This post is the third post in an ongoing series on AI in education. Read more in the series here.
As of March 2026, more than 130 AI-in-education bills are under consideration across 31 states, and at least 28 states have published official AI guidance for schools. This is part of a broader pattern of states legislating on classroom technology: 33 states have now enacted full K-12 cellphone bans. Most AI-related policy activity is focused on which generative AI tools to allow, restrict or recommend. But the question that matters most doesn't have an answer yet: Which tools actually produce learning gains for kids? Absent clear evidence of learning, edtech risks becoming an inefficient use of student time and fueling the broader backlash against technology in schools.
At Accelerate, we are working to help build that answer. We are closing out a year-long research effort with 10 grantees deploying AI-powered learning tools in real K-12 classrooms. Nine of the 10 are running quasi-experimental studies with comparison groups, partnered with external researchers — the kind of rigorous evidence base that has been notably thin in this field.
We are still months from outcome results, which we expect to land later this year. But the early implementation data point somewhere consistent that is not dissimilar from our experience with high-dosage tutoring.
School Implementation Decisions Determines How AI Gets Used
Usage varies widely
Across our cohort, the gap between intended and realized usage is wide, even within the same tool — suggesting the tool itself isn't the main story. One literacy grantee saw 98.5% of eligible students register for a generative AI tool in a North Carolina district spanning urban and rural communities, but only 28% actually use the tool. The same tool in a large urban Georgia district hit 52% utilization. A grantee delivering AI-supported math tutoring in a charter network sees engaged days per week ranging from 0.7 to 5.6 across schools using the same product. A multimodal math tool deployed across 11 middle schools in a single Florida district meets its weekly target among students who use it — but only 56% of eligible students have logged a single session. Within a single study, the gap between the highest- and lowest-implementing schools is often bigger than the gap between studies.
The schools that hit their dosage targets share a recognizable profile. They built the tool into the regular school day. They had a site-level adult actively supporting participating teachers. They also planned for the predictable disruptions: weather days, testing windows, staffing gaps. None of this is novel. It is the same implementation infrastructure that has reliably distinguished high- from low-impact tutoring sites over the past four years.
Teachers shape how a tool gets used, not just whether
Teachers also exercise meaningful discretion over how a tool gets used once it is in their classroom. In one Arizona district, a math tool purchased explicitly for Tier 2 and Tier 3 intervention is now being used most heavily for Tier 1 instruction — against the district's original guidance. A grantee's implementation design is a hypothesis about how teachers will use their tool, but teachers ultimately decide.
What This May Mean for State Education Policy
If implementation conditions, not the AI tool itself, are emerging as the dominant variable in whether AI tools can potentially produce learning gains, several practical questions follow for state and district leaders.
- What implementation conditions make usage more likely? Early implementation data suggest that adoption improves when AI tools aren't treated as optional add-ons. States and districts may set clearer expectations about when the tool will be used, who within schools will support teachers, how usage will be monitored and how leaders will respond when engagement falls below expectations.
- How is usage being tracked, and against what? AI platforms generate detailed engagement data. But that data often sits disconnected from the student information systems where academic outcomes live. States that invest in interoperability standards (i.e., AI usage data can flow into the same systems districts already use to track student academic progress) give district leaders the ability to whether the tool is working, and for whom, without commissioning a study.
- Who is accountable when tools go unused? Districts may be paying for licenses to AI tools that teachers never assign. The wide variation in usage observed across our cohort suggests this is not hypothetical. Procurement frameworks may benefit from clearer expectations about who monitors usage, and what happens when adoption falls below meaningful thresholds. Outcomes-based contracting can help with this.
- What infrastructure makes evidence possible? This is the question that matters most, and the one states are best positioned to act on. The hardest blockers our research partners have hit this year had nothing to do with the tools — they were data-sharing agreements, IRB timelines and partner cooperation. The field needs evidence about which AI tools work for which students under which conditions. That evidence won't materialize until states invest in the conditions to produce it. State investment in streamlined data infrastructure, model data sharing agreements, and research-friendly procurement may do more to advance the field than any single tool selection.
A year of funding AI in real classrooms has surfaced one finding more consistently than any other: AI isn't replacing teachers, schools or the systems around them. Those are the things deciding whether AI earns a place in classrooms at all.
