AI for EdTech Companies
You are building the systems schools deploy, which means their regulatory obligations arrive on your roadmap as feature requests.
Edtech companies sit upstream of every school and university decision about AI, which means the classification question lands on the product roadmap before it lands in a customer's governance meeting.
AI for edtech covers adaptive learning and personalisation models, assessment and feedback systems, content generation constrained to curriculum, learner analytics, multi-tenant model architecture, and the provider obligations and children's data governance that education deployments require.
You may be the provider of a high-risk system
Annex III makes several education applications high-risk, and the obligations on a provider are substantially heavier than those on a deployer. Edtech companies frequently assume the school carries the burden.
- Determining admission or assignment to an institution is high-risk. If your product influences placement, it is in scope.
- Evaluating learning outcomes is high-risk. Which reaches any system producing an assessment that goes on a record.
- Monitoring prohibited behaviour during tests is high-risk. Proctoring products sit squarely here.
- Emotion inference in education is prohibited outright. Not high-risk — banned, since February 2025, and engagement detection from faces is the common form.
- Provider obligations include the technical documentation, risk management and conformity work your customers cannot do on your behalf and will ask you for.
Your buyers' procurement will enforce this before any regulator does
A public school or university carrying Article 27 obligations cannot complete a fundamental rights impact assessment without information only you hold: what the system does, who it affects, what harms are foreseeable, what the oversight arrangements are and what evidence of performance exists. Vendors who ship that documentation are winning procurement processes against competitors still answering with a marketing deck. It is a sales asset before it is a compliance cost, and it takes far less time to produce than the deals it unlocks.
Children's data, which is a different regime
School consent is not family consent
A school agreeing to a contract is not the same as a family agreeing to their child's data being processed, and in several jurisdictions the distinction carries specific requirements. Rollouts have ended over this.
Profiling of children is restricted
Many jurisdictions have codes limiting profiling and behavioural targeting of children specifically, with age-appropriate design requirements that reach defaults, nudges and engagement mechanics.
Retention needs a defensible answer
Learner interaction data is valuable for model improvement and difficult to justify retaining indefinitely about a child. Decide the position deliberately and build to it rather than discovering it in a data protection review.
Training on learner data needs saying out loud
Whether you may train on a school's learner data is a contract question with a children's data dimension on top, and increasingly the answer is no by default. Design the architecture for that answer.
Model quality, and the evidence buyers now ask for
| Area | Common gap | What good looks like |
|---|---|---|
| Evaluation | Aggregate accuracy only | Performance by learner segment, with weak segments documented |
| Fairness | Absent or asserted | Outcome reporting across groups, shipped as a product capability |
| Efficacy evidence | Testimonials | Measured learning outcomes against a comparison, however modest |
| Multi-tenant models | One pooled model | Shared base with institution-level adaptation where cohorts differ |
| Content generation | Open-ended | Constrained to curriculum and reviewed; errors reach children directly |
| Age-appropriate design | Retrofitted | Defaults, retention and engagement mechanics designed for children from the start |
Efficacy evidence is becoming the procurement differentiator
Education buyers have been sold a great deal of technology on testimonials and are increasingly asking what evidence exists that it improves learning. A modest, honest evaluation — a comparison cohort, a defined outcome, and a straightforward account of what was and was not demonstrated — outperforms an impressive claim with nothing behind it, particularly with institutional buyers who have been disappointed before. Building the measurement into the product makes this answerable continuously rather than as a one-off study.
How an engagement runs
Classification and children's data position first, because both are architectural.
Product and classification review
Whether your product falls in Annex III, and what provider obligations follow.
Children's data architecture
Consent model, retention position, profiling constraints and training permissions.
Evaluation and model work
Segment-level evaluation, fairness reporting, and adaptation where cohorts genuinely differ.
Evidence and documentation
The pack your buyers need for their own assessments, plus efficacy measurement built in.
Operation
Drift monitoring, scheduled retraining, and documentation kept current as the product changes.
