AI agent or workflow? Do the math before you sign

The question

What we told them

Probably not — and the reason is arithmetic, not opinion.

There are two ways to wire AI into a process. In a workflow, you decide the steps in advance and the model fills in specific ones: classify this email, extract these fields, draft this reply. The path is fixed, so the system is predictable, debuggable, and cheap to run. In an agent, the model owns the path — it decides each next step and when the work is finished. That flexibility is real, and you pay for it in latency, cost, and a failure surface nobody can fully map in advance.

Here's the math that matters. Say the model gets each individual step right 95% of the time — generous, for messy real-world data. Chain ten autonomous steps and the odds that every one goes right are about 60%. Twenty steps: roughly 36%. An agent running unsupervised across a long chain isn't impressive. It's a coin flip with confidence.

Coding agents are the exception that proves the rule: they get away with long chains because tests catch mistakes within seconds, so errors don't compound silently. Your intake process almost certainly has no equivalent — nobody finds the mistake until a client does.

So the rule we give every client: if you can draw the process as a flowchart today, build the flowchart. That's most of what we build — n8n workflows with Claude handling the judgment steps inside — and it's most of what businesses actually need. Agents earn their cost in the narrower cases where the path genuinely can't be known in advance: open-ended research, diagnosing a novel problem, work where each step depends on what the last one uncovered. Intake isn't that.