Success Stories

The challenge

Teachers were using generative AI in private and asking students not to. Building principals were getting conflicting questions from parents about cheating, accommodations, and data privacy, and answering them inconsistently. The existing policy was vague enough that nobody trusted it to defend a decision, so most decisions were not being made.

The approach

We started with a policy rewrite, not a training rollout. A small working group of teachers, two principals, the curriculum director, and a parent representative spent eight weeks producing a usable AI guideline document with three tiers (permitted, permitted with disclosure, not permitted) tied to specific assignment types rather than tools. Only after that document was board-approved did we launch faculty training, in two cohorts: an early-adopter cohort of about sixty teachers who piloted classroom routines, then a general cohort the following semester. The early-adopter cohort doubled as the trainers for round two.

The outcome

The policy work was the most valuable part, and it took the longest. Once principals could point at a document and say what tier an assignment fell into, the parent conversations got shorter and calmer. Teacher confidence in talking with students about AI rose substantially in self-report; what we could not measure cleanly was actual classroom practice change, because most teachers did not want observation tied to AI use yet. About a quarter of the early-adopter cohort moved meaningfully; the rest experimented briefly and reverted. We were honest with the board that this was a culture-change project measured in years, not a tool deployment measured in semesters.

What they kept

• The three-tier assignment classification (permitted / permitted with disclosure / not permitted)
• A standing AI working group with rotating teacher seats
• Quarterly principal office hours specifically for AI questions
• Annual policy review tied to the district calendar, not to vendor news cycles

The challenge

The provost inherited five overlapping faculty committees on AI, each with a different recommendation and none with implementation authority. There was no shared picture of where the institution actually stood — which programs had quietly built things, which were paralyzed, which had budget, which had data infrastructure, and which had governance gaps. Without that picture, every prioritization conversation devolved into discipline-level lobbying.

The approach

We ran an institution-wide AI readiness assessment across six dimensions (governance, infrastructure, faculty capacity, student support, curriculum, ethics review), with structured interviews in every academic unit and parallel surveys for staff and a sample of students. The output was a heatmap, not a score — green where they could move now, yellow where they needed a precondition, red where they needed leadership decisions before any tool rollout would land. The provost used that heatmap to sequence a three-year roadmap with named owners.

The outcome

The heatmap did most of the political work. Once units could see themselves relative to peers, the lobbying shifted from 'fund my idea' to 'help us close our yellow.' Year one execution was solid on the green items — library guides, syllabus language templates, an opt-in faculty community of practice. The red items, particularly a real ethics review pathway for student-facing AI tools, moved slower than the roadmap promised because the governance committee meets monthly and decisions take quorum. We rewrote year two to be more conservative about governance timelines.

What they kept

• The six-dimension readiness frame, now run annually with a lighter instrument
• Named accountable owners on every roadmap item — no 'the committee' assignments
• Public-to-campus dashboards showing roadmap status
• Faculty community of practice with stipended facilitators

The challenge

Demand for new learning content was outpacing the team's capacity by a wide margin. Individual designers had started using generative AI on their own, producing wildly inconsistent outputs and reinventing prompts each time. Quality varied by who happened to be on the project, and there was no shared institutional memory of what had worked.

The approach

Rather than license a new platform, we worked with the team to build an internal AI Suite: a versioned library of role-tuned prompts and templates for the eight most common authoring tasks (learning objectives drafting, scenario writing, knowledge check generation, accessibility review, translation seed, debrief question banks, facilitator notes, summary generation). Each entry had a short voice-and-tone preface tied to the firm's house style, expected inputs, and known failure modes. Designers were trained on the suite, not on AI in general, and were expected to contribute improvements back.

The outcome

Authoring throughput for the covered task types roughly doubled within two quarters by team self-report, and quality drift across designers narrowed because the prompts were doing the standardization work that style guides had failed to do for years. The win was less about AI and more about the discipline of writing down good prompts as institutional artifacts. The accessibility review prompt was the biggest surprise — designers who had previously skipped accessibility passes because they were tedious now ran them consistently. What stalled: a planned translation workflow ran into legal review on data residency for a full year and never launched.

What they kept

• The versioned prompt and template library with owner and changelog per entry
• Quarterly suite review where designers nominate prompts to retire or upgrade
• Pairing new designers with the suite during onboarding before any tool training
• Failure mode notes documented next to each prompt

The challenge

Off-the-shelf AI accessibility features were arriving inside tools the district had already licensed — live captioning, text-to-speech, reading-level rewrites, image description. Teachers were unsure which features were classroom-ready for which student profiles, what the privacy posture was, and how to document use in IEPs without creating legal exposure.

