Getting Students To Engage With AI Tutors In Schools: What Actually Works

AI tutors in schools give students structured, teacher-directed, curriculum-aligned support through targeted practice, feedback and consolidation. For school leaders, the question is no longer whether the technology works, but which kind of AI tutor students will actually use well enough to learn from – especially if the goal is to narrow achievement gaps rather than widen them.

Khanmigo, the AI assistant built into Khan Academy, is a useful cautionary case. The tool was available, well-intentioned and built around a serious educational ambition, yet for many students it remained what founder Sal Khan called a “non-event.” They simply did not use it.

His image for this is a human tutor who sits at the back of a classroom and waits for students to ask for help. A few confident students make their way over; most carry on exactly as they were. That is an engagement problem, not a technology one.

This article looks at why student engagement breaks down, how different types of AI tutors create different implementation challenges, and what schools can learn from programs such as Skye, Third Space Learning’s spoken AI math tutor.

Unlocking Funding for Math Interventions

A free guide for district and school leaders on how to fund your school’s math interventions, including tips and advice on ensuring your interventions are as cost-effective as possible.

Download Free Now!

Key takeaways

• AI tutors may become part of the answer to the achievement gap.
• Adoption still depends on limits around sensitive student data, including risks of surveillance or data misuse, and on the fact that AI tutors may lack the emotional intelligence of human support.
• We cannot assume AI tutors are the answer until we interrogate the type of AI tutoring being provided to schools.
• The students who most need support are not always the students who will seek it out.
• The tools that look best in a demonstration are not always the tools that work best in schools.

Why focus on student engagement for AI tutoring?

The promise of AI tutoring is that more students, including those from disadvantaged backgrounds, receive structured, responsive support when they need it. Access alone will not deliver that promise, because a student who never engages with the tool learns nothing from it, regardless of how clever the technology underneath is.

And if AI tutors only work for students who are already confident, motivated and supported at home, they will not close the achievement gap. They may, as I have cautioned for some time, widen it.

The achievement gap for disadvantaged students that underpins all interventions

AI in education is as much an equity question as an edtech one. In math, the gap between disadvantaged students and their peers is one of the most persistent challenges in American education.

On the 2024 Nation’s Report Card, 55% of economically disadvantaged eighth-graders scored below NAEP Basic in math, up from 42% in 2015, compared with 26% of their peers. And the gap between the highest- and lowest-performing students is widening: scores rose for the highest performers and fell for the lowest.

What AI tutoring could offer

AI tutors could make something closer to one-on-one tutoring available to far more students, with extra practice, timely feedback and targeted support at a scale schools cannot reach through traditional methods. Tutoring itself has a strong evidence base: the Institute of Education Sciences describes high-quality tutoring as a research-proven approach.

Why access alone could widen the gap

But not every model of one-on-one AI tutoring will reach the students who need it most. An open, optional AI platform is most likely to be used by students who already seek help: those with quiet places to work, strong routines, confident parents, better devices and paid access to premium AI tools. The students most affected by the disadvantage gap may be less likely to access it, persist with it, or use it well.

A school-based, teacher-directed AI tutor can be a structural response to a structural problem, in a way that a freely available chatbot relying on student motivation never will be. So the question throughout is simple: does this model reach the students who need it most, or only those already likely to opt in?

Be clear what kind of AI tutor you mean

The phrase “AI tutor” covers too many different things.

It can mean an open chatbot, such as ChatGPT or Gemini, used by a student at home. It can mean one of the growing number of AI tutoring tools marketed by edtech companies. Or it can mean a structured, school-directed program with curriculum-aligned lessons, assessment and adult oversight.

Each of these is a different proposition, and they behave very differently once they are in front of students.

A chatbot answers, a tutor teaches

A chatbot that answers a student’s questions is doing something useful, but answering questions is a small part of what a tutor actually does.

A tutor diagnoses, prompts, checks, redirects, withholds answers when necessary, and keeps the student thinking. Good tutoring supports critical thinking; it does not simply produce answers on the student’s behalf. There should be no cognitive offloading or outsourcing.

In schools, tutoring also sits inside a wider system: state standards, assessment, student safety, behavior routines, teacher judgment, and school priorities.

So when school leaders evaluate AI tutoring tools, I would suggest the first question is a design one: what is the educational design here, and under what conditions does it need to work? How advanced the artificial intelligence is can come second.

What the tools that work have in common: a push and a pull

I have been writing about AI tutoring tools in schools for a while now, and watching which ones actually earn their place. The tools that work are rarely the cleverest models. They tend to get two things right at the same time.

The first is a push from the school. An adult decides the session matters, schedules it, gives it a purpose and checks what happened afterward.

The second is a pull from the program itself. The design is good enough for an AI tutor to draw a student in, keep them thinking, and show them they are getting somewhere.

