Anyone who works through the literature keeps arriving at the same conclusion: one-on-one or small-group tutoring is among the most effective interventions we know. Meta-analyses of dozens to hundreds of studies show consistent and often relatively large effects. That’s not a marginal finding; it’s one of the most robust results in educational research. What’s more, it’s one of the key ways to address inequality.
But there’s an important nuance beneath that consensus. Tutoring doesn’t just work automatically. It works especially when it happens under fairly specific conditions. High frequency, consistent attendance, well-trained tutors, and ideally embedded in the school day, these are all factors that increase effectiveness. And that brings us to what’s called “high-dosage tutoring.” That’s not a minor detail: when that level of intensity is reached, the effects are considerably larger. But that’s also where the problems start. To be clear: not all tutoring needs to be high-dosage, but the strongest evidence for effectiveness sits precisely there. And that’s exactly where things go wrong: we often organise something that looks like it, but isn’t quite.
A recent study by Shmoys and colleagues doesn’t ask whether tutoring works, but rather something that may be more practically relevant: do we even manage to organise it as intended? The answer is uncomfortably simple: no.
Across five programmes spanning three tutoring formats (face-to-face, live online, and AI-driven), not a single programme reached the “high dosage” threshold. Students received fewer sessions than planned, fewer hours than intended, and often less consistently, too. That’s already interesting in itself. Because it shows the problem isn’t a lack of good ideas; it’s in their execution.
What makes this study strong is that it looks beyond the numbers. Through interviews, observations, and surveys, the researchers try to understand why things go wrong. And what emerges is both recognisable (I once set up a tutoring programme with colleagues for a research project myself) and confronting.
The first problem is simply logistical. Think of:
- Finding time in an already packed timetable.
- Finding a quiet space.
- Having students available at the right moment.
It may all sound too mundane for words, but that’s precisely where things most often fall apart.
The second problem is quality. Not every tutor is equally strong, and short training sessions don’t always compensate for that. With AI tutoring, the problem shifts: the frustration is that the system provides insufficient support for certain aspects of learning, such as genuine comprehension.
The study finds that student engagement is strongly linked to the relationship with the tutor, which doesn’t surprise me at all. Tutors deliberately invest time in getting to know students, in small conversations, in building trust. That’s not a side issue; it’s a mechanism. And it’s precisely here that AI tutoring hits a wall. Students prefer reading with a real person over reading with a computer. That sounds anecdotal, but it points to something more fundamental. As any good teacher knows, learning is not only cognitive; it is also, and very much, relational.
Many of the problems described in the study stem from additional demands placed on teachers. They have to free up space, adjust plans, troubleshoot technical issues, and communicate with tutors, on top of everything they already do. And that may be the core of the problem. We have an intervention that works, but that only works when it’s well organised. And that organisation requires time, structure, and support that are often simply not there.
That makes this study a valuable companion to the meta-analyses that keep confirming tutoring is effective, or rather: can be. It may be a hard lesson, but it’s an important one.