The AI agent production checklist: 10 things to have before launch

Before an AI agent goes live it needs ten things: a clear job, access to real data, answers grounded in sources, scoped tool access, guardrails, human approval for risky actions, tests on real examples, logging of every run, cost tracking, and a named owner who improves it. Skip any of these and the agent works in a demo but fails in daily use.

The checklist

1. One clear job. 'Answer refund questions from the policy and past tickets', not 'help customers'.
2. Real data access. A read connection to the systems where the work happens, not a copy that goes stale.
3. Grounded answers. Retrieval over your documents, with the source shown, so people can check.
4. Scoped tool access. The agent gets only the permissions it needs, per tool, per action.
5. Guardrails. Written rules for what it must never do, enforced in code, not just in the prompt.
6. Human approval. Refunds, emails to customers and record changes wait for a person until trust is earned.
7. Evaluations. 30 to 50 real past cases with known good answers, run before every change.
8. Logging. Every run recorded with inputs, sources, actions and outcome.
9. Cost tracking. Cost per run and per month, with cheap models for easy work.
10. An owner. Someone reviews failures weekly and improves the agent monthly.

Where teams usually cut corners

Evaluations and ownership. Both feel optional in week one and become the reason the project is abandoned in month three. If you only add two things to your plan, add these.

How Infrenic handles this

Every item on this list is built into Keel, our agentic AI platform, so it is present from the 48 hour proof of concept onward rather than bolted on before launch.

FAQ

Infrenic builds AI agents on your data and tools, ships them to production and runs them, all on Keel. Send us a problem and see it working in 48 hours.