Moving an AI agent from prototype to production means dealing with seven topics the demo avoids: real data, access rights, permission to act, evaluations, monitoring, costs and maintenance. None is spectacular; each can stop the project.
1. Real data
The prototype runs on ten hand-picked examples. Production gets everything: crooked scans, empty fields, formats that change from one supplier to the next. Test on a real sample, exceptions included, before promising anything.
2. Access rights
An agent that reads the company’s systems must show each person only what they could already see. This is designed from the start: limited service accounts, permissions inherited from the source systems, sensitive data filtered.
3. Permission to act
Drafting a reply and sending it are two different things. Every outward action (sending, publishing, paying, changing a system) must follow a clear rule: who approves, above which threshold, and what can go out on its own once trust is established.
4. Evaluations
Without an evaluation set, there is no telling whether a change improves the agent or makes it worse. It takes real cases, with their right answer approved by the business, replayed before every change of prompt, tool or model.
5. Monitoring
In production, what the agent does must be visible: volumes, approval rates, corrections made by the teams, errors from the systems it calls. A drop in quality should trigger an alert, not a complaint three weeks later.
6. Costs
A single model call is cheap; multiplied by thousands of cases and reasoning steps, it becomes a line to watch. The model is chosen task by task, what can be cached is cached, and the cost per case handled is measured.
7. Maintenance
Models change, and so do processes and systems. An agent without a team to look after it quietly degrades. Before a project starts, who will run it, and with what budget, must be known.
In what order
A single process first, measured before and after. The agent in suggestion mode: it prepares, a person decides. Autonomy opened task by task, when the numbers justify it. It is slower than a demo, and much faster than a project that never ships.