Agents or automation? How to decide, step by step

The simplest system that works is usually a mix: agents where judgment is needed, plain automation where it is not. Here is how to decide for each step.

3 min read
On this page
  1. Three options for every step
  2. Questions to ask about each step
  3. Why the simplest system wins
  4. A worked example

In short

  • Rule-based steps with clean input: use plain automation. It is cheaper, faster and fully predictable.
  • Steps that need judgment or read messy input like emails and documents: use an agent.
  • Decisions that carry real risk: keep a person in charge, with the agent preparing the work.

Once a process is mapped, every step needs a decision: who or what does it? It is tempting to give everything to an agent. That makes the system more expensive, slower and harder to test than it needs to be.

Three options for every step

OptionBest whenExample
Plain automationThe rules are clear and the input is structured.When an invoice is marked paid, update the order status.
AgentThe step needs judgment or reads unstructured input.Read an incoming email, work out what the customer needs and prepare the right update.
PersonThe decision carries risk, or the rules do not cover it.Approve a refund above a set amount, with the agent's summary in front of you.

Questions to ask about each step

  1. Is the input structured? A form field or a database value is structured. An email, a PDF or a phone note is not. Unstructured input is where agents shine.
  2. Can the rule be written down completely? If yes, automate it. If the honest answer is "it depends", that step needs judgment.
  3. What does a mistake cost? Low cost and easy to reverse: let the system act and log it. High cost or hard to reverse: add an approval point.
  4. How often does it happen? Rare steps may not be worth building for at all. A person can keep doing them.

Why the simplest system wins

Every agent you add is something to test, monitor and limit. Every plain automation is something that will do exactly the same thing every time. Anthropic's guide to building effective agents makes the same point: use the simplest approach that works, and add agentic complexity only when it clearly pays for itself.

In practice, a well-built system often has one or two agents doing the judgment-heavy work, surrounded by deterministic steps that move data reliably between systems.

A worked example

Take client onboarding. A new client signs. Today, someone reads the contract, creates records in three systems, sends a welcome pack, requests documents and chases them.

  • Creating the records from the signed form: automation.
  • Reading the documents the client sends back and checking they are complete: agent.
  • Writing a personal welcome note based on the sales call notes: agent, reviewed before sending at first.
  • Approving an exception to the standard terms: person, with the agent's summary.
  • Reminding the client about missing documents on a schedule: automation.

Nothing in that list is exotic. The value comes from the whole process running as one coordinated system instead of a chain of handoffs.

Written by Achraf Ouahab

Founder of ACHRAF OUAHAB MEDIA LLC, which builds agentic AI systems that scale business operations on infrastructure the business owns.

About the founder →

Start here

Have a process in mind? Let's talk.

Your website, member portal and admin portal on one foundation you own, with the agentic AI systems built in. You ask, they do the work, you approve what matters.