AI agents for a one-person business: what actually works today

The first thing I gave an agent free rein over was my inbox. It was supposed to triage, draft replies to routine questions, and flag anything unusual. Within two days it had drafted a perfectly polite reply to a prospective client confidently offering a discount I don’t offer.

It never sent it, because I’d put a review step in front of anything outbound. That review step is the entire reason this story is amusing rather than expensive, and it’s the main thing I’d tell anyone experimenting with agents.

What an agent actually is

Cutting through the marketing: an agent is an automation that decides.

A normal automation follows a path you built. When a form is submitted, add the row, send the email, create the task. Completely predictable, and it breaks the moment reality doesn’t match the path you imagined.

An agent gets a goal and works out the steps itself. It can read a messy email and decide what kind of request it is. It can handle the case you didn’t anticipate. It can also decide to do something you’d never have chosen, which is the same property viewed from the other side.

That’s the whole trade. Flexibility for predictability. Everything about using agents well follows from taking that seriously.

Where agents genuinely earn their place

The tasks where deciding is the hard part and the output is reversible:

Inbox triage. Reading messy incoming email, categorising it, drafting replies, flagging what needs you. Genuinely good, because email is unstructured in a way rigid automations handle badly.

Research and monitoring. “Check these ten competitors weekly and tell me what changed.” Needs judgement about what counts as a change, which is exactly the agent-shaped part.

Content repurposing. Reading a long piece and deciding what the good extracts are. Structural judgement, low risk, easy to review.

Lead qualification. Reading an enquiry and deciding whether it fits, with reasoning you can check. AI lead-gen tools for solo consultants covers this in context.

First-pass data cleanup. Messy inputs into consistent formats, deciding what belongs where.

Notice the pattern: every one produces something you look at before it matters.

Where they fail expensively

Anything sending money. Obvious, and worth stating.

Anything client-facing without review. My discount incident. The model isn’t lying, it’s filling a gap with something plausible, and plausible is dangerous when it concerns your prices.

Long chains of steps. Errors compound. An agent that’s 95% reliable per step is about 60% reliable across ten steps, and the failures are silent.

Anything irreversible. Deleting, publishing, sending, committing. If undoing it is hard, don’t delegate the decision.

Tasks with a fixed correct path. If you can describe exactly what should happen, you want an automation, not an agent. It’ll be faster, cheaper and it will never improvise.

That last one covers most of what people ask agents to do. 7 Zapier automations every one-person business should steal covers the plain-automation approach, and it solves more problems than agent enthusiasm suggests.

The design rule that makes them safe

One sentence: agents draft, humans dispatch.

Everything an agent produces that will reach another person, or change something you can’t undo, passes a checkpoint. In practice that’s a review queue, a draft folder, or a notification asking for approval.

This sounds like it removes the benefit. It doesn’t, because the expensive part was never clicking send, it was reading, deciding and composing. An agent that hands me eight drafts to approve in four minutes has saved me an hour, and the four minutes is what keeps me out of trouble.

The second rule: scope the tools narrowly. An agent with access to your entire business will eventually use a bit of it you didn’t consider. Give it the minimum it needs for the task.

Building one without writing code

The mainstream automation platforms have absorbed agent-style steps, which is the practical route for most solo operators. Zapier lets you drop a decision step into an otherwise normal workflow, which is exactly the right shape: deterministic plumbing with judgement at the one point that needs it.

That hybrid is what I’d actually recommend. Not “an agent runs my business”, but “a normal automation handles the flow and asks a model to make the one call that requires reading something”.

If you’re weighing platforms, Zapier vs n8n covers the trade-off, and Zapier vs Make the other obvious alternative.

What this replaces, honestly

There’s a lot of talk about agents replacing a first hire. My experience is narrower than that.

An agent handles the reading-and-deciding layer of routine work. It doesn’t handle relationships, judgement about your business, or anything requiring context it doesn’t have. The AI stack that replaces your first virtual assistant is the honest version of that comparison, and how to hire and onboard your first contractor with AI covers when a person is genuinely the answer.

The realistic framing: agents remove a category of small decisions that used to interrupt you. That’s valuable. It’s not a member of staff.

My take

Start with one task, reversible, with a review step. Inbox triage is the obvious first candidate because email is messy enough to justify the flexibility and drafts are harmless.

Then resist expanding. The temptation after a good first result is to give it more scope, and scope is precisely where agents go wrong. Errors compound across steps, and an agent with broad access will eventually make a decision you’d have vetoed.

And before you build any agent, ask whether a plain automation would do. Most of the time it would, and it’ll be more reliable, cheaper and completely predictable. The discount my agent nearly offered is a good reminder that “it decides for itself” is a feature and a liability in the same breath.