UK AI assurance: what NCSC, ICO, DUAA and SS1/23 now expect from enterprise AI

The United Kingdom took a deliberately different road to AI regulation: no single AI statute, but five cross-sector principles — safety, transparency, fairness, accountability and contestability — applied by existing regulators through existing powers. That makes assurance evidence, not policy paperwork, the currency of compliance. The stakes are real: according to techUK's report "AI Adoption in the UK: Putting AI into Action", AI could add up to £550 billion to UK GDP by 2035 — but only organisations that can prove their AI is governed will capture that value.

What do the NCSC's secure AI guidelines actually expect?

The NCSC's Guidelines for Secure AI System Development — published jointly with the US CISA and co-sealed by agencies from 18 countries — establish a secure-by-design baseline across four phases: secure design, secure development, secure deployment, and secure operation and maintenance. Translated into enterprise practice:

• Threat-model prompt injection and data poisoning before go-live, not after the first incident.
• Know your AI supply chain: which models, which connectors, which data each one may touch.
• Log AI-specific events so an incident can be reconstructed, step by step.

DUAA: "meaningful human involvement" is a higher bar than an approve button

The Data (Use and Access) Act 2025 (DUAA) replaced Article 22 UK GDPR with new Articles 22A–22D, in force since February 2026. Significant automated decisions are now broadly permitted — but the safety valve is meaningful human involvement, and the emerging consensus is blunt: the human role must be active, not a token gesture. A reviewer who rubber-stamps machine output without the context, the data and the authority to change the outcome does not count. The ICO, which consulted on updated automated decision-making guidance in early 2026 and is preparing a statutory code of practice on AI, treats accountability for AI outcomes as non-negotiable. Ask yourself:

• Does the reviewer see the inputs and the reasoning — or only the verdict?
• Can the reviewer actually override the decision, and does the record prove it?
• Is every automated decision traceable to a model, a prompt and a data source?

Why does SS1/23 pull AI into model risk management?

For banks, building societies and PRA-designated investment firms, the PRA's supervisory statement SS1/23 — five model risk management principles in force since May 2024 — treats AI like any other model: inventoried, validated, monitored, with a named senior manager accountable. Vendor models are explicitly in scope, so "our API provider handles that" is not a defence. FCA-regulated firms face the same direction of travel under the senior managers regime.

What should a UK enterprise demand from an AI platform?

• Verifiable processing records: append-only, per-call logs of every model call, retrieval and action — evidence, not marketing claims.
• Human oversight in the execution path, capturing reviewer context and authority so DUAA reviews are demonstrably meaningful.
• On-prem deployment: keeping models, data and audit trails inside your own perimeter sidesteps the concentration concerns behind the UK's critical-third-party regime.

Regulators will not read your AI policy. They will ask for the record.