Agents vs. Automation: How to Know Which One You Actually Need

Automation is a promise: "I know exactly what needs to happen, so I wrote it down."

Agents are a different kind of promise: "I know what needs to get done, but I don't know the path until I'm on it."

Most companies deploy one when they need the other, and that's where the problems start.

Where the decision actually lives

Automation is powerful when you can specify every step before execution. A workflow that pulls invoice data, validates it against your ERP, and posts the journal entry — that's automation. The steps don't change. The logic is fully written before anything runs.

An agent is the right tool when the steps can't be pre-specified. Not because the requirements are vague, but because the task requires judgment calls that depend on what you find along the way.

Research is the clearest example. "Compile a competitor analysis" isn't a workflow — it's a goal. Which sources to consult, when to go deeper on a thread, when to abandon one — those are judgment calls. Automation can't make them; an agent can.

The heuristic I've landed on: if you can write the steps down in full before running the task, use automation. If writing the steps down requires hypotheticals that cover every possible thing you might find, you're in agent territory.

Where companies go wrong

The most common mistake: deploying an agent when automation would work fine. Agents are slower, more expensive to run, harder to evaluate, and more likely to produce unexpected outputs. If the task is deterministic, use automation. Agents are for the work that's inherently not.

The second mistake: trying to automate something that requires judgment, watching it break on edge cases, and concluding automation doesn't work for that use case. It doesn't — for that task. The problem is the match between technology and task.

The sequencing is where most of the value is

Most real workflows contain both deterministic and non-deterministic stages. Getting the split right is where the real opportunity is.

A lead enrichment workflow might be 80% automation — pull contact data, check enrichment sources, update the CRM — with an agent layer for cases that don't fit the standard pattern: a company with no website, a contact whose role doesn't match your ICP definition, a duplicate that might or might not be the same person. Automation handles the routine 80%; the agent handles the 20% that would otherwise fail silently or land in someone's queue.

If you're figuring out where AI can help your business, skip the question "where can I deploy agents?" The better question: where in my existing workflows does judgment appear? Address that specifically.

The cost of choosing wrong

Automation with uncovered edge cases becomes a maintenance liability. Someone catches the failures, debugs the cause, patches the workflow, repeats. A workflow with enough edge cases stops being an asset.

Agents that should be automations cost more per run, produce inconsistent outputs, and introduce failure modes that fixed workflows don't have. An agent can make a decision that's locally reasonable but globally wrong in ways you won't catch until downstream.

The teams getting the most out of AI right now are the ones who've gotten disciplined about matching the type of work to the type of solution — not the ones with the most agents.

Figure out where judgment lives in the workflow. The rest of the design follows from that.