Directors, Not Users

Current State

Three forces converging in Massachusetts

The PD gap

DESE defines HQPD as "systematic, purposeful, and structured over a sustained period of time." The dominant model — single-session workshops — meets none of those criteria. Coaching works (d=0.49) but can't scale. The approaches that scale don't work. TNTP 2015; GAO 2026; Kraft et al. 2018

AI guidance without implementation

DESE published comprehensive AI guidance and resources — a strong foundation. But guidance is not implementation. No diagnostic, no strategy, no configured assistant, no measurement. The state's largest direct investment — Future Ready — allocated $135K for a 50-hour professional learning experience in about 30 districts. AI tools alone produce d=0.06-0.14. AI supporting skilled professionals does considerably more. Oreopoulos & Low 2026; Wang et al. 2024

AI literacy becomes required

The Graduation Council has recommended AI and digital literacy as a required MyCAP milestone, starting with 9th graders in 2027-28 (advisory until the legislature acts). Teachers who integrate AI into their own instruction are the district's capacity for meeting this requirement. No separate curriculum needed — the capability transfers.

The Research

The evidence base

Kraft (2020) gives a realistic yardstick for causal studies of achievement: below 0.05 is small, 0.05 to 0.20 is medium, above 0.20 is large. The causal and randomized numbers below are the benchmarks this framework should be held to.

Finding	Effect	Source
Teacher quality: strongest in-school predictor of student outcomes	Established	Nye et al. 2004; Chetty et al. 2014
Coaching: effect on teaching practice	d = 0.49*	Kraft, Blazar & Hogan, 2018 (60 causal studies)
Coaching: effect on student achievement	d = 0.18*	Kraft, Blazar & Hogan, 2018
Curriculum-focused PD (STEM)	d = 0.21 avg	Lynch et al., 2019 (95 studies)
Formative assessment	d ≈ 0.20	Kingston & Nash, 2011
AI tutoring tools alone (2-year RCT)	d = 0.06-0.14	Oreopoulos & Low, 2026 (18 schools)
AI assistance for human tutors (RCT)	+4 to +9 pts mastery	Wang et al. (Tutor CoPilot), 2024
AI-assisted lesson planning: time savings (RCT)	~31% reduction	Education Endowment Foundation, 2024
Traditional workshop PD (dominant model)	Not reliably effective	TNTP 2015; GAO 2026
AI equity gap: lower-income schools, lower integration	0.07-0.11 SD gap	Campos & Singleton, 2026
Scaffolding / Structured discussion	d = 0.82**	Hattie synthesis
Feedback / Higher-order questioning	d = 0.73**	Redfield & Rousseau, 1981; Hattie

*Meta-analysis of 60 causal studies (Kraft, Blazar & Hogan, 2018). Kraft (2020) benchmarks for education RCTs: <0.05 small, 0.05-0.20 medium, >0.20 large. Coaching's 0.18 on achievement is near the top of what school-based programs produce.
**Hattie's meta-analytic methodology has known limitations (Snook et al., 2009). These pooled values are useful for comparing practices, not as predictions of what a program will produce.

The Synthesis

Four bodies of work, organized around student learning

Improvement Science

Bryk et al. (2015): "solutionitis" — jumping to solutions before understanding problems. The Diagnose-Respond-Measure-Adjust cycle applies disciplined Plan-Do-Study-Act methodology to AI-integrated instruction, adding peer comparison as the starting point and AI as a way to lower the cost of acting on what the data shows.

Evidence-Based Strategies

The strategies with the strongest evidence are well-documented: coaching, structured discussion, scaffolding, feedback, formative assessment. We have known what works for years. Where teachers already know a practice, the constraint is time — AI changes that math. Where they know it less well, the constraint is knowledge and skill, and that's what coaching is for.

Questioning + TPACK

The judgments involved in directing AI well — being clear about purpose, attending to students and level of challenge, and evaluating output against criteria — are the same judgments involved in designing good questions, tasks, and assessments (Rothstein & Santana, 2011; Mishra & Koehler, 2006). Willingham (2007) cautions that thinking skills transfer less readily than often assumed. The claim here is narrow and testable.

How Organizations Adopt AI

AI configured for specific work, loaded with relevant knowledge and built into existing workflows, produces larger gains than general-purpose tools — especially for less experienced workers (Brynjolfsson et al., 2025; Dell'Acqua et al., 2023). The planning assistant applies this pattern to instruction.

Core Concepts

Two complementary frameworks

Multiplying value, not adding it

Ask teachers what keeps them from teaching the way they know they should, and time comes up first; teachers report working about 53 hours a week (RAND, 2024). For practices teachers already know, the assistant takes on the preparation. For practices they know less well, coaching builds the skill. Together, the skills, judgment and relationships teachers already have can reach every class, every day.

