The AIRA Guide

What is AIRA, and is it mandatory?

It is the DTA's common blueprint for designing and governing government AI. It is voluntary, but review boards will use it as their checklist.

Does every AI system need the same controls?

No. Every system uses the same foundation, but the depth of oversight, traceability and evaluation should match the risk of the use case.

Does human oversight mean approving every AI action?

No. Humans set the limits, automation runs within them, and approval is reserved for irreversible or high-impact actions.

Can you reconstruct an AI decision months later?

Only if you record the inputs, sources, versions, output and approver for every decision, and can roll back a bad change.

How do you build dependable AI around imperfect models?

Evaluate on real cases before release, use checks outside the model, and abstain or hand over to a person when unsure.

How do you keep AI sovereign and avoid lock-in?

Know where every component stores data, keep the model a setting rather than a foundation, and agree an exit plan before you sign.

Should you build, buy or blend an AI platform?

AIRA says neither is right by default. Most organisations blend a proven platform with an expert team and a trusted local partner.

How do you compare AI vendors like for like?

Score 42 AIRA-based questions on evidence, not claims, weighted by risk, and make vendors demonstrate oversight and rollback live.