What people actually mean by "AI agent", and when you would want one
A prompt, a workflow and an agent are three different things. For most small businesses the boring middle option is the right answer far more often.
What people actually mean by “AI agent”, and when you’d want one
“Agent” is doing a lot of work in AI marketing at the moment, and most of that work is making a fairly simple idea sound more impressive than it is. Once you strip the word back to what it actually means, the decision about whether you need one gets a lot easier, because it turns out most businesses don’t — not yet, and maybe not for most of what they do.
It helps to think of this as a ladder with three rungs. Almost everything sold to you as “AI” sits on one of them, and knowing which one tells you what you’re actually being pitched.
Rung one: a prompt
This is the simplest thing on the ladder, and the one most people are already using without thinking about it as a category. You ask a question or give an instruction, the tool answers, and that’s the whole interaction. Draft this email. Summarise this document. Suggest three headlines. A person is in the loop for every step — they typed the request, and they’ll read the answer before it goes anywhere. Nothing happens automatically. Nothing is triggered. It’s a conversation, and it ends when the conversation ends.
This is genuinely useful, and it’s low risk, because a human is checking the output before it does anything in the real world.
Rung two: a workflow, or automation
This is a step up in usefulness and, done properly, barely a step up in risk. A workflow is a fixed sequence: when this happens, do this, then this, then this. A new enquiry comes in through the website, it gets logged, a confirmation goes out, a task is created for someone to follow up. Every step is decided in advance by a person. The AI, if there’s AI in it at all, might be doing one particular job inside that sequence — reading a message and sorting it into a category, say — but the sequence itself doesn’t change and doesn’t decide anything for itself.
The defining feature of a workflow is that it’s entirely predictable. Run it a hundred times with similar inputs and you get a hundred similar outcomes. You can test it, you can explain it to someone in a sentence, and when something goes wrong you can point to exactly which step failed.
Rung three: an agent
An agent is different in kind, not just degree. You give it a goal and some tools, and it decides its own steps to get there. “Find out why this customer’s order is late and draft a response” is an agent-shaped task — it might check an order system, look at a delivery tracking page, read past correspondence, and decide in what order to do that. Nobody wrote out those steps in advance. The tool worked them out itself, on the fly.
That’s the whole appeal, and it’s also the whole risk. The flexibility that lets an agent handle a messy task is what makes it unpredictable. Run it on ten similar cases and you might get ten different paths through the problem — most of them fine, one of them not, and you won’t always know which until after it’s done something.
Why the boring middle rung wins most of the time
For most small and medium businesses, most of the time, a workflow beats an agent, and it’s worth being blunt about why: predictability is a feature, not a limitation, when money or customers are involved. You want your invoicing to work the same way every time. You want a booking confirmation to say the same correct thing every time. You do not want a system that occasionally, creatively, decides to handle a customer enquiry in a way nobody anticipated — even if nine times out of ten that creativity would have been fine.
A well-built workflow is also easier to trust, hand over, and fix. You can explain it to a new staff member in two minutes. You can point to the exact step that broke when something goes wrong. An agent that went sideways is much harder to debug, because the question isn’t “which step failed” — it’s “why did it decide to do that”, and the honest answer is sometimes that nobody fully knows.
Where agents genuinely earn their place
None of this means agents are a bad idea. It means they’re the right tool for a narrower set of jobs than the marketing suggests. They earn their place on tasks that are genuinely variable — where no two instances look quite the same — and where a wrong step is cheap to recover from. Early-stage research is a good example: pulling together background on a topic, a first pass at summarising a pile of documents, triaging a batch of enquiries into rough categories before a person looks at them properly. If the agent gets one of those partly wrong, someone notices and fixes it, and nothing has shipped or been charged in the meantime.
The pattern to look for is this: variable input, judgement-light task, low cost of being wrong, and a human still checking the output before it matters. Take away any one of those four and an agent starts to look like the wrong tool.
What makes agents risky
Three things, specifically. First, non-determinism — the same input can produce a different path each time, which makes testing and trust harder to build. Second, compounding errors — because the agent chains its own steps together, a small mistake early on can shape everything that follows, and it may not be obvious until the end that anything went wrong. Third, tool access — an agent that can only draft text is low risk; one that can send emails, update records, or spend money on your behalf is a different proposition, and the cost of a bad run varies each time in a way a fixed workflow’s never does.
Questions worth asking any supplier pitching you an “agent”
- What exactly does it have access to, and what’s the worst thing it could do with that access on a bad run?
- Is a person checking the output before it reaches a customer, or before money moves?
- What happens when it gets something wrong — how would you know, and how would you undo it?
- Could this actually be a workflow instead? If the steps can be written down in advance, ask why they’re not being.
If the answers are vague, or the pitch leans hard on the word “agent” without explaining what decisions it’s actually making on its own, that’s worth noticing.
The right question
The question to bring to any AI decision in your business isn’t “should we use agents”. It’s “what is the simplest thing that solves this”. Sometimes that’s a person with a good prompt. Often it’s a straightforward workflow doing the same reliable thing every time. Occasionally, for the messy, variable, low-stakes edges of a job, it’s an agent. Start at the bottom of the ladder and only climb when the rung below genuinely can’t do the job — not because the higher rung sounds more impressive.