AI for a small firm: where it helps, where it fails, how to start

AI tools are now cheap and easy enough for a firm of three to use this afternoon. Much of what is written about them is still hype or fear.

This article is for the owner of a trades firm, a clinic, a shop, an office or a small agency. By "AI" we mean the chat assistants you type to (ChatGPT, Claude, Gemini, Microsoft Copilot) and the AI features inside tools you already use. Every number comes from a source listed at the end, with its date. Where we could not read a price, we say so.

Where small firms stand today

Eurostat asks EU firms whether they use at least one of eight AI technologies, among them text analysis, speech recognition, text and image generation, and process automation. Firms with fewer than 10 people are not counted.

Share of firms with 10 or more staff using at least one AI technology (Eurostat). Survey wording can change between rounds, so read the rise as a direction, not an exact growth rate.

Size matters. In 2025, 15.6 % of Slovak firms with 10 to 49 people used AI, against 43.7 % of those with 250 or more. The OECD found the same gap across its member countries: 11.9 % of firms with 10 to 49 staff against 40 % of large ones (2024 or the latest year). In 2024, 7.2 % of construction firms and 7.8 % of accommodation and food firms used AI, against almost 45 % in the ICT sector.

So you are not late. And those who have started mostly use AI for small jobs: in an OECD survey, only 29 % of small and medium firms using generative AI said they used it in their core activities. When Eurostat asked why firms do not use AI, the reason named most often was lack of expertise (7.8 % of EU firms and 5.1 % of Slovak ones), ahead of high cost (4.2 % in both). Know-how is a bigger obstacle than money.

What AI does well today

A language model predicts likely text. That covers a great deal of office work. Given the material, it does three jobs well:

• Drafting. First versions of replies, descriptions, posts, offers and letters.
• Summarising and sorting. A long email thread, meeting notes, a pile of enquiries grouped by topic.
• Reshaping. Rough notes into a clean email, a price list into a table, Slovak into English.

Two studies measured the effect. In a 2023 MIT experiment, 444 college-educated professionals did writing tasks. Those who could use ChatGPT finished about 10 minutes (37 %) faster than the group without it, who needed 27 minutes on average, and graders scored their work higher. That version of the paper is a working paper, not peer reviewed.

In the second study, an AI assistant was rolled out to 5,179 customer support agents. They resolved 14 % more issues per hour on average. The gain was 34 % for novice and low-skilled agents and minimal for experienced ones.

Read the second result twice. The tool helped most where a person was new or less skilled. That is a reason to use it when someone joins, and a reason not to expect miracles from your best person. The OECD's survey adds that small firms using generative AI named better employee performance as the main benefit, ahead of cost savings.

Where it fails

It makes things up

A language model produces likely sentences. It does not look facts up. Google's guidance for site owners says it is "critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness".

In 2024 a Canadian tribunal ordered Air Canada to pay a passenger CAD 812.02, damages plus interest and fees, after the airline's website chatbot gave wrong information about a bereavement fare. The airline said it could not be held liable for what its chatbot told people. The tribunal called that "a remarkable submission" and wrote: "It makes no difference whether the information comes from a static page or a chatbot." (Moffatt v. Air Canada, 2024 BCCRT 149). The sum was small. The lesson is not: customers treat what your chatbot says as what your firm says.

It does not know your business

It does not know your prices, your hours or this week's exceptions unless you tell it. Ask a general assistant what you charge for a boiler service, and you get a plausible number that is not yours. The cure is a one-page facts sheet that you paste in.

What you paste goes somewhere

The 2024 Work Trend Index from Microsoft and LinkedIn (about 31,000 knowledge workers in 31 markets) found that 75 % use AI at work and that 78 % of those bring their own tools. Your staff may already be pasting customer emails into personal accounts.

Personal and business versions of one product often have different terms. Anthropic's page, dated 18 August 2026, says it does not use inputs or outputs from its commercial products to train models by default, and sends readers elsewhere for consumer accounts. Google's Workspace pricing page says of chat with Gemini that your company data is not used to train models. Read the equivalent page for any tool you pick. No passwords, card numbers, health details or ID numbers in a personal account, and move to a business plan once a task is routine.

