By Andrew Leonenko
AI can summarize a document, prepare a draft, or gather the details needed for a decision. The harder design question is what should happen next. When does a helpful suggestion become an action that affects a customer, account, or business record?
For us, the answer starts with the consequences of the action—not with how capable a model appears in a demo.
Let software prepare; keep authority explicit
A model can propose the next step. The application should enforce whether that step is permitted. A request to change a record still needs to pass the same identity, access, and business-rule checks as any other request.
This separation keeps the model useful without making it the source of authority. It also makes the system easier to inspect when a result is surprising: the model suggested an action; application policy allowed or rejected it; a person may have approved it.
Match review to the stakes
A reversible draft usually needs less ceremony than a payment, a permission change, or a message sent on someone's behalf. Review should be strongest where an action is hard to undo, affects another person, or creates a real obligation.
A useful approval request shows the proposed action, its target, the reason it was suggested, and the likely effect. A reviewer should not need to read a long chat transcript to understand what they are approving.
Make a pause safe
When work waits for approval, preserve the proposal and enough context to understand it. If the workflow resumes later, check that the user still has permission and that the relevant record has not changed. An approval from yesterday should not silently authorize an action against a different state today.
Retries need the same care. A network timeout can leave the system unsure whether an action completed. Use idempotency or a status check before trying again, especially when the action has an external effect.
Keep the boundary small and understandable
The best workflow is not the one with the most confirmation dialogs. It is the one where each capability has a clear scope, important actions have a deliberate handoff, and people can see what happened afterward.
That gives AI room to reduce repetitive work while keeping responsibility legible. A system should be able to explain what it prepared, what it tried to do, which rules applied, and where a person stepped in.
For a more technical treatment, see Agent workflows need control points, not just prompts.
Andrew Leonenko
Founder, Leonenko Group LLC