AI agents vs automation: what a business owner actually needs to know
Two words used interchangeably by people selling them, and a simple test for which one your problem needs.
You are being sold "AI agents" and "automation" as if they were different products. Mostly they are the same idea at different levels of judgement, and the distinction that matters is much simpler than the vocabulary suggests.
Plain automation: the rules are known
Traditional automation does exactly what it is told, in the same order, every time. When this file arrives, put these columns in that system and email this person.
It is fast, cheap and completely predictable. It also breaks the moment reality is untidy — a missing column, an unexpected word, a date written differently.
An AI agent: the rules cannot be written down
An agent is used where the input is human and messy. A supplier writes "can you push the delivery to after the holiday" — no plain rule can extract a date from that, but a language model can read it, decide what it means, and act.
The trade is predictability. It handles the mess, and occasionally it is wrong, so anything that matters needs a check.
The test
Ask one question about the step you want handled:
- Could you write the rule down completely, with every exception? → plain automation. Cheaper, faster, will not surprise you.
- Does it need reading, judgement, or understanding messy human writing? → an agent, with a human check on anything expensive.
- Is the process itself wrong? → neither. Fix the process first.
A broken process does not improve by being automated. It gets faster at being broken.
Most real systems are both
The contact form on this site is a fair example. An AI reads an incoming message and summarises it, because messages are human and unpredictable. Everything after that — store it, email two people, move it into a pipeline — is plain automation, because those rules are perfectly knowable.
That is usually the right shape: AI where the mess is, plain rules everywhere else. Paying a language model to do what an ordinary rule does is slower, more expensive and less reliable.
What to ask whoever is selling it
- Which steps use AI, and which are plain rules?
- What happens when the AI gets it wrong — who notices, and how?
- What does it cost per run, and how does that change with volume?
- What happens when the input arrives in an unexpected shape?