AI in schools
AI in schools, deployed without losing student privacy
Most district AI pilots die in review, not in the classroom. This is the deployment model that survives the review, written for the person who has to sign it.
01 Why pilots stall
Four failure points, in the order they usually happen
Nobody can say what it did
Six weeks in, the question is what students learned. A usage dashboard answers a different question. Without a per-skill record there is no way to defend the line item, so the pilot quietly ends at renewal.
Teachers stop trusting submissions
A tool that hands over answers destroys the diagnostic value of every piece of homework. The staff room notices within a month, and the tool gets banned in the classrooms that needed it most.
Legal review finds an unclear data path
Minors work sitting in a general consumer product with a broad terms of service, on individual staff accounts, is the finding that stops a rollout. It is usually discovered late, after the enthusiasm is already public.
It needs a workflow nobody has
Anything requiring a new assessment window, a new login per lesson or a new grading routine does not survive week nine. The pilot succeeds with the enthusiasts and fails everywhere else.
02 The deployment model
Five commitments that make a rollout defensible
Each one maps directly onto a question your review process will ask. None of them are settings a teacher can get wrong.
-
01
Student-facing, hints first
The tool the students actually use defaults to a question rather than an answer. This is what keeps homework diagnostic and keeps the tool welcome in classrooms.
-
02
A teacher-facing artifact every time
Every session produces a per-skill record with the student reasoning attached, so the pilot can be evaluated on learning rather than on logins.
-
03
Over your own material
No content library to adopt. The evidence comes back in your standards, which is what makes it usable in a board report.
-
04
One data path, controlled by the school
Work goes to one vendor under one agreement, not into forty personal accounts. Not used for training, not sold, deletable on request.
-
05
No new workflow
No assessment window, no score entry, single sign-on where the district has it. If it needs a good day to work, it is not a rollout.
03 A pilot that produces an answer
Eight weeks, one question, three numbers
Most AI pilots are designed so that they cannot fail, which means they cannot succeed either. Decide the question before you start and decide what would make you stop.
The question we would use: does per-skill evidence change what teachers do the next day? Not engagement, not usage, not sentiment.
Weeks 1 to 2
One department, two teachers, one intervention block. Baseline: how long does it currently take to work out why a class got something wrong?
Weeks 3 to 6
Cards after every session. Teachers plan from the class summary rather than from the grade book. Nothing else changes.
Week 7
Ask the two teachers one question: did you teach anything differently because of a card? Collect examples, not ratings.
Week 8
Three numbers: proportion of sessions that produced a usable card, number of lessons changed as a result, and minutes per week returned to the teacher.
Decision
If teachers cannot name a lesson they changed, stop. That is a real answer and it costs eight weeks rather than a year.
The routines that make weeks 3 to 6 survivable are in AI classroom routines that survive a real school day.
04 For the review
The questions your reviewer will ask
Is student work used to train models?
No. Not for training, not for advertising, not sold. It is processed to produce the evidence card and for nothing else.
Who controls the records?
The school or program. We process on your instruction, and export or delete on request, including derived data.
What are you not certified for?
No SOC 2 report, no ISO 27001, no COPPA safe harbor seal, no signed state DPA. Listed in full rather than omitted.
The full version, including the FERPA school official exception and the COPPA school consent mechanism, is on the student privacy page. Staff-side guidance is in ChatGPT for teachers and where it breaks student privacy.
05 For the person who signs off
Two things a curriculum director asks first
Evidence you can present
Per-skill mastery per student, exportable, with the student's own reasoning attached. It survives a board slide and an MTSS meeting because the wording came from the student, not from a summary.
A privacy posture built for minors
Student work is not used to train models. No advertising, no data resale, deletable on request. Designed around FERPA and COPPA obligations, and our security page states plainly which formal attestations we do and do not hold.
Rollout, procurement and the full sign-off checklist are on the pricing page, alongside the cost-per-student estimator.
06 Pricing
Priced per program, not per seat
Annual billing is about 17% lower. There is no free tier, and the privacy posture is identical on every plan.
Classroom
$89/mo
One teacher, up to 60 students, CSV mastery export.
School
$349/mo
Up to 10 educators and 600 students, roles, SSO, scheduled reports.
District
$1,290/mo
Up to 75 educators and 5,000 students, standards mapping, SIS export, PO billing.
More AI tools for teachers
Send the pilot question to us first
Tell us what your district needs to be able to say at the end of eight weeks. If Coeducate cannot produce it, we will say so.
Student work is never used to train models. No card required to try the demo.