Access on its own is neither a push nor a pull, and that is the gap Khanmigo fell into.

What follows are the ingredients that make up each one.

The push: what the school and the tutoring program coordinator must drive

This is the half the school controls – everything the tutoring program coordinator and teachers put in place around the tool, before and after each session.

Do not rely on students asking for help

The clearest lesson from Khanmigo’s early challenge is that schools should not design AI tutoring around voluntary help-seeking, where students are left to initiate support for themselves.

Many students who need help the most are the least likely to ask for it. Some lack confidence or suffer from math anxiety. Some do not know what they do not know. Some avoid tasks that reveal difficulty, and some have learned how to look busy while doing very little thinking at all.

Anyone who has taught a class will recognize these students, because this is a normal classroom context problem that existed long before AI technology arrived. Putting the support behind a screen does not automatically solve it.

A student who avoids asking a teacher for help will not necessarily become self-regulating because the helper is digital. In some cases, the screen may make avoidance easier, because students can guess, click through, stay quiet, test the system, or give just enough of a response to move on.

This matters most for students receiving an MTSS intervention, particularly those from disadvantaged backgrounds. If the model depends on students recognizing their own gaps, choosing to log in, asking precise questions and persisting when they are stuck, it is likely to work best for students who already have many of the habits and supports we are trying to build.

The weaker model: “Here is an AI tutor. Use it when you need it.” The stronger model: “Your teacher has scheduled this session for you because it connects to something important you need to secure. An adult will know whether you completed it and whether it helped.”

Structure is what makes student engagement more likely, and it costs nothing to build in from the start.

Some students may find AI tutoring lower-stakes

Some quiet or anxious students may engage more readily with an AI tutor than with an adult or peer, and from the student’s point of view, this makes perfect sense.

In a classroom, asking for help can feel public. Getting something wrong can feel embarrassing, and explaining your thinking in front of others can feel risky.

For some students, speaking to an AI tutor may feel lower-stakes. They can try, pause, correct themselves and ask again without feeling watched in the same way.

To be clear, this is about confidence, context and participation, and has nothing to do with the debunked idea of “learning styles.” Different students respond differently to different classroom interactions, and some are more willing to attempt difficult concepts when the interaction feels less public.

That has a practical implication for how schools select students. Some tools can also lower participation barriers through real-time translation and assistive technology for students with additional needs. Rather than selecting only from achievement data, schools should also consider students who under-participate in class, rarely ask questions, or know more than they are willing to say publicly. For some of those students, AI tutoring may provide a quieter route into mathematical talk.

I would not overstate this, though. AI tutoring does not replace relationships, teacher encouragement or classroom belonging, but it can offer another route into participation for students who find the classroom a difficult place to speak up, and that is worth having.

Connect tutoring to real learning goals and state standards

Students are more likely to engage when the tutoring connects to something that matters to them.

“Use this AI tutor” is weak. “This session will help you with equivalent fractions because that is what you struggled with in Friday’s lesson” is stronger, and “this is part of your state test prep plan” is stronger still.

The goal might be state assessment prep, end-of-year exam prep, the current unit, or a specific skill gap identified by the teacher. The important thing is that the tutoring is not happening in isolation from the rest of school, and that students understand why they are doing it now and how it connects to what they are learning in class.

This matters for disadvantaged students because intervention time is precious. If a student is missing part of the school day, homeroom, class time, or another intervention to take part in a tutoring session, that session needs a clear purpose. It should connect to what the student is learning, what they need to work on next, and what the teacher will do afterward.

It also matters for teachers. If AI tutoring is linked to curriculum priorities, teachers are more likely to value it. If it becomes another platform running in the background, it will be easier to ignore, and far less likely to support stronger student performance.

Routines and room setup matter

Schools need to think about the ordinary practical details. Where will students sit? How noisy will the room be? Who will supervise? What should students do if the tutor misunderstands them? How should they speak? What counts as good participation? What happens if they test the system?

These details can look minor until they undermine the intervention.

Imagine a quiet 5th-grade student in a busy room. The room is too noisy, the student speaks softly, and the AI tutor repeatedly misunderstands them. The student becomes frustrated, gives shorter answers, and eventually stops trying to explain. From a distance, that might look like poor motivation. In reality, it is a poor setup, and it was avoidable.

How students are set up really matters.

A five-minute launch routine

A five-minute launch routine before the first session can make a difference. Students need to know how to speak, how to listen, what to do when they are stuck, and why explaining their thinking matters. They also need to know that an adult will notice the quality of their participation, not just completion.

Many tutoring tools fail because the routines around them are weak, even when the idea behind them is perfectly sound, and AI tutoring is no different in this respect.

