Lesson plans, resources and feedback drafted from the trust's own curriculum in minutes. The teacher reviews, changes and owns every word, and safeguarding stays with people.
At a glance
An illustrative scenario. The numbers are typical examples, not a client case study. The cloud services and UK rules were checked on 9 November 2026.
A multi-academy trust has usually invested years in a shared curriculum: schemes of work, knowledge organisers, assessment points and agreed ways of giving feedback. Yet much of a teacher's planning time still goes on finding the right part of that curriculum, adapting it for this class, and building resources from scratch. The same worksheet gets made a dozen times across a dozen academies.
Feedback is the other big load. Written feedback on a set of books or essays takes hours, and much of it repeats the same points. Teachers adapt resources for pupils with SEND or English as an additional language by hand, late at night. Workload is one of the main reasons teachers give for leaving, and every hour spent re-typing is an hour not spent teaching.
Many teachers already use free AI tools to help. That is the risk the trust actually has today: pupil work pasted into consumer services that may train on it, content that ignores the trust's curriculum, and no record of what was used. The DfE's guidance is clear that personal data should not go into tools like these without the right protections, and that staff stay responsible for anything the AI produces. The trust needs a safer route, not a ban that drives use underground.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
Self-reported in a short termly survey, checked against a time diary in the pilot academies.
Example3–5 hours a week back
How much teachers change the drafts. Very high means it isn't helping; near zero can mean drafts are going out unread.
ExampleStable, and reviewed half-termly
Share of sampled drafts that a curriculum lead judges true to the trust's scheme of work.
ExampleAbove 90% of the sample
Pupil names or identifiers found in prompts by the automated check and the monthly sample.
ExampleZero, with every exception investigated
Times the assistant stopped and handed a piece back to the teacher, reviewed with the DSLs for appropriateness. It's a backstop, not a detector.
ExampleReviewed every half term
Active weekly users, and whether teachers would recommend it to a colleague.
ExampleMost teachers in pilot academies using it weekly
Cross-sector rules are on the UK regulation map.
Ask your platform and model providers to show how they meet each standard: stated purpose, monitoring and reporting, security, privacy, intellectual property, design and testing, and governance. The standard on intellectual property says pupils' and teachers' work isn't used for commercial purposes, including model training, without the copyright owner's permission. The learner-facing standards, such as filtering, apply in full if pupils ever use the tool, which this design doesn't allow.
Teacher-facing use is where the DfE sees the clearest benefits. Staff check everything the AI produces and stay responsible for it. Personal data isn't put into generative AI tools unless the school has met data protection law, been open with pupils and parents, and has the right agreements in place.
KCSIE 2026 has applied since 1 September 2026. The filtering and monitoring standards ask schools to review where generative AI is used at least once a year, and to review early when a new AI tool is introduced. Safeguarding is never delegated to the assistant: concerns go to the designated safeguarding lead through the school's normal procedure.
Ofsted doesn't inspect AI tools. It looks at whether leaders made sensible decisions about them, including data protection, safeguarding and bias, and at the effect on pupils. Keep the decision record, the DPIA and the evaluation evidence where a headteacher can find them.
Each academy trust is a controller for pupil data. That means a DPIA before the pilot, collecting only what's needed, UK processing, a clear retention period, and providers acting strictly as processors. The Children's code applies to online services likely to be accessed by children. A staff-only tool is outside it, but if pupils are ever given access, the code applies and privacy settings start high.
Since 5 February 2026, organisations designing online services likely to be used by children must take account of children's higher protection needs. The automated decision reforms also apply, but this design makes no decision about a pupil: the teacher decides what to teach and what feedback the pupil gets.
Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.
