When it comes to finance, compliance or safeguarding, government has clear structures for independent scrutiny. No one asks a department to audit its own books. Yet many are still asking delivery teams or vendors to certify their own AI.
AI systems can affect thousands of people at once. When they go wrong, the damage spreads quickly and can be difficult to undo. Despite this, many programmes continue to rely on internal teams or suppliers to vouch for performance.
That is a risky assumption.
A vendor claim of fairness is not a guarantee. A lack of visible problems does not mean the system is working correctly.
As Nicholas Shearer, an ethical technologist with government experience, warns in our new ‘AI readiness in government’ white paper: “The mistake is assuming you can buy an AI tool and just put your data into it. That’s not a good idea.”
So, what does good look like?
Independent validation. Separate scrutiny. Testing for performance, fairness, transparency, and unintended consequences.
The UK’s AI Assurance Playbook recommends independent benchmarking as standard. At 2i, that involves:
• Testing real-world scenarios reflecting the diversity of service users.
• Stress-testing models under unusual or high-risk inputs.
• Ensuring audit trails show how decisions are made.
These actions protect delivery. When scrutiny comes, the worst position to be in is having to say, “We relied on the supplier.”
Assurance provides reputational protection.
AI pilots may be underway - but is your assurance defensible?
Talk to our team about how public sector leaders are embedding structured testing and governance before scaling further.
