How not to miss the train. AI for business owners and programmers in 2026
No course and no new subscription. One repetitive task, a stopwatch and one instruction are enough to start. For programmers, we show what an ordinary day with Claude Code looks like for us.
In this article
In short
AI for a small business owner starts with one repetitive task, counting the hours before and after, and checking one data setting in your chat app. According to a BGR Quantic report for Microsoft, 94.1 percent of IT decision makers in large Polish firms use generative AI, but only 11.5 percent of firms formally measure the return. For a programmer the biggest gain is an agent in the repository, a CLAUDE.md file and short sessions.
A business owner hears about AI on the radio, from the accountant and from a client who answered a quote with a suspiciously smooth email. It feels like the train is leaving and they are still on the platform. This article is about getting on this week, without a course and without buying new software. AI for small business owners starts with ordinary things.
The second half is for programmers. It describes what an ordinary day with Claude Code looks like for us.
Almost everyone uses it, few measure it
According to a report by BGR Quantic for Microsoft, 94.1 percent of people who make IT decisions in Polish firms with more than 100 employees use generative AI. Only 11.5 percent of firms formally check whether it pays back. The report was covered by telix.pl on 20 October 2026.
Those are large firms, not yours. The lesson still fits a small one. Using AI gives you nothing by itself. What counts is what shows up in hours. A small firm has an advantage here, because one person can count it in a notebook.
AI for business owners. A plan for this week
- Pick one task that comes back every week and is about writing or retyping. Answers to similar enquiries, for example, or a summary of a long email from a supplier.
- For one week, write down how long it takes today.
- Before you paste anything, check the setting in your chat app that decides whether your conversations can be used to train models. We described it in Client data in ChatGPT.
- Write one instruction for that task and use it every time. Who the firm is, what tone you answer in and what the AI must not do.
- After a month, measure the time again.
Before pasting, replace names, phone numbers and invoice numbers with "Client A" and "Invoice 1". The answer will be just as good, and your client's data stays off someone else's server.
If after a month the gain is close to zero, that is a result too. Not every task fits. It is better to learn that after one month with one task than after a year with several subscriptions.
Three mistakes we see most often
The first is one long conversation for everything. A quote in the morning, a spreadsheet after lunch, terms and conditions in the evening. The more text in a conversation, the worse the model gets at details, which is why the morning instruction stops working. One task, one conversation. More on this in why AI forgets in long conversations.
The second is not checking. AI writes in a confident tone even when it is wrong. Someone has to read the numbers, dates, names and clauses before they reach a client.
The third is private accounts. Everyone in the office has different settings and nobody knows where anything goes. A shared business account, or even one page of rules, sorts this out faster than another tool.
A programmer in 2026. How we work with Claude Code
For a programmer the biggest change is moving from a chat to an agent. In a chat you paste code in pieces and carry the fixes back by hand. An agent in the repository reads the files, runs the tests and proposes changes by itself, and a person decides what stays.
Every project of ours has a CLAUDE.md file. It holds the build and test commands and the things Claude got wrong a second time. The file is context, not a lock, so real bans go into permission rules. We wrote about this in the CLAUDE.md article.
A bigger change starts in plan mode. Claude reads the project and writes a plan, and changes code only after we approve it. A mistake shows up in the plan, not in finished code. Before every bigger task we make a commit.
Long logs and test output are read by a subagent in its own context, and only a few lines come back to the main conversation. We end a session before the window fills up. When /context shows about half, 50 to 60 percent, we start a new session and paste a short summary of decisions at the top. That is our habit. Anthropic does not give such a number.
A person reads every change before it goes to production. The agent speeds up writing code. The decision about what the client gets stays with us. How to start with the tool itself is in What is Claude Code.
What to do by Friday
The owner picks one task, starts a stopwatch, checks the setting and writes one instruction. The programmer adds a CLAUDE.md to one project and makes one change through plan mode. In a month both will know more than after ten articles about the future of AI.
If you would rather start with a conversation, book a free AI audit. For about 30 minutes we ask how your firm works, and then you get a written list of 3–5 places where AI can save time, with the tool for each and what to watch out for with client data. There is no obligation.
What would you like to change in your business?
Leave your number. We will talk about your process and see whether we can help. The first call is free.
The service this text is about
Automation and AI / OCRAutomation of document re-typing, OCR and checking rules, and your data stays with you.