AI Is Making Decisions at Work. Most Companies Have No Rules for That.
Security expert Stephen Wilson says businesses are handing AI tools more and more independence, but treating them with the same loose oversight they used when AI just answered questions.
Key points
- Most companies still govern AI tools the same way they did in 2023, when AI only answered questions rather than taking actions.
- Stephen Wilson, field chief technology officer at HashiCorp, an IBM company, identifies three stages of AI use: assistant, agent, and operator.
- At the "operator" stage, AI tools can run an entire marketing campaign, including writing posts and choosing where to publish, with no human in the loop for hours.
- Wilson says governance, meaning the rules that control what AI can see and do, must grow at the same pace as AI independence.
- Organisations that skip this step risk AI tools accessing sensitive data or publishing unapproved content with no audit trail.
When your company's AI tool could answer a question, the risk felt manageable. When it can now book meetings, post to social media, and move money, the rules need to change. Most haven't.
That is the central warning from Stephen Wilson, field chief technology officer at HashiCorp (an IBM company), who spoke with CSO Online about how businesses are failing to keep up with a shift already under way.
Why does it matter how much independence an AI has?
More independence means more potential damage if something goes wrong. Wilson breaks AI use into three stages, and each one carries a different level of risk.
The first is AI as an assistant. This is the familiar setup: a person types a question, reads the answer, and decides what to do. Humans are close to every step. The risk here is mostly about careless data handling, where a staff member might paste a password or a confidential client record into a chat window without thinking twice.
The second stage is AI as an agent. Here, a person gives the AI a goal and walks away. The tool completes several steps on its own, possibly handing work to a second AI tool before a human ever sees the result. Nobody is watching each individual step. Wilson says organisations at this stage must treat the AI like a team member with a proper identity, meaning a traceable, limited account rather than borrowed human credentials.
The third stage is AI as an operator. This is where AI runs an entire project. Wilson gives a concrete example: a business asks an AI system to design and execute a full marketing campaign. Two hours later, the drafts, social posts, and publishing schedule are done. No human approved any individual decision along the way.
"The level of governance and identity and auditing have to increase as your level of oversight decreases," Wilson says.
The practical danger is real. An AI system working at the operator stage could publish the wrong message, access records it should not touch, or make decisions based on incorrect information, with no easy way to trace exactly what happened or why.
Wilson makes one comparison that cuts through the complexity: governing AI should scale the way governing people does. One employee needs one set of rules. A team needs clearer processes. An entire business unit needs formal controls, approvals, and audit logs. AI is following the same curve, just much faster.
Most organisations are still somewhere between stage one and stage two. The window to build proper controls is open now, before the operator stage becomes routine.
