. Voluntary 70%+ engagement in a compliance context is a completion-rate and audit-defensibility story, not just an engagement story. An AI tutor with high access and low usage is often a liability dressed up as a feature.
What this looks like in practice
This is the design principle behind GenLearn: AI-led instruction that asks the questions, adapts to what the learner says, and builds practice into the flow of a course. No prompt engineering, no chatbot window learners have to remember to open. And it easily drops into the LMS or authoring tool you already use.
If your team has been disappointed by usage numbers on an answer-engine AI tutor, that's not a signal AI doesn't work in training. It's a signal you've deployed the wrong shape of AI.
Book a 30-minute demo to see what an AI-led instructor looks like inside a real compliance or skills course.
Sources: