AI agents for a small business in 2026: what works and what's a demo
Chatbot, automation or AI agent? I rate eight everyday jobs in a small business: what AI does well today, what needs a person to check it, what is still a demo.
Nikolay Tolmachev, Founder, Tolman Studios
7 min readAI & Automation
In short
15.6%of small firms in Slovakia (10-49 staff) used AI in 2025. EU average: 17% (Eurostat)
- A chatbot answers questions. An automation follows the fixed steps you set. An agent picks the next step on its own. Most of what works in a small business today is the first two, with a bit of AI inside.
- Answering common questions, booking and reminders work today. Follow-ups, invoices and review replies work if a person checks them. Answering the phone and agents that run with nobody watching are still mostly demos.
- In 2025, 20% of EU firms with 10 or more staff used AI, and 18% in Slovakia (Eurostat). Start with one boring task, keep a person approving, and tell people they are talking to an AI.
On this page
Short answer: in 2026 an AI agent can already answer your customers' common questions, book appointments and send reminders. It can draft follow-ups, review replies and invoice entries, but a person should check them. An agent that runs your phone line or handles a customer from first message to payment, with nobody watching, is still mostly a demo.
The word "agent" is everywhere this year. Meta calls its WhatsApp assistant an agent. Zapier and Make sell agents, and n8n has an agent node. So let me start with what the word means for you.
Chatbot, automation or agent: what is the difference?
Take a hair salon with 40 messages a day.
A chatbot answers. "Are you open on Saturday?" "Yes, 9:00 to 14:00." It reads your price list and FAQ and replies in the chat. It doesn't do anything else.
An automation follows steps you set. When a booking comes in, send a confirmation. The day before, send a reminder. If the client cancels, offer the slot to the waiting list. Every step is fixed in advance. No AI is needed for this, and that's why it rarely breaks.
An agent picks the next step itself. A client writes "can I move my colour to Thursday afternoon, and is Jana free?" The agent understands the request, checks the calendar, picks a slot, moves the booking and replies. Nobody wrote that exact path. It decided.
That freedom is the useful part and the risky part. An automation does the same thing every time. An agent can do the right thing on Monday and something odd on Tuesday.
Which jobs can AI do in a small business today?
Here is how I rate eight everyday jobs, as of September 2026. "Works today" means I would let it run on its own after a test week. "With supervision" means AI prepares and a person approves. "Not yet" means the demos look better than real life.
| Job | Status | Typical tools | What to watch |
|---|---|---|---|
| Answering common questions (hours, prices, address) | Works today | A chatbot on your website, Meta Business Agent on WhatsApp and Messenger | Let it answer only from content you approved |
| Booking appointments | Works today | Reservio, Bookio, Fresha, Calendly. A chatbot connected to them | The booking tool does the real work. AI only helps the client find a time |
| Reminders and no-shows | Works today | The reminders in your booking tool, or Make and Zapier | Plain automation. No AI needed |
| Follow-up after an enquiry | With supervision | Make, Zapier or n8n with an AI step, your CRM or a Google Sheet | AI drafts the message. You approve it for the first weeks |
| Invoices and receipts to your accountant | With supervision | Make, Zapier or n8n reading PDFs from your inbox, your invoicing tool | Check amounts and VAT. A wrong number ends up in your tax return |
| Replies to Google reviews | With supervision | ChatGPT or Gemini drafts, you post | Never let it reply alone to an angry review |
| Social media post drafts | With supervision | ChatGPT, Gemini, Canva | It writes generic text unless you give it real details. Realistic AI photos of real people need a label |
| Answering the phone | Not yet | AI voice agents | Accents, noise and callers who change their mind. Test with real calls in your language |
Where the task has a clear right answer and your data is tidy, AI does well. Where a mistake costs money or a customer, AI prepares and you decide.
One thing is missing from the table: the "do everything" agent. It takes a new enquiry, sends a quote, books, invoices and asks for payment on its own. It looks great in launch videos. I wouldn't trust it with a real small business yet.
How many small businesses already use AI?
Fewer than the headlines suggest. Eurostat asks firms with 10 or more employees every year whether they use any AI tool. In 2025, 20% of them in the EU said yes, up from 13.5% a year earlier. Slovakia went from 10.8% to 18%.
Small firms, with 10 to 49 people, are further behind.
- Denmarkhighest in the EU37.5%
- Austria26.2%
- Germany23.1%
- EU average17%
- Slovakia15.6%
- Italy14.2%
- Czechia13.4%
- Poland5.8%
- Romanialowest in the EU4.1%
Source: Eurostat, isoc_eb_ai, data extracted December 2025
Two notes. One-person firms and micro businesses with under 10 people are not in this survey at all, and that is where most of my readers are. Also, "uses AI" includes someone in the office who writes text with ChatGPT. The share of firms that run real agents is much smaller.
