Every conversation about AI in schools eventually splits into two camps: "AI is going to ruin education" and "AI will revolutionise everything." Both are wrong, and neither is helpful for actually building tools schools can use.
What AI should automate
Anything repetitive, mechanical, and rule-based. Attendance roll-call. Categorising support tickets. Generating standard receipts. Translating a parent's Hindi voice question into structured queries against your fee data. These are all friction points that drain teacher time without adding educational value.
What AI shouldn't touch
Anything that requires judgement, context, or care. Grading subjective answers. Coaching a student through a difficult patch. Discussing a behavioural concern with a parent. AI is bad at these, and even when it's "good enough" the trust loss isn't worth it.
The middle ground
Most useful AI features sit in the middle: AI drafts, humans approve. Examples on Schoolo:
- AI suggests a parent-teacher meeting summary, the teacher edits before sending
- AI flags students with attendance anomalies, the class teacher decides if it's a real concern
- AI drafts certificate text from a template, the principal signs off
The parent assistant question
Our AI parent assistant answers fee, attendance, and event queries in Hindi or English. It's allowed to do this because the answers are factual lookups against your school's own data — no judgement involved. The moment a question crosses into "is my child OK?" territory, the assistant routes to a human.
The right test: would you trust this AI answer at 9:47 PM when the office is closed? If yes, automate it. If no, keep the human in the loop.
An example of AI getting it wrong
Early in development, an attendance-anomaly model flagged a student as a "concern pattern" for repeated Monday absences. The actual reason: a standing Monday-morning physiotherapy appointment, already known to the school. The system had no way to know that context — which is exactly why it routes to a human teacher for a judgement call instead of auto-escalating to a parent notification or a formal record.
How we test before shipping an AI feature
Every AI feature goes through the same filter before release: can a wrong answer cause real harm if a busy teacher doesn't catch it? If yes, the feature ships in "suggest, don't act" mode by default, with a visible edit step before anything reaches a parent. We'd rather ship a feature that saves 80% of the time with a human check than one that saves 100% of the time and occasionally embarrasses the school.
Where we're deliberately slow
We get asked regularly for AI-generated report card comments and AI-driven grading. We've held off on both. A report card comment is one of the few times a year a teacher writes something personal and specific about a child — automating it away trades a small time saving for a real loss in what the comment is actually for.
For a concrete example, see how AI face attendance saves teachers hours without turning the classroom into a surveillance exercise. The same philosophy shapes how we handle behaviour tracking — coaching, not policing.