AI agent vs automation: which does your business need?

If the task has fixed steps and known inputs, you need automation, not an agent. If the task needs judgement over messy inputs, you may need an agent. And wherever a mistake is expensive, a human stays in the loop no matter which one you build. That is the whole rule, and the three-question version of it gives you the verdict for one specific task in about a minute. The rest of this article is what the words mean and why so much money is currently being spent on the wrong side of it.

We build and sell both, so we have no reason to steer you either way. We do have a reason to steer you away from the expensive one you do not need.

What the difference actually is

Automation is a fixed sequence written by a person: when an invoice is seven days overdue, send this template with these values filled in. Every run follows the same steps. It is cheap, fast, testable, and when it misbehaves you can read exactly what it did and why.

An agent is a system where a model decides some of the steps: read this enquiry, work out what the person wants, gather what is missing, and choose what happens next. That flexibility is real, and it is also the cost: agents are slower, pricier per run, harder to test, and they fail in creative ways that a fixed sequence cannot.

The confusion is commercial, not technical. "Agent" is the word that raises funding and closes deals in this market, so everything is being called one.

Why so many agent projects should have been automations

Ask the people who build agents for a living, and a pattern appears quickly: many of the projects they are brought should have been simple automations, costs run away silently when a model is making the decisions, and a true agent is justified only where a fixed workflow genuinely cannot cope. Industry analysts have reached the same place from the other direction: Gartner has predicted that over forty percent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.

The pattern behind both: a business had a task with fixed steps, bought a system where a model improvises the steps, and paid agent prices for workflow work, plus the debugging.

We have made this mistake ourselves and written it up: five times we removed the model from a working system, because the step needed to be right, not clever. The most instructive case is the follow-up sequence in our automation guide that contains no AI at all, on purpose, because a reminder must never paraphrase a time.

The decision rule, as three questions

Ask them in order about the specific task:

1. Could a capable employee write down the steps on one page? If yes, and the inputs arriving are tidy, it is automation. Templates, triggers, rules. Stop here and save the difference.
2. Is the input messy in ways you cannot enumerate? Rambling emails, voice notes, half-complete forms in three languages. This is where a model earns a place, usually doing one job: extracting order from the mess. The decisions that follow can still be rules.
3. What does a mistake cost? If the answer is real money or a customer relationship, a person owns that step, whatever you built around it. Our own systems draft; a person approves. The draft is the automation. The approval is the product.

Most tasks stop at question one. That is not a disappointing answer. It is the cheap answer.

Where agents genuinely earn their keep

Messy-input extraction, as above. Drafting work a person reviews: itineraries, proposals, replies, where the agent removes the blank page and the person keeps the judgement. Research and gathering: pulling together a dossier from many sources so a human decision takes ten minutes instead of an hour. In each case the model handles ambiguity, and something deterministic, rules, tests, or a person, handles correctness.

What agents do not earn is the middle of your money path. Nothing that quotes prices, commits dates, or moves funds should be improvised by a model, and any vendor comfortable with that arrangement is comfortable with your risk, not theirs.

The short version

Buy automation for fixed steps, an agent for genuine judgement over mess, and keep a human wherever mistakes are expensive. If you are unsure which side your task is on, that question has a cheap answer too: our two-week assessment names it in writing, at a fixed fee, credited against any build. We charge the same either way, so the answer you get is the real one.