The most useful sentence I say
I build production agents for a living, which is exactly why the most useful sentence I say on calls is: "this should not be an agent." Roughly half the agent ideas that reach me fail one of four tests, and skipping the build is worth more than any build.
Four cases where the answer is no
- The process is deterministic. If the steps never branch on judgment (export, transform, upload), a script does it for a fraction of the cost and never hallucinates. Paying agent prices for script problems is the most common waste I see.
- The error cost is catastrophic and unreviewable. An agent that sends prices to customers, files legal documents, or moves money without a human gate is not automation. It is unexploded ordnance. Some decisions get a human in the loop permanently, and knowing which ones is the actual expertise.
- There is no number attached. "We should have AI" is not a project. If nobody can say which metric the agent moves (hours saved, tickets cleared, days-to-invoice shortened), the project has no way to be finished, only abandoned.
- The volume does not exist yet. An agent that handles twelve emails a week costs more to run and watch than the assistant hour it replaces. Below a real volume threshold, a checklist beats a system.
What to do instead
Usually one of three things: write the boring script, add a small automation inside the tools you already have, or, most often, fix the process itself first, because automating a broken process gives you broken output at scale.
Why hear this from me
A vendor who says yes to everything is selling hours. One who names the edge of the map (I will tell you what I am 20% at, and when an agent is overkill) can be trusted on the rest of it. Bring the process and the number it burns; a no from me costs you one call and saves you a quarter.