Entrepreneurs keep asking me the same question: “OK, I get that AI matters, but where do I actually start in a company of ten people?” I have answered it in pieces many times, so today I am putting it all together as a case study.
Let me be clear from the start: this is a composite case, built from situations I see again and again in the small companies I work with. It is not a specific company and I am not giving any numbers. Let’s say it is an online shop with about ten people: two in customer service, one on content and marketing, one on accounting and invoices, a few in the warehouse and an owner who does a bit of everything. If you recognise yourself, it is because this story repeats almost word for word from one company to the next.
1. Start with a map: where does the time actually go?
The classic mistake is to start with the tool. “Let’s get everyone a ChatGPT subscription and see what happens.” Usually, not much happens. I start with a week of short conversations with everyone on the team: what do you do every day, what do you do ten times a day, what drives you mad, what would you put off forever if you could.
In our imaginary company, the map looks roughly like this: repetitive customer emails (“where is my order?”, “how do I return this?”), product descriptions written from scratch for every new item, supplier invoices typed by hand into the accounting software, and a weekly report the owner asks for, which someone puts together on Monday morning by copying figures from three different places.
2. Pick two or three processes, not ten
From that list I pick the processes with high volume and low risk. High volume, so the time saved actually matters. Low risk, so a mistake does not cost you a customer or a fine. I have written in more detail about what a small business can automate, but here I would start with three things:
- answers to frequently asked questions and to “where is my order”, using data pulled straight from the shop platform;
- product descriptions generated from a spec sheet, then edited by a person;
- extracting data from invoices and preparing it for the accountant.
The weekly report comes in phase two. Complicated complaints, supplier negotiations and anything involving refunds stay with people.
3. The tools: simple, and connected to what you already use
A small company does not need a custom-built AI platform. It needs three layers: a general assistant (ChatGPT, Claude, Gemini or whatever the team prefers), an automation platform that connects your apps (Zapier, Make, or n8n if you have someone technical) and integrations with what you already run: email, helpdesk, the shop platform, the invoicing software.
The models have changed a lot this year alone: OpenAI released GPT-5 in August, and Anthropic launched Claude Sonnet 4.5 at the end of September. For a small company, this matters less than it seems. The difference is not the model; it is how well it is connected to your data and how clear the rules are.
4. The pilot: AI suggests, a human decides
For the first few weeks, nothing goes out to a customer without passing through a human. The AI drafts the reply, the customer service colleague reads it, fixes it and sends it. Same for descriptions. For invoices, the extracted data is checked before it goes into the books.
The pilot has two purposes. First, you see where the system gets things wrong before it gets them wrong in front of customers. Second, people learn to work with it without pressure. Write down everything that had to be corrected. That list is worth more than any demo.
5. The rules: what it does alone and what it doesn’t
After the pilot, write one page on what the AI may do on its own and what needs approval. For example: it can answer order status questions by itself, as long as the information comes from the system; it cannot promise refunds, discounts or delivery dates; it does not answer complaints, it prepares them for a person.
Then there is data. GDPR (the EU data protection regulation) did not go away because AI showed up. You do not put customer data into tools that lack a proper data processing agreement and business settings that exclude it from model training. Employees’ free personal accounts are no place for customer addresses and orders. And there is a detail few people know: since 2 February 2025, the EU AI Act requires companies that use AI to make sure their staff have a sufficient level of AI literacy. I explained briefly what the AI Act means for a company; for a small shop, this is the most concrete part.
6. The people: training and fear
The fear is real, and it is worth naming. The customer service colleague will naturally wonder whether you are automating their job. If you do not talk about it openly, you will see quiet resistance: “it doesn’t work”, “it got it wrong again”, “it’s faster if I write it myself”.
What works: involve them in choosing the processes, show them the goal is to get rid of the boring part, run hands-on training on their real tasks rather than a generic course, and appoint someone on the team as the “AI person” the others can go to with questions.
7. What to measure
I will not give you results, because they would be made up. I will tell you what to measure, before and after, so you have your own numbers:
- time per task: how long a reply, a description or an invoice entry takes;
- response time to customers, from receiving the message to replying;
- errors: corrected replies, wrong invoice data, rejected descriptions;
- customer satisfaction, at least through a short question after the conversation;
- how much of the time saved actually goes into higher-value work.
That last point is the one most often ignored. I wrote about it when I shared what AI taught me about productivity: the time you gain evaporates unless you decide in advance what to use it for.
8. What usually goes wrong
- Bad data. If product sheets are incomplete or order statuses are not kept up to date, the AI will produce wrong information, very convincingly.
- No owner. The project belongs to “everyone”, so to no one. Two months later the automations break and nobody notices.
- Too much automation, too soon. First-week enthusiasm makes you automate things you shouldn’t.
- Made-up answers to customers. A model that does not know the answer sometimes guesses. An invented return policy is a real problem.
- Tone. Replies that are correct but cold or overly formal, and don’t sound like your company. You fix it with examples and style rules, not with hope.
9. The next step: agents, carefully
Once the first processes run smoothly, the temptation is agents: systems that carry out multi-step tasks on their own, from reading an email to checking stock and preparing a quote. The technology is moving fast; in July OpenAI launched “ChatGPT agent”, which can browse and take actions on the user’s behalf. I have written about AI agents as the next step after chatbots, and I stand by the same idea: start them on internal tasks, with limited permissions and a human who approves irreversible actions. Do not give them access to money or customers on day one.
My conclusion
In a small company, AI rarely replaces people. It takes the repetitive part off their plate and gives them time for what matters: an unhappy customer handled with care, a new supplier, a product idea. But that only happens if you introduce it well: with a map of your processes, a pilot, written rules, your people on your side and someone who owns it. I wrote more about what I see in the market in AI and jobs.
Which process did you automate first in your company, and what turned out differently than you expected?