The approach

We did not start with tools. We started with a feature-by-feature review against the department's existing accommodation language, mapping each AI feature to a specific accommodation type it could plausibly support, and flagging the ones where it could not. For the green-light features, we built short use protocols (one page each) describing setup, expected behavior, known failure modes, and IEP documentation language. Teachers piloted three features at a time for a full term before any expansion. The school psychologist and a parent advisory member sat in on every feature review.

The outcome

Live captioning and read-aloud became permanent in classrooms where they had been intermittent. Reading-level rewrites were quietly retired from the protocol after one term because they smoothed away vocabulary that the IEPs explicitly required students to encounter. Image description was useful for a smaller group than initially expected, mostly social studies materials, and the team kept it but narrowed the rollout. Honest finding: the biggest accessibility gains were not the flashy features but the simple captioning, which nobody had thought worth a project plan before.

What they kept

• One-page use protocols per AI feature, owned by the department
• IEP language templates reviewed by counsel before any AI feature is used in plan documentation
• Parent advisory sign-off on each feature before classroom pilot
• A retire list, kept current, so dropped features do not quietly return

The challenge

Hands-on skill assessment was the bottleneck. Each instructor could realistically evaluate a few students per day at the level of detail the certificate required, and intake had outgrown that capacity. Written tests were a poor proxy, and the program's employer partners cared specifically about practical proficiency.

The approach

We worked with three instructors and one curriculum lead to redesign the assessment workflow rather than replace it. Students recorded short structured demonstrations of named procedures on shared tablets. AI was used in two narrow places: transcription and indexing of the spoken procedure narration so instructors could jump to the moments they needed to evaluate, and a first-pass checklist marker that flagged when expected procedure steps appeared to be missing or out of order. Instructors then reviewed flagged segments at speed and made the final judgment. No grade was assigned by AI.

The outcome

Per-instructor assessment throughput roughly tripled for the procedures included in the pilot. The bigger surprise was that students prepared more carefully because they knew their demonstration was recorded; some used the recordings as portfolio pieces with employers. The checklist marker was wrong often enough — particularly with non-standard sequencing that experienced students used legitimately — that instructors learned to treat it as a 'look here' nudge, never a verdict. One instructor opted out entirely and we left her workflow alone; her cohorts performed comparably on employer-side outcomes, which was a useful reminder that the redesign was an option, not a mandate.

What they kept

• Recorded structured demonstrations as the primary practical assessment artifact
• Transcription and timestamped indexing for instructor review
• Final grade authority strictly with the human instructor
• Student access to their own recordings for portfolio use

The challenge

Coaches were being asked AI questions by the teachers they supported, and most felt one step behind. Off-the-shelf AI PD was either too vendor-specific or too abstract, and coaches needed something that worked for the specific job of coaching teachers — modeling, observation, feedback, planning conversations — not generic productivity.

The approach

A year-long cohort built around four full-day in-person sessions, monthly virtual practice clinics, and a shared problem bank. Each in-person day paired a concept (e.g., how a coaching conversation changes when both parties have used AI on a lesson plan) with a hands-on rehearsal in pairs and a debrief in fours. Between sessions, coaches brought real coaching situations to clinics, anonymized, and the group worked them. We deliberately did not certify anyone or build a credential — the purpose was practice depth, not badges.

The outcome

Self-reported confidence in handling AI-related coaching conversations rose substantially and held at end-of-year follow-up. The practice clinics did more of the work than the in-person days, which was a surprise — the in-person days mattered mainly for trust-building that made the clinics possible. Attrition was higher than expected (about a fifth dropped out by mid-year), almost entirely due to coaching role changes or district leadership turnover, not the content. The shared problem bank kept growing after the cohort ended, which we took as the strongest signal.

What they kept

• Monthly virtual practice clinics as a standing structure
• The anonymized shared problem bank as a living resource
• Pairs-then-fours rehearsal structure for any new coaching practice
• Explicit refusal to issue a credential or score participants

The challenge

The board was pushing for an AI strategy. The head of school suspected a strategy at their scale would mostly be theater and that what they actually needed was a few specific decisions made well. They wanted help avoiding the trap of importing a large-school playbook.

The approach

We scoped down hard. Three faculty meetings over a term: one to share a short literature read and align on principles, one to make three specific decisions (faculty AI use for planning is encouraged with light documentation; student AI use is grade-banded and assignment-specific; no student data goes into general-purpose chatbots), and one to commit to a six-month review. No tool selection. No vendor demos. The technologist was asked to maintain a single internal page summarizing the decisions and any updates.

The outcome

The three decisions held for the full school year with only one substantive amendment (clarifying the grade band on student use after a fifth-grade incident). Faculty reported that having three clear decisions was more useful than a longer policy would have been because they could actually remember and apply them. The head of school told us the most valuable part of the engagement was the permission to do less. We took that seriously.

