A 90-Day Classroom AI Implementation Playbook

The case study

This is the playbook one independent middle school used to onboard an AI math tutor to its entire 6th-grade cohort across a single semester — without losing a single family to the opt-out form and with measurable learning gains by week 12. Names are anonymized; the steps generalize.

The school had three constraints: a small IT team, a vocal parent community, and a head of school who had publicly said the school would "not be reckless with AI." The plan was built around those constraints, not around the tool.

Days 1-15: Foundation

Form the working group. Five people: assistant head of school (chair), 6th-grade math lead, IT director, a parent representative recruited from the existing technology committee, and one 7th-grade student who had used the tool the previous summer.

Inventory current AI usage. A 6-question anonymous teacher survey surfaced four tools nobody on the leadership team knew were in use. Two were retired immediately; two went into the formal review queue.

Pick the pilot tool. The group chose a single math tutor with a SafeGradeAI Teen tier and a signed school-official DPA. Selection criteria, in priority order: privacy posture, alignment to the existing math scope and sequence, teacher workflow fit, and cost.

Draft the disclosure letter. One page, two reading levels, sent to the working group for sign-off before any family received it.

Days 16-45: Pilot in two classrooms

Scope: two of the four 6th-grade math sections, three 30-minute sessions per week, all in-class with the teacher present.

Cadence: weekly 20-minute teacher debrief. A shared doc captured every hallucination, every awkward moment, and every "huh, that actually worked" observation. By week three, the log had 47 entries; 9 were action items.

Family communication: a mid-pilot survey at day 30 with three questions: comfort level (1-5), one thing they liked, one thing that worried them. Response rate 71%, net comfort score 4.2 of 5.

Student voice: the student rep ran a 15-minute focus group with eight pilot students. The most useful insight: students wanted clearer signaling about when the tool was "guessing" vs. "knowing." The team built that into the in-class scaffolding.

Days 46-75: Iteration

Prompt scaffolding. The team built a one-page "ask the tutor" guide that students kept in their math folders. It cut off-task usage roughly in half without limiting the tool itself.

Disclosure refinement. The original family letter underweighted the human-teacher role. The revised version led with what the teacher does and treated the tool as the assistant.

Interim publication. A 2-page interim findings memo went to all 6th-grade families and to the full faculty. It was honest about a hallucination incident and explained the new student-side check.

Days 76-90: Decide

The bar for "expand": measurable engagement, no unresolved safety incidents, family comfort holding above 4.0, and teacher willingness to continue at full load.

All four were met. The tool rolled out to the remaining two sections at the start of the next term. The working group remained in place as the standing AI committee.

What worked

• Naming a single chair who owned the timeline.
• Putting a student and a parent on the working group from day one.
• Treating teacher debriefs as data, not as venting sessions.
• Publishing the interim memo before families asked for it.
• Setting "stop" criteria in writing, not just "expand" criteria.

What we'd change

• Start the family communication two weeks earlier.
• Build the prompt-scaffolding guide before the pilot, not at day 40.
• Add a brief weekly check-in for students, not just teachers.

A reusable 90-day template

Days 1-15 — Foundation: working group formed, usage inventoried, tool selected, disclosure drafted.

Days 16-45 — Pilot: two classrooms, weekly debriefs, mid-pilot family survey, student focus group.

Days 46-75 — Iteration: scaffolding, disclosure rewrite, interim memo to families and faculty.

Days 76-90 — Decide: measure against pre-set criteria, expand or stop, publish the decision either way.

Either decision builds trust. The schools that lose family trust are not the ones that paused — they're the ones that never explained themselves.