Choose “Save as PDF” as the destination. Turn off headers and footers for a clean copy.

Educave Research

Brief for Decision-Makers

Buying, governing and evaluating AI in schools

August 2026Educave research · Brief

The finding

The cost of building software has collapsed. Nothing in how schools buy, govern or evaluate it has adjusted. The deciding variable this decade is not capability but governance: who owns the data, who can audit the model, who holds the evidence, and who is able to say no.

The learning evidence points the same way. A 2026 rapid review of research on young people and generative AI concludes that outcomes are set less by whether AI is used than by how everything around its use is configured: the learner, the tool design, the task, the role of teachers and peers, and the assessment. Used well it can support reflection and revision. Used carelessly it invites answer-seeking, dependence and unchecked output.

Schools are not naive buyers. They are structurally disadvantaged. Procurement precedes evidence, and by the time evidence arrives the contract has renewed. If institutions respond to cheap software by buying and building more without a shared governance layer, fragmentation does not improve. It compounds.

The evidence base

Performance paradox: Generative AI often lifts immediate task performance and confidence while leaving durable understanding, transfer and independent capability unchanged. Affective gains are the weakest available proxy for learning.
Offloading: The distinction that matters is beneficial offloading, which removes peripheral effort, against detrimental and judgement offloading, which removes the thinking and checking that learning depends on.
Thin base: Most studies are short-term, small and conducted outside the setting reading them. Under 2% of assessed education interventions have strong or moderate evidence of effectiveness, and only 11% of staff request peer-reviewed evidence before adoption.
Exposure: 98% of UK higher education institutions identified a breach or attack last year, against 43% of businesses. Secondary schools rose from 60% to 73% in one year.
Return: No institution in our field research so far could evidence what its technology estate returns, or had a mechanism for measuring impact against internal objectives.

The tenancy test

If we closed our primary vendor account tomorrow, could we still access learner data, deliver curriculum, communicate with families, process admissions and run operations? If no, the institution is leasing decision rights rather than holding enterprise value.

Five questions any leader can ask without notice

Purpose: Which stated improvement goal does this serve, and how would we know it failed?
Cognitive work: What does this tool do for the learner that the learner needs to do for themselves?
Transparency: Can we trace the provenance of every figure it produces?
Access: Does this reach learners or staff by need or by cost?
Collective learning: What has this told us that a neighbouring school could act on, and have we published it?

The first three moves

Attach every subscription to a named improvement goal, with a review date before the renewal date.
Measure learning, not use. Ask for evidence of reasoning, judgement and transfer weeks later, not task completion on the day.
Write down where the human decision stays: which judgements are never automated, and who holds them.

Disclosed interest: Educave is an independent education research and technology company. educave.ai · Free to copy, forward and use inside your institution, with attribution.