Enable AI with confidence. Keep control of the risk.
AI is now inside email, code, customer service, security operations and, increasingly, operational decisions. The question for leaders is no longer whether the organisation uses AI. It is whether anyone can say where it is used, what data it touches, what it is allowed to do, and who is accountable when it gets something wrong.
AI governance is not a separate programme. It extends the security, data, risk and assurance disciplines leaders already run, with new failure modes to account for.
In plain terms
AI governance answers four questions
Where are we using AI?Every tool, model and supplier feature, including the ones staff found themselves.
What can it see?Which data it can read, and which data must never go into it.
What can it do?Whether it only suggests, or can send, change, buy or switch things on its own.
Who answers for it?A named owner, a way to check it works, and a plan for when it goes wrong.
Layer 1 · GovernWho decides, and on what basis?
- Named executive accountability for AI risk
- AI policy and acceptable use, including external AI tools
- An inventory of AI systems, models and AI-enabled suppliers
- Risk appetite and approval thresholds by use case
- Cross-functional oversight: security, data, legal, risk and the business
Layer 2 · SecureCan the AI be misused, manipulated or leak data?
- Data classification and controls on what AI can access
- Defences against prompt injection and data poisoning
- Least privilege for AI agents and their tool connections
- Model, plugin and AI supply chain assurance
- Logging of prompts, outputs and agent actions
Layer 3 · AssureHow do we know it keeps working as intended?
- Testing and AI red teaming before release
- Ongoing monitoring for drift, misuse and abnormal behaviour
- Human oversight where decisions affect people or safety
- Internal audit and third-party assurance
- Board reporting on AI risk and incidents
Control across the AI lifecycle
Controls belong at each stage, not only at launch.
- Stage 1
Intake and approval
Register the use case, classify its risk and data, and confirm who owns it.
- Stage 2
Build or buy
Assess models, vendors and data sources. Contract for security, data use and incident notice.
- Stage 3
Test and red team
Test for prompt injection, data leakage, harmful output and unsafe agent actions.
- Stage 4
Deploy and monitor
Limit permissions, log activity, watch for drift and misuse, and keep humans in the loop.
- Stage 5
Review and retire
Reassess as models, data and regulation change. Retire systems cleanly, including their data.
The risks leaders are managing
Sensitive data exposure
Confidential, personal or client data pasted into AI tools, or surfaced by an AI assistant to people who should not see it.
Shadow AI
Staff using unapproved AI services and browser extensions outside security, data protection and contract controls.
Prompt injection and manipulation
Hidden instructions in emails, documents or web pages that hijack an AI system into leaking data or taking actions.
Over-privileged AI agents
Agents connected to email, files, code or business systems with more access and autonomy than the task needs.
AI supply chain
Third-party models, plugins and AI features added to existing software, often without a fresh risk assessment.
Unreliable output in decisions
Confident but wrong answers feeding financial, clinical, legal or operational decisions without adequate checks.
AI-enabled attacks
Deepfake voice and video fraud, convincing phishing at scale, and faster vulnerability discovery by attackers.
AI near operational technology
AI-driven optimisation or analytics connected to control systems, where an unsafe recommendation or action has physical consequences.
Legally bindingRegulation and law
What applies, and when.
European Union · AI Act
- Prohibited AI practices and AI literacy duties: applying since February 2025
- General-purpose AI model obligations: applying since August 2025
- Transparency obligations (Article 50): applying from 2 August 2026
- High-risk AI (Annex III): deferred to 2 December 2027 by the Digital Omnibus on AI, in force since 27 July 2026
- High-risk AI in regulated products (Annex I): deferred to 2 August 2028
United Kingdom
- No AI-specific statute: existing law and sector regulators apply
- UK GDPR, Data Protection Act 2018 and the Data (Use and Access) Act 2025, including automated decision-making
- Equality, consumer protection and sector rules, including FCA expectations for regulated firms
Elsewhere- A growing patchwork of national, state and sector rules. Check each jurisdiction you operate in.
Voluntary, unless required by contract or regulatorStandards and guidance
How to build and evidence good practice.
Management systems and riskISO/IEC 42001 (AI management system)ISO/IEC 23894 (AI risk management)NIST AI RMF and Generative AI Profile
AI securityUK AI Cyber Security Code of PracticeETSI EN 304 223NCSC and partners' guidelines for secure AI developmentOWASP Top 10 for LLM ApplicationsMITRE ATLAS
Data protectionICO guidance on AI and data protection
Same principles, different stakes by sector
Enterprise and technology
Copilots and assistants with broad access to email and files. Data access controls and labelling matter more than the model itself.
Financial services
Model risk management, explainability, fair treatment of customers and resilience of AI-dependent important business services.
Energy and OT
AI may advise, but safety-relevant changes to control systems need engineering approval, fail-safe design and human authority.
Healthcare and higher education
Patient and student data, research integrity, clinical safety and academic openness, often with limited central control.
What leaders must decide
- Who is accountable for AI risk, and does the board know?
- Do we know every AI system and AI-enabled supplier we rely on?
- Which data must never enter external AI services?
- Which decisions can AI make alone, and which always need a human?
- What can an AI agent do without approval, and how is that limited?
- How do we test AI for manipulation before it goes live?
- Can we detect and investigate misuse of an AI system?
- How do we assess AI features quietly added to existing software?
- Where is AI allowed near operational technology, and under whose authority?
- How will we evidence compliance as AI regulation changes?
The leadership principle
Banning AI pushes it into the shadows. Allowing it without control moves risk faster than the organisation can see it.
The goal is governed adoption: clear accountability, controlled data, bounded autonomy and evidence that it works as intended.