01Channels
02Intake
03Intelligence
04Orchestration
05Systems of record
06People
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Intake: sign-in, uploads, filesTeachers sign in with the trust identity they already use. Uploaded work is deleted after the agreed short retention period. | API Gateway, S3 | API Management, Blob Storage | Kong or your existing API gateway; Ceph or IONOS S3 Object Storage |
| Reading pupil workTest on real photos of exercise books, with children's handwriting, not typed samples. | Amazon Textract | Azure Content Understanding (includes Document Intelligence) | Open-source parsing (e.g. Docling) plus an open-weight vision model |
| Removing pupil namesNames from the class list are swapped for codes before the model sees anything. The teacher's screen puts them back. | Amazon Comprehend PII detection | Azure Language PII detection in Foundry Tools | Open-source PII detection (e.g. Microsoft Presidio) |
| DraftingBoth clouds offer a choice of models, including Claude. Test each against your own curriculum and feedback policy, not a general benchmark. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Curriculum knowledge (retrieval)Only approved, current versions of the trust's schemes of work, success criteria and feedback policy. Every draft cites the unit and version. | Bedrock Knowledge Bases | Foundry IQ, built on Azure AI Search | OpenSearch, or PostgreSQL with pgvector |
| Assistant runtimeThe first release is retrieval plus drafting, with no tools that act. Keep it that way until there's a clear reason. | Amazon Bedrock AgentCore (if tools are added later) | Foundry Agent Service (if tools are added later) | An open-source agent framework (e.g. LangGraph) on Kubernetes |
| WorkflowClass sets are processed as a batch, so a slow model never blocks the teacher's screen. | AWS Step Functions | Durable Functions or Logic Apps | Temporal, or n8n |
| Middleware and events | Amazon SQS, EventBridge | Service Bus, Event Grid | Apache Kafka or RabbitMQ |
| Link to the MIS and learning platformClass lists come in read-only, for name removal. Nothing is written back to the MIS. | Read-only API through API Gateway | Read-only API through API Management | Read-only API through Kong or your existing gateway |
| GuardrailsBlocks unsuitable content in drafts, and triggers the safeguarding stop that hands a piece back to the teacher. | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, data protectionStaff accounts only. Pupils have no access. | IAM Identity Center, KMS, Macie | Entra ID, Key Vault, Purview | Keycloak or your existing Active Directory; OpenBao or HashiCorp Vault; an HSM |
| Monitoring and auditLog prompts, sources and drafts with pupil codes, not names. Keep logs only as long as the DPIA says. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyIONOS AI Model Hub runs in German data centres: EU, not UK, residency. Check the trust's own data policy before choosing it. | eu-west-2 (London) | UK South / UK West | Your own UK data centre, or a UK or EU provider |
On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →
Many organisations keep core systems and sensitive data on their own servers. This is how the same design works across both.
Stays on-prem
The management information system, safeguarding records and any on-site file servers stay where they are. Safeguarding records are never connected to the assistant at all.
Uses cloud AI
Retrieval over the curriculum library, reading of uploaded work and drafting run in a UK cloud region, or on the trust's own GPUs with open-weight models. Pupil names are removed before any text leaves the trust's control.
How they connect
Teachers reach the assistant over the internet with the trust's single sign-on. The read-only class list feed comes from the MIS through an API gateway, over a site-to-site VPN or a private link such as AWS Direct Connect or Azure ExpressRoute where the trust has one.
If the link drops
Teaching carries on as it did before the assistant. Teachers plan from the curriculum library directly, and any class sets in the queue are drafted when the service returns. Nothing about a pupil depends on the assistant being up.
What each agent may do through its tools, what it may not, and where a person signs off. The limits are enforced by the platform, not the prompt.
Drafting assistant
Where a person signs offThe teacher reviews, edits and owns every plan, resource and piece of feedback. Safeguarding concerns are dealt with by the teacher and the designated safeguarding lead, never by the assistant. The assistant's permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.
The management information system is the record for pupils and classes. The assistant reads class lists through a read-only API, only to swap names for codes, and never writes back.
Class sets go onto a queue and are drafted as a batch. If the model is slow or a provider is down, the work waits safely and the teacher is told when drafts are ready. Each upload carries an ID, so a retry never creates duplicate feedback.
The curriculum library is the source of truth for content. Curriculum leads publish approved versions into it; the retrieval index is rebuilt from it, and old versions are removed, so the assistant can't draft from a superseded scheme.
