Building an agent that answers questions well is a weekend project. Building one that a finance team relies on to close the month is an engineering discipline. The difference is rarely the model.
What breaks between demo and production
- Silent failure. The agent answers confidently with stale or partial data and nobody notices for weeks.
- No handoff. There is no defined moment where a human approves, corrects or takes over.
- Unclear scope. The agent has access to everything, so every incident becomes a security review.
- No memory of what happened. Without an audit trail you cannot explain a decision to an auditor or a customer.
The production checklist
Before an agent touches a real process, it should have:
- A bounded scope: explicit data sources, explicit actions, explicit limits.
- Identity and permissions: the agent acts on behalf of a role, and inherits that role's access, not more.
- A human handoff point: the step where a person approves before anything irreversible happens.
- Observability: every run logged with inputs, retrieved context, output and cost.
- A rollback path: actions that can be reversed, and a defined owner who can reverse them.
- An accepted quality bar: measured against real cases, agreed with the business area before launch.
Start where the handoff already exists
The fastest path to production is to automate a process that already has a review step: invoice validation, contract triage, ticket classification. The human checkpoint is already part of the workflow, so the agent can take over the volume while the team keeps the judgment, and the trust is built with evidence instead of promises.