AI instructional coaching, honestly.

By Yechiel · Founder, PD That Works. Yechiel builds PD That Works — the workshops, the daily AI coach, and the Day 30/60/90 reports — and wrote every page on this site.

"AI instructional coach" is a phrase about three years old and already attached to a dozen different products. Some write lesson plans. Some score classroom video. Some send a teacher a message every day and remember what she said last week. They are not the same thing, and a school evaluating them deserves to know which one it's buying. Here's the category, sorted — plus what an AI coach genuinely can't do, and the research a defensible one is built on.

What "AI instructional coaching" means

Strip the marketing and it's this: software that supports an individual teacher's practice over time, in conversation, rather than delivering content to a group. That's the same job description as human instructional coaching — an individual, sustained relationship focused on one teacher's practice — with software doing part of it.

The word doing the work is coaching. A tool that answers questions is a search box. A tool that produces materials is a generator. Coaching means someone asks you about the thing you're trying to change, remembers your answer, and comes back tomorrow.

Three different products wear the name

Content generators. Lesson plans, rubrics, worksheets, unit outlines. These are genuinely useful and save real hours. They aren't coaching: a generator has no memory of your practice and never asks how the lesson went. What to ask: does it know anything about me between sessions?

Observation and analysis tools. Classroom video or transcripts, scored against a framework — talk ratio, question types, wait time. Closer to coaching, and useful for a coach's evidence. Two cautions: they require recording a classroom, which is a trust decision before it's a technology decision, and analysis is not the same as support. What to ask: who sees the output, and does anything happen after the score?

Daily follow-up coaching. A short written exchange on the days a teacher teaches, tied to one goal that teacher set, with memory across the cycle. This is the layer that matches ESSA's "sustained, job-embedded" definition and the one that fills the gap between coaching touches. It's what PD That Works builds. What to ask: what happens on an ordinary Tuesday?

A school can want all three. They come out of different budgets and solve different problems, and a vendor that blurs them is worth slowing down for.

What an AI coach cannot do

Being clear about this is the fastest way to tell an honest vendor from an eager one:

• It cannot see your classroom. An AI coach reads what a teacher writes. It does not watch fifth period, notice the kid in the back, or catch the thing the teacher didn't think to mention.
• It cannot model a lesson. Demonstration is a human act.
• It cannot have the hard conversation. A teacher whose practice isn't working needs a person with standing and judgment.
• It cannot replace a mentor for a new teacher. Induction is relational, and first-year teachers need someone in the building.
• It cannot promise student outcomes. Anyone selling test-score gains from a coaching chatbot is selling something the evidence doesn't support.

What it can do is the part no human coach has hours for: show up every teaching day, for every teacher, and remember. We make coaches better, we don't replace them — that boundary is the product, not a disclaimer.

The research a defensible AI coach is built on

Every design decision in our coach traces to a named finding. Four of the load-bearing ones — and what we built because of each:

Recording progress toward a goal, and reporting it to someone, measurably raises goal attainment. — Meta-analysis of 138 randomized trials (19,951 participants), Psychological Bulletin
What we built because of it: The daily check-in is exactly that shape: one goal, progress in writing, reported to a coach — minutes a day.

Over a third of feedback interventions make performance worse. The dividing line: feedback aimed at the practice helps; feedback aimed at the person — including praise — backfires. — Meta-analysis of 607 effect sizes, Psychological Bulletin
What we built because of it: The coach never grades, scores, or praises the person. It talks about what the teacher tried and what happened.

AI that flatters measurably degrades people's judgment — while raising their satisfaction with it. — Controlled experiments, Science
What we built because of it: The coach is free to not affirm, and teacher-satisfaction ratings are never used to tune it. Ever.

Teacher coaching is among the most effective PD forms ever measured — and more coaching hours did not mean better outcomes. Specificity beats dosage. — Meta-analysis of 60 causal studies of human coaches, Review of Educational Research
What we built because of it: One specific practice goal per 30-day cycle, minutes a day — not hour-long sessions, and no inflated promises.

Note what the last one concedes: those 60 studies are of human coaches. The AI trials support goal attainment, and that's the only claim we make.

The honest-AI stance

Three commitments, each of which costs us something:

It says what it is. "I'm an AI, and I won't pretend otherwise." No persona, no invented name, no pretending a person wrote the message.

It doesn't flatter. The sycophancy finding above is the uncomfortable one for this whole product category: the flattering assistant scores better on satisfaction surveys. So we don't use satisfaction ratings to tune the coach — a deliberate choice to be liked slightly less and useful somewhat more.

Teachers' words stay theirs. Reflections are private to the teacher; leaders see trends, never words. Unreadable to them by construction, not by policy — five teachers is the minimum for a privacy-safe pattern. That boundary is what makes a teacher's 10pm answer honest, and an honest record is the only kind worth aggregating.

You can test all three yourself: the coach answers questions about itself — how it coaches, what research shaped it, who sees what. No salesperson in between.

What to ask any AI coaching vendor

1. Does it disclose that it's an AI, plainly, to teachers?
2. Does it praise? If it scores or flatters, the feedback research says it may be making things worse.
3. Is satisfaction used to train it? A model tuned on "did you like that" drifts toward flattery.
4. Who can read a teacher's words — and is the boundary structural or a policy someone could change?
5. What does it remember, and for how long?
6. What does it refuse to claim? A vendor with no list of things it doesn't do hasn't thought about it.

Common questions

What is an AI instructional coach?

Software that supports an individual teacher's practice over time in conversation — asking about a goal the teacher set, remembering the answer, and returning the next teaching day. It differs from a lesson-plan generator, which produces materials but has no memory of your practice, and from an observation tool, which scores a recorded classroom.

Can AI replace an instructional coach?

No. An AI coach reads what a teacher writes; it cannot observe a classroom, model a lesson, or hold a hard conversation. What it does is the daily follow-up a human coach has no hours for — a five-minute check-in for every teacher, every teaching day, with memory across the year.

Is AI coaching for teachers effective?

The strongest evidence is for the mechanism rather than the product category: a meta-analysis of 138 randomized trials found that monitoring progress toward a goal and reporting it to someone raises goal attainment, and a meta-analysis of 60 causal studies found human coaching among the most effective PD forms measured. Those coaching studies are of humans; the honest claim for AI is goal attainment, not replicated coaching effects.

Do teachers trust an AI coach with what they write?

They do when the privacy boundary is structural rather than promised. In PD That Works, reflections are unreadable to school leaders by construction; leaders see engagement and themes across five or more teachers, never an individual's words.