The assistant can be switched off at any time without affecting teaching, assessment records or safeguarding. It holds nothing the trust can't do without.
Measure teacher time on planning and feedback, audit the curriculum library for gaps and versions, check data flows, and agree the acceptable use policy.
Gate · must be true to passBaseline agreed, curriculum library ready, DPIA started, DSLs and unions consulted, product safety standards checked with providers.
Two subjects in 3–4 academies. Planning first, then formative feedback once planning is trusted. Teachers keep working as normal and compare.
Gate · must be true to passTime saved, curriculum alignment above target, no personal data reaching the model, DSL and DPO sign-off.
Roll out to the pilot academies' full staff, with support, monitoring and a way to switch it off.
Gate · must be true to passTraining complete, monitoring live, filtering and monitoring review updated, rollback plan agreed.
Add subjects and academies in waves, then extend to support staff and SEND adaptations.
Gate · must be true to passEach new subject's library passes the same alignment check.
| Activity | CEO / trust board | Director of education | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | DPO | DSL lead | Trust IT |
|---|---|---|---|---|---|---|---|---|
| Business case and workload benefits | Accountable | Responsible | Consulted | Consulted | Informed | Consulted | Informed | Consulted |
| Target architecture | Informed | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Informed | Consulted |
| DPIA and provider assurance | Accountable | Consulted | Consulted | Consulted | Informed | Responsible | Consulted | Consulted |
| Safeguarding design and review | Accountable | Consulted | Consulted | Informed | Informed | Consulted | Responsible | Consulted |
| Curriculum library and content | Informed | Accountable and Responsible | Consulted | Consulted | Consulted | Informed | Informed | Informed |
| Build and integration | Informed | Consulted | Accountable | Responsible | Responsible | Consulted | Informed | Consulted |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Consulted | Consulted | Consulted |
| Run, monitor and improve | Informed | Accountable | Consulted | Consulted | Responsible | Consulted | Consulted | Responsible |
R ResponsibleA AccountableC ConsultedI Informed· not involved
Option A
If there's no architect on the programme, I'd own the design end to end: the curriculum knowledge base, how pupil data is kept out of the model, the integration with your MIS and learning platform, and the gates. I'd work closely with your DPO and safeguarding leads, and report to your director of education or trust board.
Option B
If your trust's IT team or partner already has an architect, I'd take on the retrieval, name removal and drafting pipeline under their lead, and give them a second senior opinion on data protection, provider assurance against the DfE standards, and security.
Goes wrong
Teachers keep pasting pupil work into free tools because the approved one is slower or harder to reach.
How the design avoids it
Make the approved route the easy one: single sign-on, inside the tools teachers already use, and fast enough for a Sunday evening.
Goes wrong
The assistant drafts from general knowledge, not the trust's curriculum, and quality varies academy to academy.
How the design avoids it
Retrieval from the approved library only, citations on every draft, and curriculum leads reviewing a sample each half term.
Goes wrong
Staff start to rely on the assistant to spot safeguarding concerns.
How the design avoids it
Say plainly, in training and on screen, that it isn't a detector. Teachers read every piece of work, and the DSLs own the process.
Goes wrong
Feedback drafts are pasted into books unread, and pupils notice.
How the design avoids it
Short, specific drafts, an edit-rate measure, and feedback policy that says the teacher's voice comes first.
Goes wrong
Pupil names slip through in file names, headers or handwriting.
How the design avoids it
Name removal on text and file metadata, an automated check on every prompt, and a monthly sample by the DPO.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Keep this assistant teacher-facing. If the trust adopts pupil-facing tutoring later, treat it as a separate product with the learner-facing standards, filtering and the Children's code in full.
Measure workload with a baseline and a method you could share, so your evidence is comparable with what the testbed publishes.
Keep provider contracts clear on processor and controller roles, with a data map and a current DPIA. These were common gaps the ICO found in its edtech audits.
Add the assistant to the annual filtering and monitoring review, and re-check providers against the standards each year.
Good answers are specific. Vague answers to any of these are worth following up.
Whether you need someone to lead the design or to support the architect you already have, let's talk it through.
