Handing a task to an AI agent is easy. Defining what it can do without asking is the part that protects you. Behavioural scientist Dr Gleb Tsipursky, CEO of Disaster Avoidance Experts, explains the one-page register that sets the limits.
An AI agent can work through multi-step tasks on its own and use connected tools to take action, which makes delegation more consequential than asking a chatbot for a draft.
OpenAI’s guide for developers describes agents as systems that independently accomplish tasks on a user’s behalf. For a senior EA, the question becomes what the agent has been authorised to do without asking.
The US National Institute of Standards and Technology (NIST) says organisations should define clear roles for human oversight of AI. A delegated-work register turns that principle into something you can use in a live executive workflow.
What the register looks like
The register can be one page.
Give every recurring agent workflow five fields:
- Its purpose
- What it may monitor
- What it may do independently
- What needs approval
- Who can stop it
Take board-meeting preparation, for example. Assuming your organisation has approved an agent for this, it may watch the executive calendar and a designated board folder.
It can draft an agenda, identify missing papers and prepare reminder messages. But it can’t send those messages, move a board meeting, share a paper outside the approved group or alter the final agenda without your approval.
That distinction matters because the OWASP Top 10 for LLM Applications (a widely used security reference) warns that giving an AI system too much functionality, permission or autonomy can allow damaging actions.
The result is a clear boundary for the agent, a clear point where you step in, and an owner the executive can see.
Set approval around consequence
The mistake I see most often is delegating the outcome without defining the decision rights.
“Keep the CEO’s diary up to date” sounds clear to a person who already understands relationships, sensitivities and changing priorities. But an agent may read it as permission to reschedule meetings, contact external stakeholders or act on incomplete context.
The immediate cost is usually rework. In a higher-stakes workflow, the mistake can expose confidential information or create a commitment the executive never intended. Responsibility for what the agent does stays with your organisation, not the tool.
A practical test is reversibility. If an action is easy to undo and carries little consequence, the agent can usually act within the register. If it sends something externally, changes a system of record, touches sensitive information, commits money or materially affects a person, route it back for approval.
The register should also name who can stop the workflow. You, or another designated owner, should be able to pause it when context changes or output looks wrong. IT or the system administrator needs a reliable technical stop when the problem sits below the workflow level. NIST also recommends mechanisms to override, shut down and decommission AI systems after deployment.
How to get started
- Pick one recurring, lower-risk workflow this week, using an agent your organisation has approved.
- Write the five fields, run the agent against them, and record every exception for a week.
- Tighten or expand permissions based on what actually happens.
That way, autonomy grows from observed performance rather than enthusiasm.
disasteravoidanceexperts.com
