“Agent” is the most overused word in software right now. Vendors attach it to everything from a chatbot to a scheduled spreadsheet macro. That confusion is expensive, because agents and automation cost very different amounts to build, run and trust.

Automation follows rules

Classic automation is a fixed recipe: when a form is submitted, create a record, send an email, update a status. It’s cheap to run, completely predictable and easy to audit. When it fails, it fails loudly and in the same place every time.

Its weakness is variety. The moment inputs stop looking alike — a quote request written as a paragraph, an invoice in a new layout — the rules multiply until nobody wants to maintain them.

That’s not a reason to skip it. Most of the value in a typical process still comes from plain, well-built automation: moving data between systems, triggering the next step, notifying the right person. It’s the foundation everything else sits on.

Agents handle judgment

An agent reads unstructured input, decides what it means and chooses the next step from a set of tools it’s allowed to use. That makes it good at the messy middle of a process: understanding an email, classifying a request, drafting a reply.

The trade-off is cost and certainty. Agents are slower and more expensive per run, and they’re probabilistic — which is why they need guardrails, evaluation sets and a clear path to a human.

A quick way to choose

Ask three questions about each step in the process:

• Are the inputs structured? If yes, use automation.

• Could you write the decision as a short rule? If yes, use automation.

• Does a person currently read, interpret or write something? That step is a candidate for an agent.

A useful default: Automate everything you can describe as a rule, and reserve agents for the one or two steps that genuinely need judgment. It keeps cost down and makes the whole system easier to debug.

The best systems use both

In practice, almost every system we ship is a hybrid. At Northwind Freight, plain automation receives each email and files it; an agent extracts lanes and dimensions; automation pulls rates from the pricing tool; and the agent drafts a quote that a rep approves in one click.

Rules give you reliability. Agents give you reach. Put each where it’s strongest.

Designing it this way keeps the expensive, probabilistic part small and well supervised — and that’s what lets a team trust it with real work.