コンテンツに移動

From assistance to authority: A CISO playbook on AI governance 

Helping organizations focus on containment, rather than prevention

Ross McKerchar

Access our CISO playbook on AI governance

 

If you had asked me, as recently as late 2025, if I’d ever allow our engineering workstations to be controlled by an AI agent, I’d have said something along the lines of, “Not until I retire.” Now, in late 2026, it’s part of our everyday workflow. The AI revolution – from new models and capabilities to adoption and integration – is moving faster than many of us expected, and governance is struggling to keep up.

In some ways, that’s not surprising. Governance has typically lagged behind technological change: some organizations embraced cloud services before establishing effective oversight, adopted SaaS platforms faster than they could develop management frameworks, and found shadow IT spreading beyond the reach of existing controls.

AI is no different. In fact, the governance challenges may be even greater.

Many organizations begin with acceptable-use questions, but agentic deployments add a second problem: governing delegated access and action. Employees are using LLMs in their day-to-day work. AI capabilities are appearing by default in productivity suites and business applications. Departments are building automations and agents. And many businesses, including ours, are deploying agentic systems that can retrieve information, connect to internal resources, invoke tools, and carry out actions on behalf of users.

The question is no longer "Can employees use AI?" but "How do we safely delegate authority to AI systems?"

Why AI governance is a board-level issue

When a generative AI tool drafts a document, the consequences of failure may be limited. However, when an AI system can access customer information, connect to internal systems, retrieve sensitive data, and trigger automated processes, it changes the risk profile significantly. Governance therefore becomes less about individual prompts and more about visibility, ownership, delegated authority, supplier risk, and architectural controls. Organizations need to understand what AI systems exist, who owns them, what they can do – and what to do when something goes wrong.

But by the same token, we know that risk appetites are changing. We've experienced that ourselves. On our own AI governance journey, the emphasis has been on safe and transparent empowerment to innovate, as opposed to prohibition. Containment is the new prevention.

Learning through implementation

In my view, it's important for internal IT and security teams to be on the leading edge of AI adoption. If they’re not, there will be other teams in the organization doing it for them, in ways that are not necessarily visible – which is a sobering prospect.

That's doubly true for Sophos as a vendor. We want to be ahead of our customers when it comes to technological adoption, so that we can better protect them from new threats. And so, over the last few years, and particularly since early 2026, our security teams have been navigating many of the challenges our customers are now facing: more vulnerabilities to triage, more agents to keep safe, and more code to threat-model. In addition, we’ve dealt with internal AI adoption, setting up an AI Safety Board, supplier and platform assessment, agent security research, and shadow AI discovery. At the same time, our researchers in SophosLabs, Managed Detection and Response (MDR), the Counter Threat Unit (CTU), Sophos AI, and Incident Response (IR) have been examining and dealing with how AI is changing both attack and defense across the threat landscape.

As a result of this work, we've realized that traditional governance approaches are not necessarily sufficient anymore and may even be counterproductive. Many organizations do not need another high-level AI principles document. Instead, they want practical, operational guidance on what to build and how to implement it, with the emphasis on enablement rather than restriction, and on containment rather than prevention.

Introducing the CISO playbook for AI governance

The latest Sophos CISO playbook (available here, along with our other CISO playbooks) is designed to help organizations establish governance for AI systems – particularly agentic AI – in a way that enables experimentation and innovation within permissive boundaries, and safe deployment at scale, rather than simply regulating AI usage.

The playbook focuses on practical implementation and includes a toolkit of downloadable assets that organizations can select from as needed and adapt to their own environments:

  • AI acceptable use policy (AUP) templates: A general AI AUP template, plus two further AUP templates on generative AI and building and deploying agents
  • AI Safety Board charter template: A template for a cross-functional, decision-oriented governance forum, based on the charter we use at Sophos – with an emphasis on enabling innovators
  • AI vendor due diligence questionnaire: A collection of questions to ask suppliers
  • Architectural patterns for agentic deployments: Tips on reducing the blast radius of agentic deployments, including an agentic architecture readiness worksheet
  • AI delegated authority assessment sheet: A worksheet to assess when AI systems move from assistance to authority
  • Hunting for shadow AI: Tips on discovering hidden AI in your estate and what to do when you find it, along with potential telemetry sources
  • Model Context Protocol (MCP) governance assessment checklist: An example review checklist for MCP servers, tools, and connectors
  • Preparing for the vulnerability flood: Guidance on preparing to absorb AI-scale vulnerability discovery, including a practical readiness assessment
  • Implementation guide: A short guide that ties the toolkit together and provides advice on how to use it as part of a phased approach for AI governance, including a 90-day planner

The objective is not to provide a universal framework, or legal or regulatory compliance advice. Instead, the playbook offers a set of tools, templates, and practices that organizations can adapt to their own operating model, risk appetite, and legal and regulatory requirements.

Where appropriate, we’ve included hints and tips on how AI itself can be used to help with AI governance.

Who is this playbook for?

We’ve deliberately covered a lot of bases in this playbook, reflecting our experiences and thinking at Sophos. Our hope is that organizations will adapt and tweak the assets to suit their own environments, requirements, and risk appetites, and so it’s not intentionally aimed at any particular sector or maturity level.

That being said, this playbook may be of particular interest to organizations that are:

  1. Interested in the concept of AI authority. The playbook treats authority as a critical governance variable. The key question is not which model is being used, but what the AI is allowed to do.
  2. Experiencing AI adoption now. Rather than treating AI as a future initiative, the playbook starts from the premise that AI use already exists across most organizations.
  3. Keen to connect governance to operations. Many governance frameworks focus on policy. This playbook also addresses architecture, ownership, supplier assessment, review processes, monitoring, and proactive security measures. We deliberately scoped the playbook along these lines because governance ultimately succeeds or fails through operational reality, rather than policy language alone.
  4. Reluctant to restrict innovation opportunities. A blanket ban is unlikely to address all AI use and may drive some activity out of view (although targeted prohibitions remain appropriate for unacceptable uses). The playbook is designed to help organizations embrace opportunities, as safely as possible.

AI governance as an enabling function

Perhaps the most important lesson from my own experience is that governance should enable adoption, rather than impede it. When governance becomes synonymous with bureaucracy and prohibits experimentation and adoption, employees may seek workarounds. Conversely, when governance provides clear routes to adoption, visible ownership, and practical, reasonable guardrails, organizations can gain both innovation and control. As AI continues to move from assistance to authority, that distinction will only become more important.

Some of the playbook content may not be appropriate for all organizations, and we have erred on the side of being comprehensive (although the toolkit should not be considered exhaustive). We offer this toolkit as a public resource and encourage organizations to adapt the assets, consider which to prioritize/start with, trim as appropriate, and adjust their approach over time to suit their environments and requirements.

You can access the Sophos CISO playbook for AI governance here.

Acknowledgments

Thank you to the following people for their contributions to the playbook:

  • Matt Wixey, Mindi McDowell, and Ryan Westman (X-Ops Insights)
  • Kim Fleming (Trust and Assurance)
  • Bud Ellis and Sally Adam (Product Marketing)
  • Daniel Dobrygowski (Senior Legal Counsel)
  • Rajeev Kapur (Engineering)
  • Chris Longman and Dan Hall (Product Management)