Notes from our project work on moving AI agents from a demo to a supported business process.
An AI agent is software that can plan a few steps, call other systems, and finish a task with limited human input. Most of the agent projects we see start as demos. Fewer reach day-to-day use.
The projects that do reach production tend to have the same basics in place. The agent has a narrow, written job. It can only use the systems it needs. It keeps a record of what it did. A person reviews its work at defined points.
Permissions deserve the most attention. An agent with broad access can cause real damage, and an agent with too little access is not useful. We treat agent access the same way we treat service accounts: reviewed, logged, and changed when the job changes.
Testing is also different. A single prompt test does not show whether an agent completes a multi-step task correctly. We test against real past cases and track how often the agent finishes the job without help.
Cost needs tracking from the start. A long-running agent makes many model calls, so we report cost per completed task alongside accuracy and response time.
If you are starting out, pick one process with a clear measure of success and a natural point for human review. Application support tickets and finance reconciliations are common first choices.
