Most organisations now understand that they need rules for the use of generative AI. The more difficult question is whether those rules influence what happens when employees use AI in real work.

A September 2026 study commissioned by SAS found that seven out of ten UK and Irish enterprises had written an internal policy covering employee use of generative AI. Yet only one in eight could demonstrate that the rules were being followed. Just 12 per cent of the technology leaders surveyed described their AI governance framework as well established and comprehensive.

The study was based on 100 enterprise technology decision-makers, so it should be treated as a strong warning signal rather than a complete picture of every organisation. Even with that qualification, the central problem is clear: writing a policy is easier than establishing command.

Policy establishes intention. Command establishes control.

A policy can define approved tools, prohibited uses and expectations concerning confidential information. It can state that people must check AI output and remain responsible for decisions. Those are necessary foundations, but a document cannot observe behaviour, test compliance or intervene when practice departs from the rules.

Command exists when an organisation knows how AI is being used, who has authority to use it and who remains accountable for the result. It also requires a practical way of detecting when the organisation's stated position and its actual behaviour have separated.

The Chartered Governance Institute describes governance as the framework by which organisations are directed and controlled, including who may make decisions, who has authority to act and who is accountable. Applied to AI, this definition points beyond the production of policy. Direction must be accompanied by control and accountability.

Why the gap develops

Generative AI rarely arrives through a single, easily controlled route. Employees can use public chatbots, AI features embedded in office software, specialist services bought by individual departments and tools accessed through personal accounts. An organisation may therefore have an AI policy without possessing a reliable picture of its AI activity.

The gap grows when:

  • the policy is issued but employees are not trained to apply it to real situations
  • approved tools are listed but unapproved use is not detected
  • people are told to verify AI output but no verification method is defined
  • AI-generated material enters decisions without its origin being recorded
  • responsibility is divided between technology, legal, risk and operational teams
  • managers assume that technical controls have replaced managerial judgement

These are not merely compliance failures. They are failures of operational command. The organisation has declared what should happen but cannot establish what did happen.

When AI moves from answering to acting

The problem becomes more serious as AI develops from an advisory tool into an agent capable of taking actions. An AI system may search records, draft communications, update files, move information between services or initiate parts of a workflow. The issue is no longer only whether an answer is accurate. It is whether the system has been given appropriate authority and whether its actions can be seen, checked and reversed.

Before an AI agent is allowed to act, managers should be able to answer five questions:

  • What purpose has been assigned to it?
  • What information and systems may it access?
  • What actions may it take without approval?
  • How will its output and behaviour be verified?
  • Who retains responsibility when something goes wrong?

If those questions cannot be answered, the organisation has granted capability without establishing command.

Turning policy into an operating method

A usable AI policy needs an accompanying management method. One practical structure is:

  • Purpose: State what the AI is intended to achieve and what remains outside its role.
  • Context: Give the system the relevant information, constraints, standards and organisational circumstances.
  • Prompt or instruction: Define the task clearly, including boundaries, required output and any point at which human approval is mandatory.
  • Verification: Check claims, sources, calculations, omissions and actions using evidence independent of the AI's own confidence.
  • Evaluation: Judge whether the result is useful, proportionate and consistent with the purpose, then record what should change next time.

This approach connects policy to individual pieces of work. It makes the manager's role visible at the points where intention can otherwise be lost: when the task is framed, when authority is granted, when the result is checked and when responsibility is exercised.

Making oversight scalable

There is an obvious managerial difficulty. AI is introduced partly to save time, yet a powerful and interconnected system may create a growing burden of supervision, verification and control. AI can reduce the cost of doing work while increasing the cost of knowing what has been done.

The practical answer may be a parallel AI oversight system. The operational AI performs the work while a separately configured oversight AI monitors whether it remains within its purpose, permissions and limits. It can examine evidence, identify unusual behaviour, maintain an audit trail and escalate exceptions to a human manager.

  • Operational AI: Performs the assigned work within defined authority.
  • Oversight AI: Observes, challenges, records and escalates exceptions.
  • Human commander: Sets purpose, boundaries and risk tolerances, resolves significant disagreements and retains responsibility.

This allows managers to command by exception rather than inspect every routine action. It does not remove accountability. The organisation and its leaders remain responsible for the design of the system, the authority granted to it and the response to warnings.

Independent checking is an essential safeguard

An oversight system is not independent merely because it runs a second check. If the operational work and its verification are performed by the same model, the second result may repeat the same assumptions, blind spots and failure patterns. Using the same model twice is repetition, not independent verification.

For consequential work, using a different AI model for the checking system is one useful aid to independence. Preferably, the checking system should also have independently written instructions, separately controlled permissions and access to different evidence sources where appropriate. It should conduct its initial assessment before seeing the first model's conclusion, then report disagreement rather than being pressed into consensus.

The greater the consequence of an AI-supported action or decision, the greater the required independence between the system producing the result and the system verifying it.

A further audit layer may periodically test both systems through sampling and adversarial checks. However, endlessly adding monitors merely creates an oversight regress in which every monitor needs another monitor. The aim is a small number of genuinely different controls, not an unlimited hierarchy of similar systems.

Different model names do not automatically provide true independence. Models may share training material, architecture, suppliers or assumptions. Organisations should therefore assess independence across the model, instructions, evidence, permissions and provider. Common-mode failure is the danger: several systems can agree and still be wrong for the same underlying reason.

Evidence of command

An organisation should not judge its AI governance by the existence of a policy alone. It should look for evidence that command is operating. Depending on the scale and risk of the work, that evidence may include:

  • a current register of approved AI tools and their owners
  • clear access permissions and limits on connected data
  • records showing where AI contributed to important outputs or decisions
  • sampling of AI-assisted work rather than reliance on self-certification
  • incident reporting that includes near misses and policy workarounds
  • training based on realistic decisions rather than a policy-reading exercise
  • named accountability for monitoring, escalation and improvement
  • independent verification, which may include a different model, for consequential outputs and actions
  • periodic audit sampling designed to detect common-mode failures

The objective is not to monitor every harmless interaction. Controls should be proportionate to consequence. A brainstorming session and an AI-assisted decision affecting a customer should not be governed in exactly the same way. Command depends on understanding the difference.

The managerial question

The most useful question is not, 'Do we have an AI policy?' It is, 'Can we demonstrate that AI is being used within its intended purpose, authority and limits?'

A policy can prohibit the wrong behaviour. Command creates the conditions in which the right behaviour is understood, observable and correctable. As AI becomes more capable and more deeply connected to organisational systems, that distinction will become increasingly important.

Organisations do not retain command by preventing all delegation. They retain it by deciding what may be delegated, defining the conditions of delegation and keeping human responsibility where it belongs.

Further reading
The distinction between delegating work and surrendering responsibility is explored in Orchestration and Command in the Age of AI by R T Scanlon.

Sources

For further information: ai-humancapability.com