An illustrative scenario. There are five of you. The owner is sales, finance and tech support, sometimes on the same day. One Thursday she reads that “companies without agents will fall behind,” and by the end of the meeting someone has asked for quotes on three: one for sales, one for admin, one for customer service. None of them is crazy. Together they’re a second job nobody decided to take on.
The thesis in one line: with five people you don’t need an agent per department, you need one well-chosen agent, because the scarce thing isn’t budget, it’s the owner’s attention.
AI for micro and small businesses: the scarce resource is your attention, not your budget
First, the scale we’re talking about. According to Spain’s national statistics office (INE, Central Business Register, published December 11, 2025), on January 1, 2025 Spain had 3,310,824 active companies; 54.4% had no employees at all and 81.6% had two or fewer. The typical Spanish company is tiny. Meanwhile the European statistic on AI adoption (Eurostat, 2025 data) only covers companies with 10 or more employees: across the EU, 19.95% use AI, and among small ones, 17%. That’s an EU figure, not a Spanish one, and it says nothing about firms like yours.
What we take from that is an inference, not a data point: most of the frameworks, case studies and agent offers in circulation were designed for companies with a tech team, and the numbers don’t even look at firms your size. It’s not that AI is useless at five people. It’s that almost nobody has written the rules for that scale.
And one mechanism holds up without any figures. An agent isn’t installed, it’s supervised. Someone reads what it produces, corrects it, decides what happens when it’s unsure. At five people, that someone is you. Every agent you add isn’t one more tool, it’s one more task on a calendar that had no room left. So the useful question isn’t “what could AI do?” but how much supervision you can afford, and the honest answer is usually: one agent’s worth.
What should a five-person company automate first?
Three conditions, all at once. The work repeats every week, so the effort pays back. A mistake shows up early and costs little, so being wrong isn’t a drama. And it has a clear owner, a named person who reviews the result. If one fails, you wait. For ranking several candidates, use the four criteria in which processes to automate with AI; this is the version for when the whole team fits around one table.
| Candidate | Now? | Condition for it to work |
|---|---|---|
| Replying to repetitive email (standard quotes, appointments, usual questions) | Yes | The agent drafts, a person sends; start with a single type of email |
| Extracting data from invoices and delivery notes | Yes | Every field can be checked against the original document |
| Answering WhatsApp after hours | Yes | A closed list of questions and a handoff to a person when it’s unclear |
| Preparing quotes from your template | Yes, with your sign-off | Nothing reaches the customer without you looking at it |
| Deciding prices, discounts or exceptions | Not yet | They depend on your judgment, and it isn’t written down |
| Handling complaints and claims | Not yet | A mistake here costs customers, not minutes |
| Several agents coordinating with each other | Not yet | It triples the supervision before the first one works |
Notice that the first “yes” on the list is drafting, not sending. In a small company the draft with human review is the most sensible starting point: you get most of the saving with an error only you can see. If that’s your first case, replying to repetitive emails is the packaged version of that work.
Why one well-chosen agent and not one per department?
Because every agent has three costs and only one shows up on the invoice. There’s the cost of setting it up, which is the one that gets budgeted. There’s the cost of watching it, which is an hour someone stops spending on something else. And there’s the cost of fixing it when something changes (the model provider, a connected tool, an invoice format), which arrives unannounced and always at a bad time. In a large company a team absorbs those last two. In yours, you do.
With two or three agents running, the problem stops being technical and becomes one of management: who looks at what, when, and with what permissions. We’ve written about how to solve it with five controls and no platform team in AgentOps for small businesses. The short version: don’t open a second front until the first one runs without you having to think about it.
What should you leave out for now?
Anything that depends on your judgment and isn’t written down. How much discount you give a ten-year customer, who gets longer payment terms, when a complaint deserves an exception. It’s not that AI can’t touch it; it’s that you can’t ask it to decide something you couldn’t explain in a sentence yourself. If it doesn’t fit in a sentence, it isn’t a process yet, it’s a habit, and the first job is to write it down.
Don’t raise autonomy before you have evidence, either. The five-step ladder in levels of agent autonomy works just as well for a small company’s agent: start by observing and proposing, and move up only when weeks of reviewed cases justify it.
The line to take away: in a five-person company, the best agent isn’t the most capable one, it’s the one that fits the attention you have. Pick one, make it really work, and leave the other two for when the first one is running on its own.
Sources (accessed October 21, 2026): INE, Central Business Register (DIRCE) at January 1, 2025, press release of December 11, 2025; Eurostat, Use of artificial intelligence in enterprises, 2025 data, EU enterprises with 10 or more employees. The INE figures are for Spain; the Eurostat figures are for the EU. The reading about the scale of current offers is ours, not the sources’.