Articles · AI in education · District leadership · Equity

Before you buy AI: the questions every district should ask

A practical checklist for district leaders: the BUILT Questions, three pass-fail gates, and the contract terms to require before AI tools reach students.

By Dr. William Hickman · 6 min read

The vendor’s best slide usually says something like “99 percent accurate.” It is a real number. It answers a question nobody at your table asked. Accurate for whom? Tested on which students? And what happens to the child in the other 1 percent?

I have sat on both sides of that table, as a superintendent and as a builder of AI tools for schools. Here is what I have come to believe: in a school system, buying an AI tool is a civil rights decision, whether you treat it as one or not. Public schools are bound by federal protections against discrimination based on race, color, national origin, and disability, and those duties do not disappear when a decision is made by software.

That does not mean stop buying. It means asking better questions than the sales deck answers.

Govern before you buy

Before any demo, settle who decides. At minimum, that means three things.

An oversight group that includes instruction, technology, data privacy, special education, English learner services, and a family representative, with authority to approve tools and a standing date on the calendar.

An approval workflow that covers classroom-level tools as well as district platforms, because that is where most AI actually arrives: a teacher finds a free grading assistant, a department switches on a new feature. Keep the small-scale workflow light. If it is so burdensome that teachers hide their tool use, you have governance that governs nothing.

An incident protocol that says, in writing, who is notified and how fast when a system fails, data is exposed, or a pattern of recommendations turns out to be unfair.

The BUILT Questions

In Built for Someone Else I describe how every technology imagines a default user, and how AI fails most often on the people outside that picture. To make that idea usable in a vendor meeting, I break it into five questions.

  • B, Builders. Who designed it, and who was missing from the room? Were teachers, special educators, and people who have taught English learners involved?
  • U, Users. Who was the imagined student? Usually it is a motivated kid with a quiet room, a working laptop, strong English reading, and no disability. Make the vendor describe what the tool does with every student outside that picture.
  • I, Inputs. Whose lives are in the training and test data, and whose are absent or distorted?
  • L, Limits. Where does it fail, and on whom was it tested? Ask for error and flag rates by race, English learner status, and disability status.
  • T, Trace. When it gets a student wrong, can the student and family see the decision, challenge it, and have a person correct it?

Get written answers and score each one green, yellow, or red. A vendor who cannot tell you how a product performs for different groups does not know how it performs. That is not a trade secret. It is a gap in their testing, and you are about to inherit it.

If answers are missing, require a local pilot, reported by group, before full rollout.

Why insist? Because the failures are rarely random. In 2023, a Stanford team led by Weixin Liang and James Zou ran essays written by non-native English speakers through seven popular AI detectors. On average, the detectors labeled those essays AI-generated more than 61 percent of the time, while doing nearly perfectly on essays by American eighth graders. Clear, simple writing looked like a machine. The students working hardest to belong paid the price.

Watch the demo, not the deck

My procurement rubric for schools starts with three gates. Each is pass or fail, scored on a live student account with the slides closed. A product that fails any gate is not scored at all.

  1. Teacher in the loop. Can a teacher assign, see, pause, and redirect the work? Is the product designed around the teacher or around her absence?
  2. A guardrail against doing the thinking. When the student account asks for the answer, does the tool ask a question back, give a hint, or require an attempt first? Can the student switch that guardrail off?
  3. The Monday-morning dashboard. Does it name the children who are stuck, on what, and what they tried? Or does it report minutes, streaks, and percent complete?

That third gate has a twist worth knowing. A 2026 CEPR discussion paper that followed 26,811 secondary students over thirty months found that about 80 percent of AI-related learning losses came from students who finished assignments unusually fast with high homework scores. On a typical dashboard, those are the students listed at the top, in green. Ask how the product flags them.

A product that passes all three gates is then scored 0, 1, or 2 on each letter of my MAKER test. Does it Make something new, Adapt to the child, Keep them thinking, End with something to show, and Reach others? Eight to ten points: pilot it. Five to seven: pilot only, with the missing pieces written into the contract as conditions. Four or fewer: decline.

Put it in the contract

Promises in a sales meeting are worth exactly what the contract says. Before you sign, require:

  • A data processing agreement that satisfies FERPA and your state’s student privacy law.
  • Plain answers on what student data is collected, where it is stored, who can see it, whether it is used to train the vendor’s models, and when it is deleted.
  • Independent accessibility verification against WCAG 2.1 Level AA, checked by your own staff with a screen reader and a keyboard, rather than a self-reported badge.
  • Documented bias testing, plus results by student group reported to you at least once a year.
  • Accommodations honored by default. A student whose plan allows movement breaks should never be flagged for moving.
  • No discipline, no accusation of cheating, and no change in placement based on an AI flag alone.
  • An appeals path that gets a family a human answer within a set number of days.
  • The right to leave, and to get your data back, if the tool does not perform fairly.

Build before you buy

One more rule, and it is the cheapest one on this page. Nothing is purchased for students until at least one teacher who will use it has built something with it, on paid time, and brought what she built to the leadership meeting. A teacher who has been confused by a tool and found her way out can recognize a child stuck in the same place. A vendor webinar after the contract is signed cannot do that.

When the board asks why this is taking a month, the answer is simple: thirty days of due diligence on a three-year contract is not obstruction. It is the minimum.

What to do this week

  1. Ask for the list. Have your technology department list every AI-powered tool in use, including free tools teachers adopted on their own. Most leaders are surprised.
  2. Change who reviews. Starting now, any tool that screens, flags, scores, or monitors students goes to your special education, English learner, and civil rights leads before it goes to the board.
  3. Print one page. Put the five BUILT Questions and the three gates on a single sheet and bring it to your next demo.
  4. Ask one family. Before your next major technology decision, walk a parent of a student with a disability and a parent who speaks a language other than English through the tool. Then listen.
  5. Rescore one renewal. Run your largest AI renewal through the gates with students’ actual work in hand, and ask the teacher who uses it most whether she would buy it again today.

Dr. William Hickman (Dr. Philip William Van Hickman) is an inventor, former superintendent, and the founder of MindAble Solutions. He is the author of eleven books, including The Intelligent Classroom. About him · Official site