Schools Are Buying AI Faster Than They Can Write Rules For It

Last updated: September 19, 2026

Almost every article about AI in schools is an opinion. This one has a purchase ledger under it.

For the category with the sharpest consequences, see our companion piece on what districts pay for AI detectors and what the research says about them.

We traced what districts bought from generative-AI classroom vendors between FY2020 and FY2026. Superintendents were not asked. This is what their finance offices paid for.

What districts actually bought

We counted only AI-native classroom tools. MagicSchool AI, Brisk Teaching, SchoolAI, Diffit, Curipod, Khanmigo and Class Companion. Products that did not meaningfully exist before ChatGPT.

Civic IQ spend data, K-12 school agencies. The fiscal year runs July to June. Share is measured against districts reporting any spend that year, so it controls for our own coverage growing.

From 5 districts to 281 in three years. From under $50,000 to over $5 million.

Which vendors

California leads on district count at 63. New Jersey follows with 38, then Texas at 28 and Pennsylvania at 25.

What this looks like in one district at a time

The totals hide how ordinary each purchase is. These are individual records from the same data.

• West Allegheny School District in Pennsylvania approved MagicSchool AI for the 2026-27 year, capped at $23,000.
• Gurnee SD 56 in Illinois paid $3,960 for a Diffit site license.
• Penns Grove-Carneys Point in New Jersey approved Diffit as a supplemental program, not to exceed $4,950, on Title I funds.
• David City Public Schools in Nebraska set aside $3,400 for Gaggle monitoring in 2026-2027.
• Chester County IU 24 bought Curipod through a statewide consortium deal rather than its own bid.

Every one sits below the dollar line that would trigger a formal review. That is the story in five records.

The number that explains everything

The finding that matters most is not the growth rate.

The median district spent $7,999 on AI tools in FY2026. The median purchase order was $4,399.

Of 281 districts buying, 171 spent under $10,000. 230 spent under $25,000. Of 574 line items, 459 came in under $10,000.

Most bid thresholds sit between $25,000 and $50,000. Below that line a purchase needs no formal bid. It needs no board vote. No vendor security questionnaire. No privacy review.

It is a departmental budget line or a card swipe.

So the answer is not that administrators are reckless. These tools are priced below the line that would force a district to have a rule. A $4,399 site license routes around the whole approval process. That process was built for $400,000 purchases.

The buying system is working as designed. It was built for a world where serious technology was expensive. That is a structural problem. Telling districts to try harder will not fix it.

The rules gap, in survey data

Our spend data shows the buying. Independent surveys show the governance lagging, and the gap is wide whoever is asking.

• About 6 in 10 US public K-12 teachers use AI for their work, but only 18% have had formal guidance from administrators. 34% have had none at all. (Gallup with the Walton Family Foundation, fielded February 9 to March 2, 2026, 2,069 teachers, probability-based panel)
• 88% of districts have AI work underway, but only 56% have an acceptable use policy and just 34% have a data-privacy policy covering AI. (CoSN State of EdTech 2026, 607 districts)
• Only 13% of educators say their district has an AI policy clearly communicated to students and teachers. (EdWeek Research Center)
• 35 states and Puerto Rico have issued K-12 AI guidance. Only six require districts to adopt a policy: Idaho, Maryland, Oklahoma, Virginia, Ohio and Tennessee. (AI for Education tracker)

Guidance is not policy. Policy is not an enforced buying control. Most states have issued the first. Few districts have the third.

The evidence gap is worse than the policy gap

Districts are buying ahead of their rules. They are buying much further ahead of the evidence.

A 2026 Stanford SCALE review screened more than 800 papers on AI in K-12. It found only 20 with strong causal evidence. Of the student-facing ones, the authors are blunt: none ran in US K-12 schools. (Stanford SCALE)

The Institute of Education Sciences reached the same conclusion in August 2026. Strong causal studies remain, in its words, extremely limited.

The best US K-12 trial we found ran Khanmigo across 18 Tennessee middle schools for two years. It found roughly 0.04 to 0.08 standard deviations a year. The authors note the gains look like Khan Academy practice without AI. Nothing in the data, they say, requires the AI tutor to explain it.

A pre-planned trial with about 1,000 high schoolers found students given unrestricted GPT access performed 17% worse on a later unaided exam than students who never had it.

Two cautions when a vendor shows you a study

First, the widely cited meta-analysis reporting a 0.867 effect size was retracted in April 2026, and it is still circulating in sales decks.

Second, the most impressive studies use the wrong population. The Harvard physics AI tutor result ran on undergraduates. The MIT study everyone quotes about cognitive decline used adults aged 18 to 39. Neither tells you what happens in a 7th-grade classroom.

What goes wrong when you buy before you write rules

This is not hypothetical. The bill has already come due in several districts.

• The FTC took action against Evolv Technologies in November 2024 over its AI weapons screening claims. In Utica, New York, the scanners missed a seven-inch knife later used to stab a student. Raising sensitivity produced a 50% false-alarm rate.
• A Baltimore County student was handcuffed at gunpoint after an AI system flagged his crisps packet as a possible firearm. Security had already reviewed and cancelled the alert. The principal, unaware, called police anyway.
• An Associated Press investigation found monitoring false alarms leading to student arrests. In Lawrence, Kansas, nearly two thirds of 1,200 alerts over ten months were non-issues. More than 200 were flagged homework. In Polk County, Florida, about 500 alerts over four years produced 72 involuntary hospital stays.
• AI writing detectors falsely flag non-native English speakers at a 61% rate, per Stanford researchers who strongly caution against using them in educational settings.

