What Does a School With an AI Tutor Show at Renewal Time?

The Problem

Anthropic, OpenAI and others are shipping educator and student tools for free or at steep discounts, buying their way into a market estimated at $6tn (Financial Times).

At a price near zero, adoption is easy. The hard part comes one term later, when someone asks what students can actually do now. All the school can produce is transcripts and usage counts. How many students used it how often is knowable. Whether that use was learning or ghostwriting is not, at least not from those logs.

For teachers it is stranger still. Submitted work clearly got better. Yet sit the same student down with a comparable problem and the result sometimes contradicts the submission. Right now that gap lives entirely in a teacher’s intuition, with no way to record it.

Existing learning analytics products are a generation behind this problem. They track logins, time on page, submission rates. Every one of those metrics predates AI in the workflow. When a student spends 40 minutes with a model and 5 minutes on the assignment, time on page says nothing.

Why Now

Free distribution created a clock. Free and discounted programs are usually scoped to a term or a school year, and when they lapse, a renewal meeting happens. That meeting needs outcomes, not usage, and no product on the market builds that document.

The technical preconditions are in place. Schools already hold assignments, grades and attendance in an LMS, and AI tools emit transcripts. Both sides expose APIs. What is missing is not data but the layer that joins them and renders a judgment.

Policy is moving the same direction. The debate has shifted from blocking teen access toward opening it safely (OpenAI), and any decision to open comes with a reporting obligation attached. At the same time, kids are far from uniform in how they treat these tools (WIRED), which means average usage numbers hide everything worth seeing.

How to Build It

The core move is not collecting new data. It is joining two streams that already exist.

flowchart LR
    A[AI tutor transcripts] --> C[Concept tagging]
    B[LMS assignments and grades] --> C
    C --> D[Auto-generated unaided checks]
    D --> E[Per-student concept mastery call]
    E --> F[Renewal meeting report]

Keep the MVP narrow: one subject, one term, roughly twenty concepts. Tag which concepts a student wrestled with in the transcript, then auto-generate short unaided checks on those same concepts. The delta between concepts they got help on and concepts they can now clear alone is the report.

The stack can be ordinary. LTI plus vendor APIs for the LMS side, a small model for concept tagging. The hard part is not the model but who builds the concept map. Early on you build it by hand with teachers, and as schools accumulate, that map becomes the moat.

Success Criteria

The first assumption is whether renewal meetings genuinely need this artifact. Ask three schools what they submitted at last term’s meeting and you will know within a week.

The second is whether teachers read the report as surveillance. The moment it looks like a per-student scoreboard, adoption stalls. Start at the class level, showing where the whole room has a hole, and open individual calls only when a teacher asks.