What an AI agent can and cannot do for a small business

I build these for clients and I run several inside my own business, so this is written from operating them rather than from watching a demonstration. The demonstrations are misleading in a specific way: they show the part that works and end before the part that does not.

The useful question is not whether an agent can do a job. It is what happens on the twentieth attempt, when the input is malformed and nobody is watching.

What are they reliably good at?

Work where the input is messy, the output is structured, and a wrong answer is cheap to catch.

Reading a hundred enquiry emails and tagging each with what the person actually wants. Turning a pile of meeting notes into a task list. Pulling the details out of supplier invoices into a spreadsheet. Drafting a first version of something a person will edit anyway. Watching a feed of mentions and deciding which three need a human today.

The common shape is that a person still sees the result, and the cost of the agent being wrong is that the person corrects it. That is a good trade, and at volume it is a large one.

Where do they fail?

Work where a wrong answer is expensive and nobody is positioned to catch it.

Sending anything to a customer without review. Making a commitment about price, availability or a date. Anything touching money. Anything where being confidently wrong is worse than being silent, which describes most of what a business says to the outside world.

The failure is not that the model is stupid. It is that it does not know what it does not know, so it produces a plausible answer at exactly the moment a person would have stopped and asked someone.

• Reading messy input and producing structured output · hand it over

  Tagging, extracting, sorting, summarising. This is the sweet spot and it is worth real money.

• Drafting something a person will edit anyway · hand it over

  The first version, not the sent version.

• Deciding what deserves a human's attention today · hand it over

  Triage is a genuinely good use, because the human still decides.

• Sending anything to a customer unreviewed · keep a person on it

  The confident wrong answer goes out at the same speed as the right one.

• Committing to a price, a date or an availability · keep a person on it

  A plausible invention here is a liability, not an error.

• Anything that moves money · keep a person on it

  No exceptions worth the saving.

What does one actually cost to run?

Less than people expect to operate, more than people expect to build properly.

The running cost of the models themselves has fallen far enough that for most small-business workloads it is not the constraint. What costs is the surrounding work: connecting it to the systems that hold your data, deciding what happens when it fails, and building the place a person checks its output.

That last part is where budgets go. An agent with no review surface is simply unaccountable, and the cost arrives later in a form nobody attributes to it.

What is the honest failure mode?

Automation that moves work rather than removing it.

I have built things that looked like a success on the day they shipped and turned out to have created a new daily job: checking whether the automation did the right thing. If the checking takes as long as the task did, the project was a lateral move with a subscription attached.

The test I now apply before building anything is simple. What is the work after this exists, honestly stated, including the checking? If that number is not clearly smaller than the work today, I say so rather than build it. It has cost me a couple of projects and saved several clients more than the projects were worth.

Where should a small business start?

With the task somebody currently does that is high volume, low judgement and visible when wrong.

For most businesses that is enquiry handling: reading what came in, working out what it is, routing it, and drafting the first reply. It is the right starting point because it is measurable, because the failure is cheap, and because the improvement shows up somewhere the owner already looks.

I have written about the wider version of that in a marketing team of one, which is the account of running this stack on my own business rather than a client's.

What would I tell you not to do?

Do not start with the thing that is most impressive to describe. Start with the thing that is most annoying on a Tuesday.

The businesses that get value from this are the ones that automated something dull and kept going. The ones that do not are the ones that tried to build the clever thing first, discovered how much of the work is plumbing and review, and concluded the technology was overhyped rather than that the project was badly chosen.

If you want the boring version built properly, that is what I do.