Can AI Coach a Teacher? The Honest Answer Is No

Why not

Coaching runs on decisions only a person who knows you can make. Deciding what a particular teacher should work on in a particular year, with the particular twenty-two children in front of them, depends on knowing all three. So does knowing when to push and when to leave something alone for a month. And the feedback lands in the first place because of who it is coming from. A stranger with a clipboard and a friend who has watched you teach for three years can say the identical sentence and only one of them gets heard. We are not going to build a company on pretending otherwise.

But "can AI coach a teacher" is one question wearing a coat over about five smaller ones. Some of those smaller questions have an answer that is genuinely yes. Sorting them out is the whole point of this piece, and it is more useful than either the sales version or the eye-roll.

Coaching works. That part is settled.

Start with what is not in dispute.

In 2018, Matthew Kraft, David Blazar and Dylan Hogan published a meta-analysis in Review of Educational Research combining 60 studies that used causal research designs. Pooled across all of them, teacher coaching moved instructional practice by 0.49 standard deviations and student achievement by 0.18.^[1]

In education research those are large numbers. Coaching is one of the better-evidenced things a school can do for teaching quality, and it outperforms the workshop model most professional development still runs on.

Then comes the sentence in that paper that gets quoted far less often. The authors found that average effects from effectiveness trials of larger programs were only a fraction of the effects found in efficacy trials of smaller ones.

That finding has a limit worth naming, and we named it two weeks ago in Instructional Coaching for Small Schools. The smaller programs in the review were often run close to the researchers who designed them, so size was not the only thing that differed, and none of this proves small is better. What it does say is that coaching is not durable when it gets spread thin.

Read that as a practical fact rather than a research caveat. The method is not the problem. Reach is. When a coaching program has to serve a lot of teachers, the ingredient that gets thin is the individual attention, because there is only so much of a human coach to go around. This is the same arithmetic we described from inside a building in The Quiet Problem in Private Schools, just measured.

So the real question was never whether a machine can replace a coach. It is whether anything can carry the part of coaching that keeps evaporating when a coach is stretched across forty teachers.

What happened when a model was left to do the whole job

In 2023, Rose Wang and Dorottya Demszky ran the experiment that everyone talking about AI and teacher growth should have to read first. They gave a general-purpose chatbot three jobs on real elementary math transcripts: score the segment against standard observation instruments, name the instructional highlights and the missed opportunities, and suggest an actionable next move for student reasoning. Expert math teachers judged what came back.^[2] The setup is worth naming plainly, because it is a real design rather than a straw man: one teacher, one model, and nobody in between.

The output was relevant. It was also, in the researchers' assessment, rarely novel or insightful. The finding worth carrying around is this one. 82 percent of the model's suggestions pointed to places in the transcript where the teacher was already implementing that suggestion.

Picture the human version. Someone watches your lesson, sits down across from you, and advises you to try asking more open-ended questions, referring to the exact minute you asked one. You would not schedule a second session.

Bound this fairly. That was a general-purpose model, used with no special setup, in 2023, on one subject. It is not proof that no machine can ever help a teacher. What it is evidence about is a shape, the one where the machine is the entire loop and there is nobody to catch it when it says something useless. It is clean proof of something narrower and more useful: fluency about teaching is not the same as knowing what happened in your room. A system can produce paragraphs that sound exactly like coaching and contain nothing you can act on.

What a machine did move

Now the other side, because an honest answer owes you both.

Demszky, with Jing Liu, Heather Hill, Dan Jurafsky and Chris Piech, later ran a randomized controlled trial with 1,136 instructors in a large online course. The tool did one thing. It gave automated feedback on a single practice, uptake of student contributions, which is the habit of taking what a student actually said and building the next move on it rather than moving past it. Instructors who got the feedback improved their uptake by 13 percent, with early signs of better course satisfaction and assignment completion among their students.^[3]

Bound this one too, and hard. Those were adult learners in an online course, not eight-year-olds on a rug. One practice, not teaching as a whole. Do not let anyone tell you it means an app raised test scores in an elementary classroom, because it does not.

