The problem

Schools trust SGAF with their governance records. Models make mistakes, providers change their terms, and model versions are retired on a schedule nobody controls. AI in a product like that has to be governed like any other risk, not added and hoped about.

Model type

SGAF uses large language models from one provider through a private API, matched to the job. A small, fast model handles classification at volume, such as checking page after page of a school website. A mid-tier model handles synthesis, policy judgements and a second check on automated results. The latest version of that mid-tier model writes board report commentary and analyses minutes. Model choices live in one registry, so an upgrade is a one-line change with a known reach. When a retired model version broke a live tool, the fix went into that registry, and the tools that write for boards now move to new models on their own validation schedule.

Much of the product uses no model at all. The data pipelines, the risk scoring and the release checks are deterministic code, so the same input always gives the same answer.

How outputs are evaluated

  • An evaluation harness. Fixed test cases run through five deterministic graders and one model acting as a judge, against a frozen baseline. A drop of 10 per cent or more fails the run, and the same harness compares models before a switch.
  • An accuracy benchmark. The website checker runs against real school websites, and every failure it reports is checked by a person to find false positives.
  • People before publication. Extracted improvement plans are reviewed by an operator before a school sees them, board papers are reviewed before they are saved, and nothing reaches a governance statement until the board adopts it.
  • Labels. Findings that came from a model are marked, so a board reads them as written judgements, not facts.
  • Review after release. AI help is judged on real use, not the demonstration.
  • Independent review of the code. Changes to sign-in, permissions or data access are audited by a model from a different provider before they merge.

The impact assessment

A data protection impact assessment covers the AI processing described in SGAF's data processing agreement. Personal data is kept out wherever it is not needed: the minutes review redacts names, email addresses and phone numbers before any model reads the text, and it warns about confidential items first. Anything that would rate a named school has to pass an impact assessment before it is shared beyond me.

Vendor vetting

The AI provider is held to written terms. Customer data is never used to train a model, the provider typically keeps inputs and outputs for no more than 30 days, it holds independent security certification, and transfers out of the UK are covered by recognised safeguards. Every approved sub-processor is named in the data processing agreement, and schools get notice of any change with 14 days to object. Plugins and integrations are security-vetted before they touch the system.

The result

AI adds speed without becoming the decision-maker. Every model output is either tested, reviewed by a person or labelled, and the rules are written down where anyone can check them.