Most enterprises now have an agent pilot that works. Far fewer have an agent that runs every day inside a production workflow. The distance between the two is rarely the model.
The pilot proves the wrong thing
A pilot is designed to answer one question: can a model do this task? It runs on a curated dataset, a friendly user group, and a single integration — often read-only. Production asks different questions. Can the agent act on systems of record safely? Who is accountable when it is wrong? What does each completed task cost?
When a pilot succeeds, it is tempting to treat the remaining work as scaling. In practice, it is a second project with a different shape.
The model was never the bottleneck. The fourth integration was.
Identity is the first wall
Enterprise systems grant access to people, through roles designed years ago. Agents need narrower, auditable permissions that can be revoked without breaking a workflow. Programs that design agent identity early — service principals, scoped tokens, per-action logging — move faster later, because security review becomes a checklist rather than a negotiation.
Exceptions decide the economics
A demo handles the common case. Production is defined by the long tail: the malformed record, the missing approval, the customer who replies in two languages. Every unhandled exception becomes a manual task, and enough of them erase the business case. Designing exception paths — and routing them to a named person — is the work that turns an agent into an operation.
You cannot govern what you cannot see
Agents make many small decisions. Without tracing, replay, and task-level evaluation, nobody can say whether quality is improving or drifting after a model update. Observability is not a later phase; it is what allows risk and audit teams to approve a release at all.
Autonomy should be earned through measured performance, not granted at launch.
What production actually requires
- 01A workflow owner who signs off on outcomes, not just a sponsor who funds the pilot.
- 02Scoped agent identities with per-action audit logs.
- 03Risk tiers that define which actions run autonomously and which require approval.
- 04Task-level evaluation sets that run before every release.
- 05Tracing, cost-per-task reporting, and a team accountable for run.
None of this is exotic. It is the same discipline enterprises already apply to any system that touches money, customers, or safety. Agents simply make it impossible to skip.