Directors, not users

A user asks for a quiz and uses whatever comes back. A director starts from what her students need — the standard they haven't mastered, the misconception last week's exit ticket turned up, the level of thinking the unit calls for. She picks the instructional move that fits, has the assistant produce the material, and judges it the way she'd judge a colleague's work. The user gets a quiz. The director gets an assessment designed to tell her something. The difference is professional judgment, not prompting skill — the increasingly deliberate planning, monitoring, and evaluation that Zimmerman (2002) and Schraw (1998) call self-regulated learning. Nearly everyone starts out as a user of any new tool; the point is not where teachers begin but how far the work helps them go. The metacognitive skills that develop — clarity of purpose, managed complexity, critical evaluation, iterative refinement — transfer across professional practice, which is why this framework positions alongside TPACK (Mishra & Koehler, 2006) as complementary: TPACK describes the knowledge teachers need; self-regulated learning describes how they deploy it.

Theory of Change

The chain of claims

Each link is grounded in existing evidence. The chain itself — the specific combination — is the hypothesis being tested.

1

Start with a data-driven diagnostic that identifies specific gaps by grade, subject, and subgroup. Bryk: make the work problem-specific

2

Target the highest-leverage strategies for that district's specific needs, organized around the adopted curriculum. Lynch et al. (2019): curriculum-focused PD has larger effects; Kraft et al. (2018): coaching improves both practice and achievement

3

Equip teachers to direct AI toward those strategies using assistants configured for their curriculum and standards. Cross-sector evidence: configured AI with skilled users produces larger gains (Brynjolfsson et al., 2025; Wang et al., 2024)

4

Multiply existing expertise — strategies that have always been known to work become practical and sustainable because AI removes the preparation bottleneck. The multiplier thesis

5

Measure continuously, disaggregated by subgroup, and adjust quarterly. Bryk: we cannot improve at scale what we cannot measure

Intellectual Honesty

Where the evidence stands

This framework does not overstate its claims.

Established

Teachers are the most important in-school influence on student learning (Nye et al., 2004; Chetty et al., 2014)
Coaching improves both teaching practice and student achievement, with smaller effects in larger programs (Kraft et al., 2018)
Professional development built around curriculum materials is associated with larger effects than generic PD (Lynch et al., 2019)
Feedback, discussion, explicit instruction, metacognitive strategies and formative assessment have consistently positive evidence (Hattie, 2009; EEF; Kingston & Nash, 2011)
One-time workshop PD rarely changes practice (TNTP, 2015; GAO, 2026)
Most teachers have not received AI training from their school or district (Gallup-Walton, 2025)

Emerging (credible but early)

Student-facing AI tutoring produces small-to-medium gains, limited by engagement (Oreopoulos & Low, 2026)
AI that supports a human tutor improves student mastery, especially for less experienced tutors (Wang et al., 2024)
AI-assisted lesson preparation reduces planning time by roughly a third without lowering quality (EEF, 2024)
In other professions, AI assistance raises productivity most for less experienced workers (Brynjolfsson et al., 2025; Dell'Acqua et al., 2023)
Schools serving more disadvantaged students are integrating AI more slowly (Campos & Singleton, 2026)

Evidence that cuts the other way

Open AI access can raise immediate performance while weakening learning (Bastani et al., 2025, PNAS)
People who rely on capable automation tend to miss its errors (Parasuraman & Manzey, 2010; Dell'Acqua et al., 2023)
Teacher-sourced supplementary materials are often low in quality (Polikoff & Dean, 2019)
Broad syntheses like Hattie's overstate effects relative to causal studies (Kraft, 2020; Slavin, 2018)
The coaching programs behind the 0.18 estimate generally involved more contact than monthly team cycles

What this work will test

Teachers reinvest time saved through AI in high-impact instructional practice
The assistant's output is consistently accurate, aligned and rigorous
Team coaching supported by a prepared assistant approaches the effects of more intensive coaching
The full system improves achievement more than AI tools alone
Director judgment developed with the assistant grows, rather than erodes, and shows up in teachers' questioning and assessment design
A district can run the cycle independently after one year

Under ESSA's evidence framework, this approach meets Tier 4: Demonstrates a Rationale — a well-specified logic model informed by high-quality research, with a planned evaluation. That is the honest starting point. Moving to Tier 3 or 2 requires a comparison group, so the measurement design includes one from the outset.

The evaluation design will be pre-registered with the Registry of Efficacy and Effectiveness Studies (REES) before the first engagement begins. It uses staggered starts and comparisons with peer-district trends, corrects for regression to the mean, and measures time use and output quality directly instead of relying on self-report. Single-district results are reported descriptively; the confirmatory test pools districts. All results are published.