A common example is an enquiry form. Code receives the message, a model labels it as a new project, a support question or spam, and code sends each label down its own branch. For new projects the model drafts a reply, and a person approves it before it goes out. The model answers narrow questions. It never decides what happens next.
For most business tasks it's the better place to start. A fixed path makes a known number of model calls, so the cost is predictable, and each model step can be tested against real past cases before anyone relies on it. OpenAI's guide to building agents says that when a task doesn't clearly need an agent, "a deterministic solution may suffice."
The part to design carefully is where a person steps in. Steps that only read or draft can run on their own. Steps that send, pay, delete or publish should wait for a human to approve them until they've built a long record of being right.
Updated September 2026
Check it yourself