What is worth taking is the shape of the win rather than the size of it. The system did not out-think anybody. It watched for one specific, observable thing, described it accurately, and came back with it every single time. Narrow, concrete, and relentlessly consistent turns out to be a real contribution. Broad and wise turns out to be a bluff.

Five parts of the job, and who should do which

Here is the honest breakdown. A coaching cycle is really five separate jobs, and they do not have the same answer.

Notice what happened. A machine can do this, and better than any of us. Your memory of third period at 2:45 on a Thursday is a summary written by someone who was also running the class. A recording is not.

Describe it without a verdict. A machine can do this too, arguably more easily than a person can, because it has nothing at stake in the room and no history with you. Describing rather than judging is the distinction the whole series keeps returning to, and it is laid out in Coaching vs. Evaluation.

Decide what matters most right now. Human. This depends on whether you are in your first year or your nineteenth, what your class did last September, what you are already carrying, and the child in the third row whose situation is not in any transcript.

Choose the next move, small enough to try on Tuesday. Mostly human. A system can lay out options. A person who knows you picks the one you will actually do.

Come back and ask how it went. A machine can do this, and this is where it earns its place. Step five is the step that disappears in every school, every year, for reasons nobody is at fault for. The workshop ends, the observation gets filed, everybody's calendar closes over the top of it. We made the same argument about the first weeks of the year in The First Weeks of School Set the Whole Year, and about doing it alone in Reflection Is the Most Underrated Teaching Skill.

Three of the five are recall, description, and showing up. Those are machine work, and they are also, not coincidentally, the three that thin out first when a human coach is stretched. The other two require knowing a person. Those stay with people, permanently, and any product that says otherwise is selling you something.

Five questions worth asking any tool that says it coaches

Take these to any demo, including ours.

Who reads what it says before it reaches me? If the answer is nobody, you are not being coached, you are receiving output. Ask them to name the human and describe what that person actually does.

Does it describe, or does it score? A number attached to your teaching is an evaluation no matter what the marketing calls it. Ask to see the exact thing a teacher receives.

What happens the day a lesson goes badly? Ask what the system does with your worst twenty minutes. Who sees it, whether it follows you into a review, and whether you get to say it was not a fair sample. If nobody can answer that cleanly, you will teach the version of yourself that looks good, and the tool will have made you worse.

Does it come back? Ask what happens on day eight. If there is no answer, you have bought an observation, not a cycle.

What happens when it is wrong? Ask what you do when the description does not match what happened in the room. Whether you can say so, whether a person hears it, and whether anything actually changes as a result. A system that cannot be corrected is not in a relationship with you.

How eduimpact fits, honestly

We do steps one, two, and five. That is the whole claim.

eduimpact watches, describes what happened in plain language rather than a rating, and then comes back to ask how the change landed, and keeps asking. Steps three and four, deciding what matters and choosing your next move, stay with you and with the coach you have, if you have one.

Here is the part that is easy to miss. The system does not arrive knowing your school. It learns it, and the same loop does the learning whether there is a coach down the hall or you are doing this on your own.

A person reviews what eduimpact produced and grades it against how teaching is meant to look where you are. The model is graded, never the teacher. eduimpact organizes what that review flagged into focal points, so whoever leads the next coaching conversation opens it with a prioritized place to start instead of a blank calendar. The conversation itself stays human. And what sharpens the system is not the conversation. It is the grading. Every response marked wrong, or right, or nearly right goes back into the knowledge base you own, so the next cycle starts from a model that has already been told where it missed.

That is the loop, and it means the tool is less useful in month one than it is in month six, which is a strange thing to advertise and also the honest description of anything that has to learn your building before it can be worth much in it. Your content stays yours.

A person reviews what the system surfaces, so you are never working from a black box. Nothing is rated. Nothing is ranked. Nothing goes to whoever decides your contract. Student data can stay inside your building. And there is a private journal in it that no other person can open. Not your coach, not your principal, not us. The system itself does read what you write there, and that is deliberate: it is how the response you get can account for the week you are actually having instead of the one on paper.

It is an extension of coaching, not a coach. If your school has an instructional coach, this gives them reach into weeks they cannot physically get to. If your school has nobody in that role, it is a second set of eyes and an honest description of your own teaching, which is more than most teachers have ever been handed.