Eurostat also asked a question to the firms that thought about AI and decided against it: why not?
- said they lack the know-how
- 71%
- Eurostat, EU, 2025
- were unsure about the legal consequences
- 53%
- Eurostat, EU, 2025
- worried about data protection and privacy
- 49%
- Eurostat, EU, 2025
These three worries are fair. The rest of this post is about them.
What can go wrong?
It makes things up. An AI that can't find the answer may still give one. In February 2024 a tribunal in British Columbia ruled on a case against Air Canada. Its website chatbot had given a customer a wrong refund rule, and the airline had to stand behind it. The airline argued that the bot was responsible for its own words. The tribunal didn't accept that. Treat what your bot says like what your staff say.
It touches personal data. Names, phone numbers, sometimes health details. Under GDPR you need a data processing agreement with each tool and a line in your privacy notice. Also check where the data goes. Make lets you choose an EU data zone when you sign up, but its own help page says some AI features still run through the US. n8n says its cloud stores data in Frankfurt. You can also run n8n on your own server.
It has to say it is an AI. Since 2 August 2026 the EU AI Act (Article 50) requires that people know they are dealing with an AI system, unless that is obvious. I covered this in what the AI Act asks of your chatbot. In practice it is one line in the first message.
It can cost more than the advertised price. Zapier bills its Agents as a separate add-on, on top of the normal plan. Meta's Business Agent is free to start, and Meta has said paid plans will follow. I explained what that means in my post on Meta Business Agent. Plan as if every free AI tool gets a price one day.
Where would I start?
The chat on my own website runs on Google's Gemini and answers only from my site's content. Writing good answers took longer than connecting the model. That is usually the case, and it is why I start small with clients.
- Pick one task that comes up at least 10 times a week. Booking questions and reminders are good first candidates.
- Write down how you do it now, step by step. If you can't describe it, a machine can't do it either.
- Build the plain automation first. Add AI only to the step that has to read or write text.
- Let AI draft and a person approve for the first two to four weeks. Note every mistake.
- Add the AI line to the first message, and a way to reach a person.
- Sign the data processing agreements and update your privacy notice.
- After a month, decide: keep it, fix it, or switch it off.
You may not need an agent at all. If you get five messages a day and answer them over coffee, a better FAQ page and a booking link do more for less. My post on where AI chatbots help and where they don't has a quick self-test for that.
If one task already takes up your evenings, here is what I set up in AI and automation. We would start with that one task.
Questions people ask
What is the difference between an AI agent and a chatbot?
A chatbot answers questions in a chat. An agent can also act. It can look up a free slot, write to your calendar or send an email, and it decides on its own which of these to do next. Many tools sold as agents today are chatbots with one or two actions added.
Can an AI agent answer my business phone?
AI voice agents exist, and the demos sound good. With Slovak or German, background noise and callers who change their mind, I would not let one take calls alone yet. Test it on 20 real calls first and keep a quick way to reach a person.
Which tool should I start with: Make, Zapier or n8n?
All three connect your apps and can add an AI step. Zapier is the easiest to start with. Make is cheaper for many steps and lets you pick an EU data zone when you sign up. n8n can run on your own server. Pick the one that already connects to your booking and accounting tools.
Am I responsible if the AI tells a customer something wrong?
Most likely yes. In 2024 a Canadian tribunal held Air Canada responsible for wrong refund information from its website chatbot. I'm not a lawyer, but I would treat what your bot says like what your staff say.
Do I have to tell customers they are talking to an AI?
Yes. Since 2 August 2026 the EU AI Act requires that people are told they are dealing with an AI system, unless that is obvious. One line in the first message does it.
Sources
- Use of artificial intelligence in enterprises (data extracted December 2025), Eurostat, checked 8 September 2026
- Artificial intelligence by size class of enterprise (dataset isoc_eb_ai), Eurostat, checked 8 September 2026
- Introducing Meta Business Agent, Meta / WhatsApp Business, checked 8 September 2026
- Regulation (EU) 2024/1689 (AI Act), Article 50, EUR-Lex, checked 8 September 2026
- AI providers and data zones in Make, Make, checked 8 September 2026
- Zapier pricing (Agents are billed as a separate add-on), Zapier, checked 8 September 2026
- n8n pricing (cloud data stored in Frankfurt), n8n, checked 8 September 2026
- Moffatt v. Air Canada, 2024 BCCRT 149, Civil Resolution Tribunal, British Columbia, 14 February 2024
Facts and prices are as of the dates in the sources.
Nikolay Tolmachev
Founder, Tolman Studios
I build websites, AI chatbots and ads for small businesses, working from Slovakia. Before that I spent over 4 years managing IT projects. I hold PMP and Certified ScrumMaster certificates.