What happened
In its Q3 2026 AI Workflow Index, Zapier compared two ways to run AI work. One is a scheduled agent that uses a model to check for work and decide what to do. The other is a step workflow that starts when something happens, and calls a model only when reasoning is needed. Zapier modeled 2,944 AI workflows from 360 companies. On its central estimate, the scheduled agent cost 2.1x as much overall. For jobs with fewer than 30 completed runs a month, which were 61% of the sample, it cost 8.3x as much. For jobs that ran 30 or more times, the multiple was 1.2x. These are modeled estimates at October 2026 list prices, not bills. Zapier also notes that a scheduled agent can still spend tokens finding out there is nothing to do, and that for most workflows in the study, letting software run the path cost less.
Primary source: Zapier, Q3 2026 AI Workflow Index
My take
I would default to a fixed workflow, and use an agent only for the judgment call inside it. For a 10 to 50 person team, a lot of useful jobs live in that low volume group: a weekly check, a form that arrives a few times a month, a renewal that almost never fires. A scheduled agent still spends tokens looking for work when the queue is empty. Software can wait for the event and then run the steps.
The plain test I would use is this. Write the steps for one real job, as you would hand them to a new teammate. If you can write every step for one real run, it is a workflow. The step where you would write "it depends" or "use judgment" is where the agent goes. Keep that step narrow. Do not hand the model the whole path just because one decision is hard. Build the path first. Call the model when a person would have to decide, not to watch an empty queue.
Next step: Book a call at codedre.com/call.