It is weak at judgement and at sums

A model has no stake in your decision. It cannot know your cash position, your history with a customer or how a remark will land. It can also get sums wrong, so let a spreadsheet do the arithmetic and ask the AI for the formula.

Put it in one picture: AI takes the first draft, you keep the decision.

AI takes the first draft

You keep the decision

A reply drafted from your facts sheet

What you promise and what you charge

A pile of messages sorted and summarised

Reading the one that is a complaint

Notes turned into a clean email or table

Names, prices and dates checked before it goes

The odd row found in a long list

What to do about it

Eight places to use it

Start from the work, not the tool.

Customer replies and enquiries

The best first use. Most of what arrives is a repeat: prices, hours, availability, "do you cover my area?". Write a one-page facts sheet with services, prices, hours, area, payment terms and what you do not do. Paste it at the start of a chat, then the customer's message, and ask for a reply in your voice. Read it, fix it, send it. A starting instruction:

A chatbot on your website is different: it talks to customers before you read anything. Limit it to your facts sheet, say plainly that it is a bot, and show how to reach a person.

Writing and marketing

AI is good at first drafts and variations: five headlines, a service description from your bullet points, a Google Business post. Give it facts from real jobs. Never ask it to invent a testimonial, a case or a number.

Google's guidance says that using AI to generate many pages without adding value for users may violate its spam policy on scaled content abuse. AI content as such is not banned. Ten pages of generic text will not help you. One page with your real prices, area and photos might. More in our article on the Google profile.

The website

Making it. We use AI to build sites ourselves. Our builder reads the address of your current site and drafts a new one; its page promises a first draft within an hour, no obligation to try, with no payment card. A person then finishes design, speed, forms, domain and launch, as we explain in Your own draft in a builder, finished by a pro.

Answering visitors. A site search or chat that answers only from your own pages can save a phone call. Add it after the plain contact form works, not instead of it.

Being found. Google says "there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", and that "all existing SEO fundamentals continue to be worthwhile". So the work is the usual one: a correct Google profile, clear pages, a site that loads fast on a phone.

Bookkeeping and documents

AI is good at pulling the date, supplier, total and VAT out of invoices and at summarising a ten-page contract in five lines. It is early days: in 2025 only 4.6 % of EU firms with 10 or more staff, and 3.3 % of Slovak ones, used AI for accounting, controlling or finance. Treat extracted figures like data entered by a new junior colleague. Check each against the original before it goes into the books, and keep your accountant in the loop.

Sales and CRM

According to the OECD, marketing and sales is where small firms draw most on generative AI. In 2025 Eurostat counted 6.9 % of EU firms with 10 or more staff and 10.4 % in Czechia using AI for it; Slovakia was at about 5 %. In practice: a summary of a call, a follow-up drafted from your notes, and a question like "which quotes are more than a week old?". If you keep the ten-column sheet from our CRM article, paste a row and ask for the follow-up. The next-step date stays yours.

Scheduling

A booking page that shows real available slots does the main job without any model; see online booking or the phone. AI adds small things: suggesting a slot for "Thursday afternoon or Friday morning", softening a reminder, drafting a reschedule message. Do not put an AI between the customer and the calendar: a system that knows what is available beats a model that guesses.

Internal knowledge

Every firm has things only one person knows: how to quote a bathroom, what to tell a new patient. Write them down in a short, dated document, then let the assistant answer staff questions from it. The writing is the real work; the AI is a search box that talks, and it repeats an outdated rule as confidently as a current one. The OECD found that in each G7 country it surveyed, under 30 % of small and medium firms using generative AI said their staff take part in AI training, so an hour of shared practice is more than most offer.

Data and reporting

Export your sales or enquiries to a spreadsheet and ask which rows look odd and what changed since last month. It is good at that, and less good at arithmetic: let the spreadsheet compute and check a few totals by hand.

What it costs

Prices change often. These were read from the vendors' pages on 5 November 2026, in the currency each page showed.

For one person the tool costs about 20 dollars a month; a team of three on a team plan pays 60 on yearly billing, 75 on monthly. If you already pay for Google Workspace or Microsoft 365, ask first what AI your plan includes.

