The AI roadmap that ends at rollout

Most district AI roadmaps have the same shape. Form a committee. Write a policy. Vet tools for privacy. Buy licenses. Train staff. Report usage to the board.

Every step on that list is reasonable, and all of them are about the tool. The roadmap ends at rollout. In the implementation stages Dean Fixsen’s team at NIRN describes, that’s Installation. Initial and Full Implementation, when the new practice becomes routine in classrooms, never make it onto the page.

Why tool-anchored plans stay in flux

A roadmap anchored to a tool inherits the tool’s lifespan. And in this market, that lifespan is short. The platform you vetted in spring has a new model by fall. The feature you trained on moves behind a different tier. A new product shows up that the committee hasn’t reviewed.

Each change reopens the plan. Re-vet. Retrain. Rewrite the guidance. Staff experience that as churn, and churn is the enemy of behavior change. People don’t change how they teach while the ground is still moving. They wait.

It’s the pattern I see in the AI workshops I’ve run since 2022: the plan keeps everyone busy and no one committed. The dashboard shows activity. The classroom looks the same.

A roadmap should be anchored to what doesn’t change every quarter. What you need students to be able to do mostly doesn’t.

Measure the evidence, not the login

Usage data answers “are people using it?” It doesn’t answer “did the work students do get better?”

In 2019, when I was director of STEAM, my district adopted a new math program. We ran walkthroughs to see what it looked like in classrooms. Every classroom gave the program its allotted time. What happened in that time varied widely. Some classrooms skipped the manipulatives, some skipped sections, and some rewrote the formative assessments the first time they used them. The transition to a new program with all new modes of teaching math had veteran teachers feeling frustrated. The teachers rewriting the assessments weren’t going rogue because they didn’t care about student outcomes. They were falling back on comfortable routines they had developed over years of teaching math without discussions. By the usage numbers, the adoption was complete. In classrooms, it looked like 10 different programs. The district focused our professional development on helping teachers better understand why we were shifting math instruction. Teachers had chosen this program and they wanted the outcomes, but changing their behaviors was not as simple as handing out a new text.

Hall and Hord’s Concerns-Based Adoption Model has a name for what we saw. Its Levels of Use separate teachers following the steps (Mechanical use) from those using a program as a matter of course (Routine) and those adapting it for their students (Refinement). Logged time can’t tell those apart. AI plans are repeating the mistake: counting time with the tool instead of looking at the work.

Here’s a measure that does. Each semester, pull ten assessments that feed a grade or a placement, and ask one question of each: would this still show us the student’s thinking if a tool produced the artifact? Track the share that would. That’s the number for your change-management dashboard, and for your board: “4 of 10 assessments hold, up from 2 last semester.”

Ruben Puentedura’s SAMR model is a heuristic, not a validated scale, but it’s a useful second lens on the same sample. It sorts tasks by how much the technology changed them:

• Substitution or Augmentation: the task is the same, maybe with a functional improvement.
• Modification: the task was significantly redesigned.
• Redefinition: students are doing something that wasn’t possible before.

Read the two together. A task can be redesigned and still not show the student’s thinking, so the evidence share is the number that matters. SAMR tells you whether redesign is happening at all. If tasks are moving and the evidence share is rising, teachers are redesigning the evidence, whichever tool is in use.

Neither measure is precise. Ten tasks is a small sample and the judgment is rough. But it’s a rough measure of the right thing, which certainly beats a precise measure of the wrong one. And the sample describes the system, not individual teachers.

That measure can live beyond a vendor switch. Logins don’t.

Build the roadmap backward

Start with a decision the district makes on student evidence: a course grade, a placement, a graduation requirement. Then work backward.

1. Which evidence does that decision rest on? Name the assessments.
2. Does that evidence still hold? Could the tool produce it without the student’s thinking?
3. What would have to change about the task? That’s your Modification target.
4. What do teachers need to make that change? That’s your training plan, and the tool is one line item in it.

Here’s one run through. A 10th-grade lab report feeds the semester science grade. Does it hold? Not as written: a tool can write the analysis section from the data table. The Modification: keep the report, and add a five-minute conference where the student explains one anomaly in their own data. Now the grade rests on something the tool can’t supply, and the training plan is how to run those conferences in a class of 28.

Tools still show up in this roadmap. They show up at step four, in service of a task, instead of at step one in search of a use.

Why staff wait, and what ends the wait

Behavior change in schools is slow for good reasons. Teachers have seen initiatives come and go, and waiting one out is a rational response. A roadmap earns their commitment when it’s clearly going to outlast the current product cycle.

CBAM’s Stages of Concern explain the wait. Each new tool resets staff to self concerns: what is this, and what does it mean for me? Consequence concerns, whether it’s helping students, only come once those settle. A plan anchored to the tool keeps resetting the clock. A plan anchored to student evidence starts at the consequence question and stays there.

Anchoring to evidence tells staff the plan is about what students should be able to do, and that conversation was going to have to happen anyway: what counts as proof that a student learned it.

Where to start

Pull your current AI plan and highlight every step that names a tool, a license, or a training on a tool. Then highlight every step that names a student task or a piece of evidence. If the second color barely shows up, that’s the gap. Pick one decision the district makes on student work and answer the four questions above for it. That’s the first page of a roadmap that doesn’t end at rollout.