What you receive
A product that survives your buyers' procurement and your regulators' reading.
Classification position
Whether and where your product is high-risk, documented with the provider obligations that follow.
Children's data architecture
Consent model, retention, profiling constraints and training permissions built in.
Segment-level evaluation
Performance by learner segment, with weak segments documented rather than averaged away.
Fairness reporting
Outcome reporting across groups, shipped so buyers can run it on their own deployment.
Buyer documentation pack
What a school or university needs to complete its own impact assessment.
Efficacy measurement
Built into the product, so the learning outcome question is answerable continuously.
Is this the right starting point?
Worth being direct. There are situations in edtech where custom AI work is the wrong spend, and those are listed rather than buried.
Worth doing if
- Your product may fall within Annex III and nobody has established the position.
- Buyers are asking for compliance evidence you cannot currently produce.
- Evaluation is aggregate accuracy and nobody knows which learners the model serves badly.
- Children's data retention and training permissions were never designed deliberately.
- You are asked for efficacy evidence and have testimonials.
Do something else if
- You want emotion or engagement inference from students.
- The product needs distribution and customers rather than better models.
- Contracts do not permit the data use the intended architecture requires.
- There is no willingness to document where the model performs poorly.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
Are we the provider of a high-risk system?
Quite possibly, and it is worth establishing rather than assuming your customers carry it. Annex III makes systems determining admission or assignment to institutions, evaluating learning outcomes, and monitoring prohibited behaviour during tests high-risk — which covers a large share of edtech products. Provider obligations are substantially heavier than deployer obligations and include technical documentation, risk management and conformity work your customers cannot do for you. Separately, emotion inference in education is prohibited outright rather than regulated, so engagement detection from faces is not a compliance project.
What documentation do our buyers actually need?
Enough for them to complete their own assessment. A public school or university deploying a high-risk system must produce a fundamental rights impact assessment covering what the system does, who it affects, what harms are foreseeable, what oversight exists and what complaint mechanism applies — and they cannot write it without information only you hold. Vendors shipping that pack are winning procurements against competitors answering with marketing material. It costs far less to produce than the deals it unlocks.
Can we train on learner data?
It depends on your contracts and it has a children's data dimension on top, and the answer is increasingly no by default. School consent to a contract is not family consent to processing a child's data, several jurisdictions restrict profiling of children specifically, and indefinite retention of learner interaction data is hard to justify. The practical consequence is architectural: design for a world where you may not pool learner data across institutions, because retrofitting that is far harder than building it.
How do we answer the efficacy question?
With a modest, honest evaluation rather than testimonials. Education buyers have been disappointed by technology sold on claims and are increasingly asking what evidence exists that it improves learning. A comparison cohort, a defined outcome and a straightforward account of what was and was not demonstrated outperforms an impressive unsupported claim with exactly the institutional buyers you want. Building the measurement into the product makes it a continuous answer rather than a one-off study you have to commission.
Should we build one model across all our institutions?
It depends on whether your cohorts genuinely differ, and in education they frequently do — curriculum, age range, language background and prior attainment all shift the distribution. A single pooled model is dominated by your largest customers and serves the others with a model fitted to somebody else's learners; per-institution models starve on small cohorts. A shared base with institution-level adaptation captures most of both, at real operational cost. The contract position on data pooling may settle this before performance does.
Related verticals
Organisations in edtech usually share data, buyers or regulators with these. All fourteen are listed on the Public Sector & Education page.
AI for K-12 Education
Teacher workload, attendance analytics and accessibility — with detection and emotion inference declined outright.
Read more →AI for Higher Education
Student support, research administration and admissions — with the integrity detection evidence stated plainly.
Read more →AI for Vocational Training Providers
Content adaptation, completion modelling and funding administration — with competence certification left to assessors.
Read more →Tell us what the problem looks like.
Thirty minutes, no charge, no deck. We will tell you whether this is an AI problem, a data problem, or a process problem — and we will say when the honest answer is to buy something rather than build it.