Every one of those was a buying decision that preceded a policy decision.

Our own data points the same way. Monitoring and safety vendors such as Gaggle, Securly, GoGuardian, Lightspeed and Bark reached 798 districts and $16.7 million in FY2025 in our sample. That is far more districts than the classroom tools that get the press. The surveillance layer scaled first, and scaled quieter.

The backlash has started

In September 2026 New York City adopted a one-year moratorium on student-facing generative AI from pre-K through 8th grade. It covers roughly 600,000 students, about two thirds of enrollment. (Chalkbeat)

The same policy limits device time by grade band. It also requires every high school student to complete two AI-literacy modules. Days later Los Angeles Unified restricted student access for the 2026-27 year, reversing its own earlier position.

This is what happens when adoption outruns governance. You do not get careful rules. You get a moratorium. A blanket ban is the only tool left to a district that never built a graduated one.

So should AI be used in schools?

On current evidence: some of it, for some purposes, under conditions most districts have not established yet.

The strongest case is teacher-facing tools that cut admin load. Lesson drafting, levelled materials, translation, first-pass feedback. That is where the buying is clustered and where risk to students is lowest. The benefit also does not rest on unproven learning claims. A tool that saves a teacher four hours a week is worth buying even if it never moves a test score.

The case is weaker for student-facing instruction, where US K-12 causal evidence is effectively absent and one rigorous trial found harm to unaided performance.

It is weakest for high-stakes automated decisions. Discipline, threat detection, cheating accusations, special-education planning. Those carry the documented failure record. They also fall hardest on the students least able to push back. Oklahoma has already barred AI as the main basis for grading or discipline.

Arguing about AI in schools in the abstract is the wrong exercise. The useful question is which uses, bought under what controls.

What a district should actually do

The binding constraint is procurement rather than pedagogy. So the fixes are procurement fixes.

1. Set an AI review trigger with no dollar threshold. Any tool that touches student data or produces student-facing output gets reviewed at any price. This is the highest-leverage change. It addresses the $4,399 problem head on.
2. Write the acceptable use policy before the next renewal. The renewal cycle is a natural forcing function, and 56% of districts already have something to build on.
3. Require a data-processing agreement covering training use. Only 34% of districts have a privacy policy covering AI. The AFT and Microsoft National AI Safety and Privacy Standard offers contract language districts can adopt directly.
4. Separate teacher-facing from student-facing approvals. They carry different risk and should not share a review path.
5. Ban AI as the sole basis for a high-stakes decision. Keep a human who can override, and who is accountable for the outcome.
6. Ask vendors which population their evidence came from. If the study ran on undergraduates or adults or in another country, it is not evidence about your middle school.

Districts are already buying the governance they skipped. Several open bids below are for AI use and governance policy work. That is the market correcting itself one contract at a time.

Frequently asked questions about AI in schools

Are schools buying AI tools?

Yes, and fast. Districts buying AI-native classroom tools went from 5 in FY2023 to 281 in FY2026. Spend went from under $50,000 to over $5 million over the same period.

Why do schools buy AI tools without a policy?

Price. The median district spent $7,999 and the median purchase order was $4,399. Most bid thresholds sit between $25,000 and $50,000, so these buys never trigger a bid, a board vote or a privacy review.

What AI tools are schools buying?

MagicSchool AI leads by a wide margin. Brisk Teaching, Diffit, SchoolAI, Curipod and Khanmigo follow. Nearly all are teacher-facing tools for lesson drafting, levelled materials and first-pass feedback.

Do AI tutors improve student test scores?

The evidence is thin. A Stanford review screened more than 800 papers and found only 20 with strong causal evidence. None of the student-facing ones ran in US K-12 schools.

How many districts have an AI policy?

In CoSN’s 2026 survey, 88% of districts have AI work underway. Only 56% have an acceptable use policy and 34% have a data-privacy policy covering AI.

Which states require districts to adopt an AI policy?

Six: Idaho, Maryland, Oklahoma, Virginia, Ohio and Tennessee. Thirty-five states and Puerto Rico have issued guidance, which is not the same thing.

Should schools use AI?

Some of it, for some purposes. The case is strongest for teacher-facing tools that cut admin load. It is weaker for student-facing instruction, and weakest for automated decisions like discipline or threat detection.

Methodology

District purchasing figures come from the Civic IQ spend database: line-item purchase order and checkbook records published by agencies, cleaned and vendor-matched.

This analysis covers agencies typed as School, with purchase dates between July 1, 2019 and June 30, 2026. Line items are over $0 and at or under $2,000,000, which excludes parsing outliers.

Vendor matching uses anchored patterns on cleaned vendor names. We removed false positives by hand. One was a school-supplies vendor named School Aids Inc., which matched a naive SchoolAI pattern.

Three limits are worth stating plainly.

• This is a sample rather than a census. 3,440 districts appear in the spend data, out of roughly 19,000 nationally. It skews toward states with strong transparency-portal rules.
• FY2026 records are still arriving, so FY2026 totals are a floor. That is why we report share of reporting districts alongside raw counts.
• Spend data shows purchasing rather than policy. The governance figures come from outside surveys. We are not claiming to observe which of these districts had a policy.

Related reading: our analysis of the largest K-12 edtech vendors by spend, our look at the ESSER budget cliff that reshaped district technology budgets, and our guide to selling into school districts.

Civic IQ tracks procurement signals across 100,000+ agencies, $14T+ in tracked spend, 1.5M+ documents per month, all 50 states. The analysis above uses the same spend data our customers use to find where agencies are already spending.