What they kept

• Three named AI decisions on a single internal page
• Twice-yearly faculty meeting time reserved for AI decision review
• A standing rule that no new AI tool gets adopted institutionally without a written decision attached

The challenge

Faculty were divided. Some viewed generative AI as legitimate scholarly infrastructure (literature mapping, summarization, code assistance, translation) and were already using it openly. Others viewed it as corrosive to the formation of an independent scholar and were quietly failing students they suspected of using it. There was no shared norm, no disclosure standard, and no defensible appeals process. PhD candidates were anxious and asking different supervisors different questions and getting different answers.

The approach

We did not try to settle the substantive debate. We worked with the dean's office and an elected faculty group to produce a process: a per-program disclosure norm (each program defined its own AI use expectations and published them annually), a candidate-facing template for declaring AI assistance on any submitted work, a supervisor conversation template for the start of each academic year, and a referral path for cases where supervisor and candidate disagreed. None of these adjudicated whether AI use was right or wrong; they made the conversation legible.

The outcome

Within a year, every program had a published norm. The norms varied widely, which was the point — a quantitative methods program defined permissive use; a critical theory program defined restrictive use; both were now defensible because they were written and visible. Candidate complaints dropped because expectations became findable. What did not resolve, and we were honest that it would not: a handful of senior faculty refused to engage with the process and continued enforcing personal standards on their own candidates. The dean's office accepted that as a culture-of-disagreement cost and did not force compliance.

What they kept

• Per-program annual AI norm publication
• Candidate disclosure template tied to every submitted artifact
• Start-of-year supervisor conversation template
• A referral path when supervisor and candidate disagree on AI scope

The challenge

After two semesters of rising informal reports of AI use in submitted essays, the writing program had drifted into adversarial mode — proctored in-class writing, AI detectors with known false positive rates, escalating academic integrity caseloads. Faculty felt like fraud investigators. Students felt under suspicion. Pass rates were holding but engagement metrics and end-of-term evaluations were dropping.

The approach

The writing program coordinator led a redesign with full and adjunct faculty representation. The core move was to shift the assessed artifact from the final essay to the writing process: shorter staged submissions with required reflection, in-class drafting that captured working notes, and a final portfolio that included revision history and a self-analysis of where the student had used assistance (human, AI, or otherwise) and why. AI detectors were retired from the program's official toolkit. Faculty got a redesigned rubric and three half-day workshops on responding to the new artifacts.

The outcome

Academic integrity caseloads in the writing program dropped substantially in the first redesigned semester, primarily because the redesigned assignments made AI-only submissions visibly thin against the rubric. Student end-of-term comments shifted from defensive to substantive. Faculty load went up — responding to staged submissions and portfolios takes more time than grading final essays — and the program had to negotiate adjunct compensation adjustments, which they were partially able to do. The honest tradeoff: the redesign improved learning and morale and cost the institution real money.

What they kept

• Staged submissions with required reflection
• Final portfolio including revision history and a self-analysis of assistance used
• A program-level disclosure norm published in every syllabus
• Three annual faculty workshops on responding to portfolio artifacts

The challenge

Teacher candidates were graduating into K-12 classrooms where AI was already present, and the education school's program had not meaningfully changed. Methods courses still treated technology integration as an optional add-on chapter. Mentor teachers in placement schools were sending mixed signals. Newly graduated teachers were reporting back, sometimes within their first month, that they felt unprepared for the AI questions students and parents were raising.

The approach

Rather than adding an AI course, we worked with methods faculty to integrate AI into existing courses at three deliberate touchpoints: a foundations-of-the-profession discussion early in the program about AI's role in teaching and learning, embedded practice in the subject-specific methods sequence (an English methods course working differently from a math methods course), and a capstone assignment in the clinical year requiring candidates to document an AI-aware decision they made in placement and defend it. Mentor teachers received a short companion document, optional but supplied to every placement.

The outcome

Candidate self-reported preparedness for AI conversations rose meaningfully by the end of the clinical year. The methods integration worked best where the methods faculty member was personally invested; in two subject areas the integration was largely cosmetic, and we flagged that openly to the dean as something not to paper over. The capstone defense became a surprisingly strong artifact — many candidates wrote sharper, more situated arguments about AI use than their faculty had expected, partly because they had been in classrooms recently and faculty had not. The mentor teacher document was used by maybe a third of placements; we have not solved that yet.

What they kept

• AI integration into existing methods courses rather than a standalone AI course
• Capstone defense of an AI-aware decision made in placement
• Subject-specific approaches in subject-specific methods sequences
• An annual review of which methods courses are integrating substantively and which are not