The AI school needs a learning budget

An AI tool can help a child finish more work while leaving that child less able to do the next problem alone. For a school owner, that creates a measurement problem before it creates a technology opportunity.

A 2025 study of nearly a thousand high-school mathematics students found that a general-purpose AI interface improved practice grades by 48%. Once access was removed, that group’s grades were 17% lower than those of the control group. A tutoring version with learning safeguards largely mitigated the negative effect. These are relative changes in grades, not percentage-point changes in a pass rate. ^[1]

The commercial temptation is easy to understand. More completed work is visible. Independent understanding takes more effort to establish.

The design around the tool matters

A separate programme in Edo State, Nigeria, offers an encouraging result. Over six weeks, students used AI in structured after-school sessions with teacher guidance. The researchers reported an English learning gain of 0.24 standard deviations. They explicitly attribute the result to the full programme, not to the chatbot in isolation. ^[2]

That distinction is central to the business case. The school buys or builds a delivery system: devices, access, lesson design, teacher time, supervision, assessment, and the tool itself. A licence price captures only one component.

The two studies cannot be ranked as if they were competing products tested in the same classroom. They differ in subject, setting, design and outcome measure. Together, they give us a better question to test in India: under which conditions does a specific programme improve independent learning at an acceptable total cost?

Put a cost against the claim

Consider an illustrative pilot with 200 students. Assume an all-in incremental cost of ₹500 per student per month for three months. The budget is ₹3 lakh. That assumption needs a real quotation and a staffing plan before it becomes an investment proposal.

Suppose a matched comparison suggests that 20 additional students reach a predefined proficiency threshold because of the programme. The incremental cost per additional student reaching that threshold would be ₹15,000. If only five additional students do so, the figure becomes ₹60,000.

The exercise does not establish what an AI pilot should cost. It shows how strongly the conclusion depends on the outcome definition and the comparison. Without a credible comparison group or other evaluation design, we cannot confidently attribute the additional proficiency to the programme.

The same discipline applies to teacher time. A tool that saves lesson-planning minutes might free capacity for feedback. Whether it does depends on how the school uses those minutes. Time saved is an operational measure; better learning is a different outcome. We should record both.

What an Indian pilot should establish

Start with a specific bottleneck: reading comprehension, mathematical reasoning, feedback turnaround, or teacher preparation. Define the baseline and the outcome before choosing the tool. Assess pupils without the AI when the claim concerns independent skill. Compare total delivery cost with a credible alternative, including additional human tutoring.

Then look at the distribution. An average gain can conceal pupils who benefit substantially and pupils who become dependent on answers. Language, age, device access, and prior attainment belong in the interpretation.

My working hypothesis is that the stronger school proposition will be the ability to show when AI helps, when it does not, and what the improvement costs. That gives parents an intelligible result and gives management a decision it can revisit. It is a more demanding standard than reporting logins or completed worksheets.