AI is genuinely useful in school administration for pattern-finding across data a person cannot read in full — spotting a fee ledger that has drifted, drafting text a human then edits, and turning a question into a query. It is not suitable for judgements about individual children, because those get treated as facts by whoever reads them.
We build school software and have declined to ship several AI features other vendors offer. That is a position, and the reasoning is below so you can disagree with it.
Every education software vendor now has an AI page, and most of them describe the same three things: predicting which students will fail, predicting which families will not pay, and writing report card remarks. We have built none of them, and it is worth explaining why rather than quietly not mentioning it.
This is not scepticism about the technology. There is real, unglamorous value in applying it to school administration. It is a distinction about what kind of question is being asked, and about who reads the answer.
The useful line is not between prediction and description, or between statistics and machine learning. It is between an output a person will check and an output a person will act on directly.
A model that flags forty fee accounts as anomalous produces work: someone opens them and finds three genuine errors. Being wrong thirty-seven times costs nothing but time. The output is a queue.
A model that says a child is likely to fail produces a belief. It is read by a class teacher, mentioned to a parent, and remembered. Nobody checks it, because there is nothing to check against until the result exists — by which point the belief has already influenced how the child was taught.
All three of these are in production somewhere, and none is controversial.
The pattern is consistent: the output is either verified before it acts, or it is a starting point that a person completes.
Each is technically straightforward. That is rather the problem.
These get past the demonstration quickly.
Phrased the way they arrive, answered so each one stands on its own.
No. We use it where the output is verified before it acts or is a draft a person completes. We decline it where the output is a judgement about an individual child that will be read as a fact, which is a narrower objection than being against the technology.
No. School data is used to run that school’s instance and nothing else. It is not pooled, not used to train anything, and not shared.
Customers have asked. We have said no and explained why, which is a better use of the conversation than shipping something we would not want used on our own children.
Checked by the Product desk on 7 September 2026. Corrections go in the page, dated, rather than quietly.
Pricing is on the page, in the open, with the counting rule written next to it. Nobody has to book a call to find out what it costs.
See pricingForty minutes with an implementation engineer — your classes, your fee heads, your grading scheme, entered as we go. No slides, no sample school.