The one-line difference
Workflow automation follows steps that are known in advance. If the form says X, do Y. When an invoice arrives, create a record, notify finance, file the PDF. Every run takes the same path.
An AI agent decides the next step itself, because the right step depends on judgment: understanding an unusual customer message, reading a non-standard document, choosing which of six paths applies.
The test: could you write the rules down?
Sit with the person who does the work today and try to write their process as if-then rules. Three outcomes are common:
- The rules fit on a page. That's automation. Most processes land here, and most people over-estimate how much judgment theirs really needs.
- The rules fit, except for one step such as "work out what the customer is actually asking" or "pull the totals out of whatever format the supplier sent". That's automation with one AI step inside it, which is the most common shape of a good build.
- Every case is a judgment call. That's where an agent earns its cost, with guardrails, escalation to a person, and monitoring.
What each costs to own
Automation is cheaper to build and close to free to run: a workflow that fires a thousand times a month costs much the same as one that fires ten. When it breaks, it breaks loudly and in the same place, which makes it easy to fix.
An agent costs more on both counts. Each run calls a model, so running cost scales with volume. Behaviour can drift as inputs change or models update, so it needs a test set, monitoring, and periodic tuning. None of that is a reason to avoid agents; it's a reason to use them only where the judgment is worth paying for.
Why we usually start with automation
Automating the predictable steps first is cheaper, it works immediately, and it shows exactly which one or two steps genuinely need an agent. That's a far better basis for the decision than guessing upfront, and it means the agent, if you need one, is built around a process that already runs.
Two adjacent cases are often mistaken for either. If you need a system to answer questions rather than run a process, that's a RAG chatbot. If what you need is invoices and contracts read and posted into your systems, that's document processing.
Three examples
- Client onboarding (signed proposal, create folders, send access requests, schedule kickoff): pure automation.
- Lead routing (read the enquiry, decide whether it's a fit and who should take it, then create the CRM record and notify): automation with one AI step to read and classify the enquiry.
- Support triage across a messy inbox (understand the request, look up the order, decide whether to answer, refund, or escalate): an agent, with a person reviewing anything it isn't sure about.