Skip to main content
Back to Resources

The AI Safety Crisis You Already Have

Oct 2026

The AI Safety Crisis You Already Have

Public debate about AI safety is increasingly framed around frontier systems: are they becoming too capable, and what happens when AI exceeds human control? Those questions matter. But they can distract leaders from the risk already inside their organisations.

AI agents are already doing real work — retrieving information, accessing systems, preparing decisions, communicating with customers, and increasingly taking actions. Each connection creates value. Each connection can also extend an agent's authority beyond what anyone intended.

The urgent issue is not whether an abstract future system might become uncontrollable. It is whether the organisation can account for the agents already operating today.

A Management Problem of the Present

Can your organisation name its agents? Explain what each is meant to do? Does each have a responsible owner, defined permissions, appropriate human oversight, and a way to be stopped?

If the answer is uncertain, this is not a safety problem for the future. It is a management problem of the present.

AI agents are not employees — but once they perform consequential work, they must be managed like consequential workers: a defined purpose and accountable owner, explicit permissions, a performance plan, human oversight and escalation, and continuous assessment through to retirement.

The Time-to-Explain Challenge

Every enterprise should assume a material agent will eventually make a poor recommendation, take an unwanted action, or behave outside its intended purpose. The test is not whether every error can be eliminated. It is whether the organisation can explain, contain, and learn from an error while it still matters.

Conventional logs rarely connect the full chain of accountability: the agent's purpose, its authority, its performance expectations, the policy that applied, the human owner, the decision history, and the corrective action.

Five Practical Disciplines

The paper sets out five disciplines for building an accountable AI workforce before scale becomes sprawl:

  • Make the AI workforce visible
  • Match control to authority
  • Define the job
  • Create the evidence chain
  • Prepare to intervene

These are not barriers to innovation. They are the conditions that let leaders approve wider deployment with confidence.

The organisations that benefit most from AI will be those that can see, measure, govern, and stand behind every agent that matters.