Students need to be taught how to use it as a learning activity, not just shown where to log in. Those example classroom interactions matter because students need to know what good participation sounds like. Strong systems are built with educators using those examples and clear scoring criteria aligned to the curriculum, so expectations are clear from the start.

Teach students how to get the most out of the intervention

Schools should assume students need to be taught how to engage with AI tutoring. The goal is to think, explain, respond to feedback and improve, and students will not arrive knowing that unless someone tells them.

Schools should also anticipate predictable behavior.

• Some students will say as little as possible.
• Some will guess repeatedly.
• Some will test boundaries.
• Some will try inappropriate language.
• Some will look as though they are participating while doing little thinking.

Every one of those behaviors will be familiar from ordinary lessons, and each is a teaching opportunity in exactly the same way. The answer is the same as elsewhere: clear routines, adult supervision, high expectations and follow-up when students do not engage properly.

A simple first-session routine might look like this. The teacher explains the purpose of the tutoring, links it to a real math goal, models one strong response and one weak response, then tells students what adults will check afterward: completion, effort, mathematical explanation and progress from check-in to check-out.

Those example classroom interactions matter because students need to know what good participation sounds like, and that depends on good implementation from the adults.

Keep an adult in the loop

AI tutoring should not mean removing the teacher.

Teachers still need to decide which students receive support, what the learning goals are, how sessions connect to face-to-face teaching, and what happens next. They need to know whether students are engaging properly, whether the content is appropriate, and whether the tutoring is improving understanding.

Without adult oversight, AI tutoring can produce a false sense of progress. A student may complete sessions without learning securely. Another may guess repeatedly. Another may need a different explanation, more concrete examples, visual explanations, or direct adult intervention.

Teacher-directed tutoring is one of the conditions that makes AI tutoring more likely to work.

The AI can deliver structured student interactions, targeted feedback and immediate prompts, while the teacher makes sense of what is happening and connects it to the wider curriculum.

The teacher is not being replaced here. They are delegating the part of the labor an AI tutor can handle, and the expert remains in the loop at all times.

Bring teachers with you

Teacher engagement matters as much as student engagement, because if staff hear “AI tutor” and think “replacement teacher,” implementation will be fragile from the beginning.

School leaders need to be clear with their staff. AI tutoring is not a substitute for high-quality teaching. It is a way to provide additional structured practice and feedback where schools currently struggle to provide enough, which makes it an argument about teacher workload and equity.

Teachers and school leaders overseeing rollout need to know which students are using the tutor, why those students have been selected, what content they are working on, how it links to classroom learning, what information teachers will receive, and what they should do when students are not engaging or not improving.

Teachers should also have a stake in what is being taught. They should know why a topic matters for a particular child, how it is being taught, and whether the methods and approach match the ones used in school.

If teachers cannot see the connection between the tutoring and their teaching, they are unlikely to value it.

What a classroom teacher needs to see

A classroom teacher does not need another dashboard just for the sake of having one. They need usable information. Which students attended? Which students engaged properly? Which questions did they struggle with? Which concept needs reteaching? Which students are ready to move on?

What a math or intervention leader needs to see

A math or intervention leader needs a wider view.

• Is support reaching the right students across different educational settings within the school?
• Are disadvantaged students receiving enough sessions, especially given what we know about the importance of high-dosage tutoring as well as high-impact tutoring?
• Are the students with the greatest learning gaps attending consistently?
• Are there classes or grade levels where engagement is weaker?
• Is the program supporting existing curriculum priorities?

Give teachers a real role here: identifying learning gaps, setting goals, reviewing progress and deciding what happens next, rather than sitting as passive monitors of a product.

The pull: what a well-designed program provides

This is the half that depends on the AI tutoring program’s design – whether it is built to draw students in and keep them thinking, or simply left open for them to use.

AI tutoring tools need educational guardrails

The best AI tutoring tools are not answer machines. An answer machine gives students the next line, whereas a tutor decides what the student needs to think about next.

In math, students often need a hint, a worked example, a prompt to explain their reasoning, or a carefully chosen follow-up question. If the tool gives the answer too quickly, it can remove the very thinking the student needs to do.

This is why effective intelligent tutoring systems need educational guardrails. The AI should not be the curriculum designer. It should work inside a curriculum-aligned sequence built by people who understand the subject, state standards and the common errors and misconceptions students are likely to bring with them.

A weak AI tutor can accept vague reasoning. It can generate a plausible but unhelpful explanation, choose examples that do not expose the mathematical structure, or move students on before they have understood the key idea.

Good tutoring has a shape. It models, guides, checks and gradually releases responsibility, helping students build conceptual understanding and deeper understanding without hiding the difficulty of the content.

Some AI tutoring tools are open-ended chat interfaces based on large language models. Other AI tutoring tools are structured programs where the AI delivers educational resources designed by teachers and subject experts. Schools need to know which one they are buying.

