The one test that separates operational AI from everything else
After the AI says it did something, go look. If a lead exists, an appointment sits on the calendar, an invoice carries a number, or a status changed — and the record shows who or what changed it — that was operational. If the only artifact is a message in a chat window, it was conversation.
This is a harsher test than it sounds. A great deal of software described as agentic produces confident summaries of work that never touched the database.
What operational AI has to have
- A registry of actions with defined inputs, outputs and side effects.
- Identity resolution: which user, which organization, which role.
- Permission checks evaluated server-side, before the action runs.
- Confirmation gates for destructive, financial and external-send actions.
- A persisted result the operator can open, and an audit record of the attempt.
Current capability, declared capability and future capability
Three things get blurred in AI marketing and they should be kept apart.
- Current capability: the action is registered, permission-checked, executed against production data and verified by a persisted record.
- Declared capability: the action appears in a manifest, catalog or tool list. That means it can be discovered — not that it has been proven end to end.
- Future capability: on a roadmap. Useful to discuss, never presentable as something the system does today.
What operational AI is not
- It is not unsupervised authority. Scope is defined by role, not by the model's judgment.
- It is not a replacement for the person accountable for the outcome.
- It is not deterministic automation; rules fire the same way every time, an operator interprets a request first.
- It is not a guarantee of correctness. It is a guarantee that the attempt is recorded.