How I decide whether a workflow needs automation, AI, or neither

Every audit I run starts with a list of workflows someone wants automated, and about a third of them shouldn’t be. Not because automation is hard, but because the workflow is either too rare, too variable, or too cheap to justify a system. The expensive mistake isn’t building automation badly — it’s building it where it doesn’t belong.

Here’s the filter I actually use, in order.

Question one: is the path deterministic?

Walk through the workflow with the person who does it. If every branch can be written as a rule — “if the invoice total matches the PO, approve; otherwise flag” — you have a deterministic path, and deterministic paths belong in plain automation. No model, no prompts, no inference cost. n8n or Make will run that rule a million times without drifting.

People reach for AI here because it’s current, but a rules-based workflow with an LLM in the middle is strictly worse: slower, costlier, and now it can be wrong in creative ways. Save the model for where rules run out.

Question two: does the variation carry judgment a model can make — and a human can cheaply verify?

Some workflows resist rules: a support ticket phrased forty ways, a scanned document with coffee on the total, an inbound lead that might be a partnership inquiry. That variability is where AI earns its keep — classification, extraction, drafting.

The second half of the question matters just as much: can a human verify the output faster than they could produce it? A drafted support reply takes seconds to approve and minutes to write — great AI economics. A generated legal clause takes longer to verify than to write — terrible economics. The verify-to-produce ratio predicts whether an AI step will survive production better than any benchmark.

Question three: does the volume pay for the system?

A workflow that runs twice a month doesn’t repay a build, no matter how annoying it is. My rough line: if the work costs less than an hour a week and touches one person, document it in a checklist and move on. Systems have a maintenance tax, and every automation you own is a thing that can break on a Tuesday.

When a client’s support triage passed all three questions — variable inputs, cheap human verification, hundreds of tickets a day — we built an agent for it, and it cut response time 74%. The workflows that failed the filter got checklists. Both outcomes were wins, because the goal was never “add AI.” It was “stop losing hours.”