Make progress visible quickly

Students are more likely to engage when they can see that what they’re doing is helping them.

That does not mean making the work easy. It means designing the learning process so students experience genuine movement: from uncertainty to partial understanding, and from partial understanding to success.

Motivation often follows success rather than preceding it, which is particularly important for students who have experienced repeated failure in math. For some students, intervention can feel like another reminder that they are behind. A well-designed tutoring session should reverse that experience quickly, so that students feel, “I can do more now than I could at the start.”

That is why diagnostic skill check-ins, progress tracking and skill check-outs matter. They give students and teachers a visible measure of progress within the session. The student can see improvement, and the teacher can see whether the session had the intended effect.

Students are more likely to persist when the session gives them evidence of improvement.

In an independent analysis of Skye, Third Space Learning’s spoken AI math tutor, by Educate Ventures Research, researchers analyzed 9,320 student sessions and found students improved within the session from 34% correct on diagnostic check-in questions to 92% on check-out questions. In sessions that included confidence prompts, 63.8% of students ended more confident than they began. Educate notes they are not yet estimating standardized attainment impact; that would require a controlled design.

Be cautious about personalization and gamification

There is an appealing argument that AI tutors can offer personalized support and round-the-clock help, which is especially useful for remote learners or students stuck on homework late at night, by adapting every question to a student’s interests.

A football fan gets football questions. A gamer gets gaming questions. A student who likes baking gets baking questions.

Used sparingly, that may reduce friction at the start of a task, particularly for students who are reluctant to begin. But as a model for learning, I am cautious.

The risk is that we confuse interest with understanding. Students do not only need to solve problems in contexts they already like; they need to recognize the underlying mathematics across unfamiliar contexts, which is why word problems and problem-solving are introduced in the early elementary grades and become increasingly important as students progress through school.

The same caution applies to gamification. Points, badges and streaks can enhance engagement, but gaming is not the same as learning, and a student can be very engaged in beating the system while doing little mathematical thinking.

The form of engagement worth designing for is the feeling of becoming more successful at something that matters. Points and badges are decoration around that.

What Skye shows about the direction of travel for effective AI tutoring

Skye is a useful example of what we have discussed in this article about engagement with AI tutors, because it shows what happens when the AI is placed inside an explicitly designed tutoring structure.

Skye sessions are school-scheduled and teacher-directed, so students do not self-access without teacher intent. Lessons follow an “I do, we do, you do” structure created by qualified math teachers, which is better suited to complex subjects because it breaks learning into manageable steps.

Skye does not generate its own curriculum content; it delivers curriculum-aligned lessons built by people with subject expertise. Each lesson begins with a diagnostic skill check-in and ends with a skill check-out.

That addresses several of the risks discussed above. It does not rely on students voluntarily seeking help. It does not make the AI the curriculum designer. It builds in assessment, makes progress visible, and keeps the school in control of who receives support and why.

That is a different model from giving students access to a chatbot and hoping the right students use it well.

To see exactly how AI tutoring with Skye works, you can watch AI tutor Skye in action.

Six questions that separate AI access from AI tutoring

Before adopting any AI tutoring tool, school leaders should ask:

1. Does the tool wait for students to seek help, or is it scheduled by the school?
2. Does it generate content freely or deliver a designed curriculum?
3. Does it use worked examples, scaffolding, immediate feedback and checks for understanding?
4. Does it show learning progress, or only usage data?
5. Does it support disadvantaged students, or mainly the students most likely to opt in?
6. How will adults monitor engagement, progress, student safety and whether it engages students in explanation rather than shortcutting thinking?

The questions hold wherever the tool is used – in school, at home, or across other educational settings – and they keep the focus on the quality of the learning, and the teacher behind it, rather than the novelty of AI-generated content or access to AI systems.

AI tutoring in schools is most likely to improve learning outcomes when it is structured, teacher-directed, curriculum-aligned, consolidation-focused and available to the students who need it most, so schools are choosing a safer, more effective learning experience rather than just more access.

Do you have students who need extra support in math?

Skye – our AI math tutor built by experienced teachers – provides students with personalized one-on-one, spoken instruction that helps them master concepts, close skill gaps, and gain confidence.

Since 2013, we’ve delivered 2,295,000 math tutoring sessions to 196,000 students, guiding them toward higher math achievement.

Discover how our AI math tutoring can boost student success, or see how our math programs can support your school’s goals:
– 3rd grade tutoring
– 4th grade tutoring
– 5th grade tutoring
– 6th grade tutoring
– 7th grade tutoring
– 8th grade tutoring

This blog was originally written by Neil Almond, a deputy headteacher in the UK. It has since been checked and published in the US.