The costs that matter are not on the price page: an hour or two to write the facts sheet, the minutes spent checking every output, an hour of practice with your staff, and usage fees if you wire AI into other tools. A plan at no cost is fine for practice with material that is not sensitive.

A plan to start in a week

Five days, one task.

1. 01Day 1: choose one taskRepeated, made of text, easy to check
2. 02Day 2: write the facts sheetPrices, hours, area, what you never promise
3. 03Day 3: test on ten old casesCompare each draft with what you really sent
4. 04Day 4: use it for realYou read every output before it goes
5. 05Day 5: count and decideKeep it, change it or drop it

Day 1. Choose a task that repeats weekly, is made of text, and where a mistake is caught in a minute. Answering the five most common enquiries qualifies. Setting prices does not. Write down how long it takes now.

Day 2. Write the facts sheet and a few rules: tone, length, what to say when unsure, what you never promise. Choose the account: business if staff will use it.

Day 3. Take ten real enquiries from last month. Let the assistant draft each reply from your sheet and compare with what you sent. If facts are wrong in more than a couple of the ten, the sheet is the problem. Fix it and run them again.

Day 4. Use it on live enquiries. You read every draft in full before it goes. Note the time per reply and every correction.

Day 5. Count: time per reply with and without, corrections, any new complaints. Then keep it, change the instruction or drop it. If you keep it, add a second task next week, not before.

What not to automate

• Decisions about people. Hiring, firing, pay, shifts. AI can draft the advert. The decision is yours.
• Prices and exceptions. Quotes, discounts, refunds.
• Complaints and bad news. An upset customer needs a person who has read the message.
• Advice that carries risk. Medical, legal, tax or safety advice.
• Anything sent unread. Replies, posts, public answers to reviews.
• Anything invented. Testimonials, reviews, before-and-after photos. That is not automation, it is misleading people.

The test: if a mistake would cost you a customer or money, a person reads it first.

How to tell whether it pays

Be careful with the claim that firms using AI do better. The OECD notes that the link between AI use and productivity appears to be explained, to a considerable extent, by firms that are already more digital and more competitive being more likely to adopt AI in the first place. Measure your own.

Take three numbers for one task, before and after:

1. Minutes per item, including checking, over a week.
2. Corrections per ten items. How many drafts needed a fact fixed?
3. An outcome. Time to first reply, quotes sent per week, enquiries answered within a day.

Then the sum: minutes saved per item, times items per month, divided by 60, times the value of your hour, minus the cost of the tool. An invented example: a reply takes 10 minutes today and 6 with a draft and a check, and you answer 80 enquiries a month. Four minutes times 80 is 320 minutes, about 5.3 hours. At 30 euros an hour that is about 160 euros against a tool that costs about 20. The figures are made up. Use yours.

Count checking as work: if it takes as long as writing the reply yourself, the task is not a fit. Compare per person too: in the NBER study newcomers gained far more than experts. Repeat the test after a month.

Questions

Will AI replace my staff?

The evidence we have points to a helper. In the NBER study an AI assistant made support agents faster, most of all the newer ones. What was measured was how much each agent got done, not how few agents were needed. For a firm of ten or twenty, the more useful question is which dull tasks it can take off people.

How do I stop it making things up?

You cannot reduce it to zero. Give it the facts, tell it to say "I will check" when a fact is missing, ask it to name the fact it used, and check that. For anything that reaches a customer, a person reads it first.

Which tool should I pick?

The one that fits what you already use. If your mail and documents are in Google Workspace or Microsoft 365, start with the AI in that suite: it works on files you already have. If not, pick one assistant and learn it for a month. In our view a good facts sheet matters more than the choice between the big assistants.

My staff already use it on their own accounts. What now?

That is common. Do not ban it, or you will only stop hearing about it. Write a half-page rule: what may be pasted, what may not, which account to use, and that a person checks everything that leaves the firm. Then give your people a business account.

Where to start

Pick the one task, write the sheet and run the week.

If you want a draft of your website first, the builder carries no obligation to try. If you want something built into your site, such as a form that sorts enquiries before they reach you, write to us through the contact form and we will say honestly whether it is worth building.