AI systems for education and training providers
The Department for Education's own recommendation is that personal data isn't put into generative AI tools. That single line rules out most of what's currently being sold to schools — and it's the reason we build systems that run on your infrastructure, where the data doesn't have to go anywhere at all.
Where the week actually goes
Teacher workload has improved slightly but remains poor: a substantial minority of teachers and leaders still report their workload is unacceptable and outside their control. The Department's own analysis points at the same handful of tasks.
Reporting, planning and resource preparation
Marking and feedback, lesson planning, resource creation, report writing and parent communications are consistently identified as the highest-potential areas for workload reduction. They're also the tasks that reliably eat evenings.
New statutory evidence obligations
Keeping Children Safe in Education requires filtering and monitoring effectiveness to be reviewed at least once each academic year, led by a named senior leader with the DSL and IT, covering every internet-connected device — with a record that the review happened. Somebody has to produce that record.
KCSIE 2026 in force since 1 September 2026Re-mapping evidence to a new inspection framework
Ofsted's framework changed in November 2025 — handbooks replaced by toolkits, report cards instead of a single overall grade, and a five-point scale. Every provider is re-mapping how it evidences quality against a structure that didn't exist two years ago.
Funding compliance in FE and apprenticeships
Audit findings recur with depressing consistency: insufficient evidence of off-the-job training, incomplete initial assessments, missing commitment statements, unsigned agreements. The consequence is clawback, and the cause is almost always document completeness rather than delivery quality.
What we'd build for a school, trust or training provider
All of these support staff. None of them assess, grade or make a judgement about a student.
Funding evidence completeness checking
Checks apprenticeship and learner files against the funding rules, flags missing off-the-job records, absent initial assessments and unsigned agreements — before an audit finds them rather than after.
The least contested, highest-value build in FE.Compliance evidence assembly
Assembles the filtering and monitoring review record, inspection evidence packs and policy gap analyses against the current frameworks, so the evidence exists as a by-product rather than a project.
Resource and differentiation drafting
Drafts teaching resources and differentiated versions from your own schemes of work and materials, for a teacher to review, adapt and own. Built on your content, not scraped from the open internet.
Report comment drafting
Turns teacher-supplied evidence and notes into draft report comments in your house style. The teacher edits and approves every one; the system never generates an assessment of a child it has no evidence for.
Parent communications and translation
Drafts letters and translates communications into families' first languages, which is high-volume, genuinely valuable, and completely non-judgemental work.
Admissions and enquiry triage
Routes inbound enquiries, drafts acknowledgements and tracks what's outstanding — administrative handling only, with admissions decisions left entirely to the people who make them.
Triage only — see the limits below.The rules that govern AI in education
Education carries a specific combination: children's data, statutory safeguarding, and assessment integrity. All three constrain the design.
- <strong>The DfE recommends personal data is not used in generative AI tools.</strong> It's a recommendation rather than a prohibition, and the guidance contemplates use with protective measures — but it is the clearest signal in the sector, and it's why we build so that pupil data never needs to leave your environment.
- <strong>KCSIE 2026 has been in force since 1 September 2026.</strong> It replaced KCSIE 2025 as the operative guidance. The 2026 edition brings AI explicitly into safeguarding scope, including digitally altered and entirely AI-generated imagery.
- <strong>An AI tool cannot be the sole marker.</strong> JCQ's position is that centres may permit AI to assist with marking, but a human assessor must review the work in its entirety, determine the mark, and remain responsible for the grade awarded.
- <strong>Assessment and admissions uses are high-risk under the EU AI Act.</strong> Annex III, point 3 covers AI determining access or admission to education (3(a)) and evaluating learning outcomes (3(b)). Following Regulation (EU) 2026/1744 those obligations apply from 2 December 2027. For UK providers the Act generally applies only where output is used in the EU.
- <strong>Emotion recognition in education settings is a prohibited practice</strong> under the EU AI Act, subject to narrow medical and safety exceptions. That covers the “engagement detection” and attention-monitoring products currently being marketed into classrooms. We won't build them.
- <strong>Ofsted's framework changed in November 2025</strong> — inspection toolkits, report cards, and a five-point scale from Exceptional to Urgent improvement, with safeguarding graded separately as met or not met.
What we won't build for an education provider
Students can't meaningfully consent to being assessed by software, and the consequences of getting these wrong follow a young person for years.
Sole marking, grading or anything feeding a qualification
JCQ is explicit that an AI tool cannot be the sole marker and a human must determine the mark. We build tools that help a teacher mark faster; we don't build tools that mark.
Admissions and access decisions
High-risk under the EU AI Act, and life-shaping for the individual. A system can handle the administrative process around admissions. It has no role in who gets in.
Safeguarding decisions
AI may surface a flag from a record. The designated safeguarding lead makes the judgement, records it and owns it. This is the area where an automated inference creates dangerous false reassurance.
Determining that a student used AI dishonestly
AI-detection tools are unreliable and produce false positives that fall hardest on students writing in a second language. An allegation of malpractice must be a human judgement on evidence, following your centre's procedures.
Behaviour prediction or risk-scoring of pupils
Profiling children to predict future behaviour is something we decline on principle as well as on law. It embeds existing bias and it follows a child through their record.
Common questions.
Can teachers use AI to mark student work?
+
As an aid, yes. As the sole marker, no. JCQ's position is that a centre may decide whether teachers use AI tools to help mark work, but a human assessor must review the work in its entirety, determine the mark, and remain responsible for the grade awarded. In practice that means AI can help with consistency checks, first-pass feedback drafting and flagging where a mark scheme has been applied unevenly — with the teacher making every judgement that counts.
Can schools put pupil data into AI tools?
+
The Department for Education recommends they don't. Its guidance states that personal data should not be used in generative AI tools, while contemplating use with appropriate protective measures where necessary. That recommendation is the single clearest reason to build on infrastructure the school controls: if the system runs inside your environment and the data never leaves it, the question largely resolves itself. Any build touching pupil data should also have a Data Protection Impact Assessment behind it.
Can AI detect whether a student used AI?
+
Not reliably, and we'd advise against treating any detector's output as evidence. Detection tools produce false positives, and those false positives fall disproportionately on students writing in a second language or with an atypical writing style. An accusation of academic malpractice is a serious matter for a young person and should rest on a human judgement about the whole picture — drafts, process, and the student's own account — following your centre's malpractice procedures.
What does KCSIE say about AI?
+
The 2026 edition, in force since 1 September 2026, brings AI explicitly into safeguarding scope — including digitally altered and entirely AI-generated imagery — and treats generative AI as a risk factor in online safety. It also requires filtering and monitoring effectiveness to be reviewed at least once each academic year, led by a named senior leader with the DSL and IT, covering all internet-connected devices, with a record kept. Until September 2026, KCSIE 2025 remains the operative guidance.
How would this work across a multi-academy trust?
+
One school first, then extended. We build and prove a workflow at a single setting, then roll it out with school-level differences configured rather than rebuilt — which avoids the common trust problem of a dozen settings running slightly different versions of the same thing that nobody centrally understands. Central team workflows like policy version control and compliance returns usually justify themselves fastest, because the duplication across settings is the whole problem.
Start with the administrative load.
Tell us what's taking staff time that has nothing to do with teaching. If it involves judging a student, we'll tell you it's the wrong place to start.