AI Risk Management: From AI Inventory to a Practical Risk Register
The UK Government has published a practical starting point for AI risk management. For a business, the first question is simpler: do you know enough about the AI you already have to manage what could go wrong?
AI adoption often arrives before formal governance. Employees may use ChatGPT, Microsoft Copilot, Claude, Gemini, meeting assistants, embedded SaaS features or AI-enabled automation without a deliberate organisation-wide plan.
You cannot manage AI risk until you know what AI you have.
The AI Risk Management Toolkit, published by DSIT on 8 September 2026, is guidance rather than law. It does not itself create legal compliance, and it does not mean every SME must adopt another formal framework.
The EU AI Act timetable for high-risk AI has changed under the 2026 AI Omnibus, while AI literacy remains an obligation for relevant providers and deployers. Article 4 now focuses on taking measures to support AI literacy rather than requiring a prescribed or “sufficient” level. Microsoft is also introducing stronger governance controls for AI agents in Power Platform and Copilot, including authentication, approved identity providers, anonymous-access and agent-sharing controls.
Businesses therefore need visibility not only of the AI tools people use, but also the AI agents, identities, permissions, data and actions being created around them.
From AI discovery to AI risk management
Connect the AI estate — including AI tools, embedded AI, AI agents, identities, permissions, connectors and data sources — to purpose and data, risk, ownership, controls or treatment, monitoring and evidence. Discovery tells you what AI you have; risk management helps you understand what could go wrong; governance decides who owns it and what controls are needed.
What the toolkit is for
The guidance supports people involved in the design, operation, procurement and delivery of AI-enabled products. It describes risk identification and assessment, treatment, monitoring and reporting across the AI lifecycle, with appropriate records, responsibility and multidisciplinary involvement.
AI risk is not just cybersecurity
Depending on the use case, consider security, privacy, accuracy, legal and regulatory context, fairness, transparency, operational resilience, suppliers, financial exposure and reputation. The proportionate questions depend on the data, people, permissions, autonomy and consequences involved.
The EU AI Act timetable has moved — the governance problem hasn't
The main EU AI Act regime is now in its enforcement phase, but the 2026 Digital Omnibus changed important high-risk implementation dates. High-risk systems under Annex III and Article 6(2) now apply from 2 December 2027. AI systems covered by regulated products under Article 6(1) and Annex I apply from 2 August 2028. Other provisions, including relevant transparency and enforcement requirements, apply on their own timetable.
Article 4 still requires providers and deployers to take measures to support AI literacy for staff and other people dealing with AI systems on their behalf. The amended wording does not prescribe a particular or “sufficient” level. Practical evidence might include AI guidance, approved-tool rules, data-handling expectations, role-appropriate training or briefings, human-review expectations, escalation routes and records showing what measures were taken. These are governance examples, not a statutory checklist or guarantee of compliance.
AI governance is moving beyond chatbots
Microsoft's enterprise AI ecosystem increasingly includes agents that can access data, connectors and business processes. Microsoft lists enhanced Power Platform and Copilot agent security and governance controls for general availability in September 2026, including Microsoft Entra ID authentication, approved identity providers, anonymous-access controls, agent sharing controls and enforcement of administrative policies.
Microsoft lists these capabilities for general availability in September 2026. Availability should be verified in the individual tenant before relying on them operationally.
An AI inventory therefore needs to ask what AI agents exist, who created and owns them, who can access them, what data and connectors they use, what actions they can take and whether they are still required. Entra ID, tenant security, authentication, access controls, Purview or DLP, endpoint security and other Microsoft 365 technical controls may sit within Altitude IT or another specialist technical implementation.
A practical AI risk register
An illustrative register might record the AI system or tool, business purpose, owner, data involved, risk, likelihood, impact, existing control, required action, action owner and review date. It is not a prescribed statutory template, and existing risk, security, privacy or compliance registers may already provide a suitable home.
Where discovery fits
An AI risk register built from assumptions is not very useful. The AI Discovery & Readiness Review starts by establishing the AI estate you actually have — including the tools, licences and, increasingly, AI agents being used across the organisation — then assesses risk, cost and waste, governance, opportunity and readiness before practical next steps are prioritised.
You cannot govern AI — or AI agents — you cannot see. The DSIT toolkit helps organisations think systematically about AI risk; the EU AI Act creates legal obligations where applicable; Microsoft governance controls provide technical mechanisms that can help control parts of a Microsoft AI estate; and ISO/IEC 42001 provides an organisational AI management-system framework where appropriate. These are different things, not one regulatory regime.
Altitude AI provides practical business and technology guidance, not legal advice, regulatory opinions or certification. ISO/IEC 42001 certification is separate, and EU AI Act applicability requires case-specific consideration.