Skip to main content
Applied AI controls

An AI agent should have less authority than its owner

An agent can only be useful in production if the team can describe exactly what it is allowed to do, what it must ask a person to approve, and how the work can be traced after the fact.

01

Turn permissions into a workflow contract

List the systems, records and tools the agent can use. Then state the least access needed for each action. A permission should be connected to a task, not granted because it might be useful later.

The same contract records who owns the workflow and which system remains the source of truth when information differs.

02

Put approval where the consequence begins

A person should approve an action that changes a customer position, moves money, changes access, closes a case or creates a compliance obligation. The system can prepare the context and route the work, but the accountable role makes the decision.

The approval should be visible in the normal operating system, not hidden in a separate technical log that users never see.

03

Plan for incomplete work

Production systems have unavailable services, missing records, duplicate events and conflicting data. An agent needs a limited retry rule, a clear stop condition and a queue or person who receives the unresolved task.

Retain the inputs, tool calls, output, approval and final outcome for each run. This makes investigation, quality review and improvement possible without exposing more data than the workflow requires.

This article provides general information. It is not legal, regulatory or security assurance advice for a particular organisation or workflow.

Related

Each article covers one operating question that comes before an AI workflow reaches production.

All Codefremics insights

Practical guidance on applied AI, platforms, payments and the controls that make digital systems usable in production.

WhatsApp us