A teacher stops you in the hallway to show you something. She has been using an artificial intelligence tool to draft parent emails, and it is good. It is really good. She has cut 40 minutes off her Thursday nights. Two other teachers on her team have started using it. She wants to know if she can show the rest of the campus at the next faculty meeting.
You have about four seconds to answer, and every option is wrong. Say yes and you have approved a tool nobody has reviewed. Say no and you have told your most resourceful teacher that solving her own problem was a mistake. Say “let me look into it” and you have said nothing at all.
The problem is not that you lack a policy. The problem is that the decision arrived before the policy did, and it will keep doing that. Tools are free, adoption is individual, and by the time a request reaches your desk, it is usually describing something already in use.
What helps is not a longer approval process. It is a short set of questions you can run in four minutes, in a hallway, out loud.
Before You Continue
Of these five questions, which one does your campus currently answer worst?
A. What problem are we actually solving, and does it need AI?
B. Whose data goes in, and where does it go?
C. Who is accountable when the output is wrong?
D. What judgment or relationship do we lose by automating this?
E. How will we know it worked, and what would make us stop?
Hold your answer. We will come back to it at the end.
Why a Checklist Isn’t Enough on Its Own
You probably already know to ask most of these questions. The problem is not knowledge, it is timing. By the time these questions show up, a tool is already saving someone real time, and asking them feels like standing in the way of something that works. That is exactly when they matter most. These five are ordered and short enough to survive that moment, not just describe it.
The Five Questions
1. What problem are we actually solving, and does it need AI?
Most AI requests are not really about AI. They are about a workload nobody has fixed. The teacher drafting parent emails at 9 p.m. has a communication load problem. AI is one answer. A shared bank of message templates is another, and it is free, private, and does not hallucinate a student’s name.
Ask what the request would look like if the tool did not exist. Sometimes the honest answer is that you would have to add staff, and the AI is genuinely the better option. Sometimes the answer is that you would fix a process you have been avoiding for three years. Both are useful to know before you sign anything.
The failure mode here is adopting a tool and never naming the problem, which leaves you unable to tell later whether it worked.
2. Whose data goes in, and where does it go?
This is the question people think they have answered because they read the terms of service. Terms of service tell you what the vendor promises. They do not tell you what your staff typed.
The gap between those two things is where campuses get hurt. A tool can be fully compliant and still be a problem if teachers are pasting in behavior notes, health information, or names attached to grades. The vendor did not do that. Your workflow did.
So ask it in two parts. What does the vendor collect and retain, and separately, what will a tired adult on a Thursday night actually paste into the box? Design for the second answer.
One important reordering: when a tool is already in use, which is most of the time, this question comes first. You cannot evaluate a problem statement, a pilot design, or a success measure until you know what has already left the building.
3. Who is accountable when the output is wrong?
AI systems produce confident, fluent, well-formatted text that is sometimes incorrect. That is not a bug you can train out with a better prompt. It is a property of the technology, and it means every use case needs a named human between the output and the person affected by it.
The test is simple. Imagine the output is wrong and a family is upset. Who do they call, and what does that person say? If the answer is “the tool made an error,” you have not assigned accountability. You have distributed it, which is the same as removing it.
Be specific about where the human sits. Reviewing a draft before it sends is a real checkpoint. Being generally responsible for AI use on campus is not.
4. What judgment or relationship do we lose by automating this?
Every task you automate was doing something besides the task. Writing a behavior plan forces you to think about a student for 20 minutes. Drafting a difficult parent email makes you choose your words, which is often where you decide what you actually think. Research on professional judgment backs this up: the thinking happens in the doing, not before it, so skipping straight to a polished draft can quietly erode the skill you meant to support.
This is not an argument against automation, it is an argument for knowing what you are trading. Scheduling, formatting, and first-pass summaries carry no hidden value and should have been automated years ago. Others, like this one, carry the entire value in the doing. Ask which you are looking at, and keep a human step in the middle until you know.
5. How will we know it worked, and what would make us stop?
Campuses are good at pilots and bad at ending them. A tool gets adopted, the enthusiasm fades, the license renews anyway, and three years later nobody can say whether it helped.
Set the measure before you start, and make it small. Not “improve communication,” something like “teachers report saving at least an hour a week, and families report no drop in response quality.” Then set a review date and, more importantly, a stopping condition. What result would make you turn this off? If you cannot name one, you have not made a decision. You have made a purchase.
Try It
An assistant principal forwards you a free AI tool that drafts behavior intervention plans from teacher-entered notes. Three teachers are already using it. They say it saves an hour a week and the drafts are good.
What do you do first?
A. Approve it building-wide. Teachers are already getting value.
B. Ask the three teachers to stop until it clears review.
C. Ask what student information has already been typed into it.
D. Send it to your district’s technology office to check the terms of service.
Reveal the recommended first move and the tradeoffs
Recommended: C. The tool is already in use, which means the data question is no longer hypothetical. You cannot assess risk, notify anyone, or write a corrective step until you know what left the building. Everything else here is reversible. Disclosure is not.
B is defensible and many campuses would do it, but stopping use without knowing what was entered leaves you with an unmeasured exposure and three frustrated teachers who solved a real problem.
D is necessary but not first. Terms of service tell you what the vendor claims. They do not tell you what your staff typed.
A is the one to avoid. Existing use is evidence of a need, not evidence of a safe tool.
The pattern: when a tool is already in use, question two comes before question one.
Use It
Copy this into your next leadership meeting agenda. It is designed to be filled in on paper, not in a browser. Question two asks about student data, and a web form is the wrong place to type that.
Send It Up
A filled-in template only helps if it reaches someone who can act on it. Attach a recommendation, not just findings. “Here’s what I found” asks the reader to do more work. “Here’s what I recommend, approve, pilot, or decline” gives them a decision to react to, which moves faster in a full inbox.
If you do not have a standing contact for this, send it to whoever owns instructional technology purchasing for your district, by name if you know them, by role if you do not.
Check Your Thinking
1. A vendor’s terms of service say no student data is retained. Have you answered question two?
Check your answer
No. You have answered half of it. The vendor’s policy governs what happens to data after it arrives. Question two also asks what your staff will enter in the first place, which is a workflow question you control and the vendor does not.
2. Which question is hardest to answer after adoption rather than before?
Check your answer
Question five. Once a tool is in daily use, any measure you invent afterward tends to confirm the decision you already made. Naming the success measure and the stopping condition in advance is the only version of this question that produces a real answer.
3. Look back at the five questions you ranked at the start. Is your weakest one still the same?
For most campus leadership teams, the ranking shifts between question one and question three, from “we adopt things without a clear problem” to “we adopt things without a clear owner.” Both are worth fixing. The second one is more urgent, because it is the one families experience.
Next Step
Pick the tool your staff are already using most, the one that arrived without a decision. Run the five questions on it this week, fill out the template, and name a recommendation. Then send it to whoever owns technology purchasing for your district, before the next license renewal makes the decision for you by default.
