The owner of a car service, a salon or a shop does work every day that nobody enjoys: answering reviews, writing the same thing on Instagram, turning Viber notes into a table, reading through a contract. The free versions of ChatGPT, Gemini or Claude can do part of this in minutes. But every one of these tasks has a trap, and you need to know it before you press “send”. Below are five tasks, and for each: what to give the AI, what you get, and what to check yourself. Plus a separate section on the data you must not paste into a chat.
This is not an article about whether “AI will replace the marketer”. We take everyday jobs you can try today without paying anything. We checked the facts about settings on October 27, 2026 against the help pages of Google, OpenAI, Anthropic and Meta; where there is no primary source, we say so plainly.
1. A draft reply to a review
What to give the AI. The text of the review, what really happened and what you are ready to do about it. Ask for “a short, calm reply, no stock phrases, no promises I did not make”. The draft will be ready in a minute.
The trap. AI loves to promise: a discount, “this will never happen again”, “we will definitely look into it”. If you are not going to do that, delete it. A customer who read a promise will mention it in public later.
What to check. The reply is published in the company’s name, so read it as your own. In Google’s help the steps are: business.google.com, the section for reading reviews, the reply icon next to the review, the text, the button that sends the reply. Google checks the reply against its content policies: usually within 10 minutes, sometimes longer. That help page says nothing about AI suggestions or AI replies. Our rule is simple: the draft comes from AI, the version is yours. Paste your own text, not the draft as it is.
2. What is wrong, from 50 reviews
What to give the AI. 30–50 reviews or complaints from chats and calls (remove names and phone numbers) and the request: “split these into 5–7 topics, how many times each one appears, two verbatim quotes for each”.
An example. The Ukrainian branch of Karcher, according to a WebPromoExperts blog post dated September 24, 2024, uploaded its call-center reviews to an ordinary AI chat and asked it to sort them by topic. It turned out that about 30 % of requests were about service: where to get it and how to bring the equipment in. After that the company added separate communication about service and a loyalty program. The post does not state the number of reviews; “30–50” and “5–7 topics” are our practice, not figures from the case. It is a 2024 source, so we take only the idea of the method from it, not the figures.
What to check. This is the shortest check of all: take two or three quotes the AI gave and find them with search (Ctrl+F) among your own reviews. Not verbatim: the AI paraphrased, and the counts cannot be trusted. Then recount the topics by hand on at least twenty reviews.
3. Answers to common questions
What to give the AI. Write down the 10 questions you hear every day: “how much does it cost”, “is there a slot on Saturday”, “what is the warranty”. Add the price list, schedule, address and terms. Ask for short answers in your tone, two or three sentences each.
The trap. AI may round a price, change a deadline or add “free diagnostics” on its own. Check every number against your price list.
Where to keep the result. Meta Business Suite has “Saved replies”: templates you paste into a conversation on Facebook or Instagram. It is not an auto-responder: according to practitioners’ descriptions, you open the conversation, paste the template and press “Send” yourself. We could not confirm the menu names in Meta’s help (the pages do not render), and the path to this section differs between desktop and the app, so check it in your own account.
4. A table from chaos
What to give the AI. Notes from Viber, notes on your phone, a photo of an invoice. Ask for a table with the columns “date, service, amount, status”. Copy the finished result into Google Sheets or Excel.
The trap. Do not ask AI to add up the amounts. It composes text rather than working like a calculator, and it can be wrong in the total just as confidently as in a row. Calculate with a formula in the table itself.
What to check. Compare the number of rows with the number of records. AI sometimes loses or duplicates rows. This is not a figure from a study but what we see in practice; we give no “in so many percent of cases”. One more thing: we did not check whether a long conversation or many photos fit into the free limit. Free-plan limits change often, so split the work into parts.
5. A contract in plain words
What to give the AI. A lease or supply contract without company details, surnames and account numbers. The request: “Explain in plain words the penalties, deadlines and termination. What is unfavorable to me here? What questions should I ask the other side?”. You will get a list of points worth asking about.
The trap. AI is not a lawyer. It may cite a law article that does not exist or has already changed. This is the same tendency to invent that we described in the article “How AI Hallucinates: What It Can Make Up About Your Business”.
What to check. Open every article it mentions on the official legislation site (for contracts under Ukrainian law, zakon.rada.gov.ua, the site of Ukraine’s parliament) and read the current version. And show a large contract to a lawyer anyway: but with a list of questions, not empty-handed. The consultation will take less time.
6. What not to paste into a chat and which settings to turn off
Do not paste passwords, card numbers, or customers’ surnames and phone numbers into a chat. Replace them with “Customer 1”, “Customer 2”. This is not overcaution: in the Gemini help Google says directly not to enter confidential data or information you would not want specialists to see. The reason is that some chats are reviewed by people for quality checks; such chats are stored for up to three years.
What the services say about training on your chats (help pages as of October 2026):
- ChatGPT. For personal Free, Plus and Pro accounts the “Improve the model for everyone” switch is on by default. The path in the web version: account menu, “Settings”, “Data controls”. If you turn it off, new conversations are not used for training but stay in your history; the setting applies to the whole account on all devices. “Temporary Chat” does not go into history and does not train the model, but may be kept for up to 30 days for safety.
- Gemini. Turn off the setting that saves activity history (we could not confirm its exact English name; in Google’s Ukrainian help it reads “save activity history”). Then future chats do not appear in history and are not used for training unless you send feedback. But they still sit in the account for 72 hours, which the service needs to work.
- Claude. An “Incognito chat” is not used to improve Claude, even when model improvement is turned on. For ordinary chats on the free, Pro and Max plans, training happens only if you allowed it in the privacy settings (the option is called “Model Improvement” there). Claude is available in Ukraine, except for certain occupied and frontline regions: the list is on Anthropic’s supported countries page.
Where sources disagree. Anthropic’s page names the Claude switch “Model Improvement”, while third-party reviews write “Help improve Claude”. So look for it in the privacy section and go by the meaning, not the words. One more thing: OpenAI’s business plans (Business, Enterprise) and the API do not train on customer data by default — but those are paid plans, and we are covering the free ones.
Which figures are missing here, and why
In the US, 58 % of small businesses say they use generative AI (U.S. Chamber of Commerce survey, 2025). Other American surveys give anything from 9 to 55 %, depending on what counts as “using”. We found no Ukrainian data. So we do not build on a popularity figure: what matters more to you than how many people use AI is what exactly you do with it and how you check it.
Summary
You can hand AI drafts, grouping and rewriting: a reply to a review, 10 ready answers, complaints sorted by topic, a table from notes, an explanation of a contract. You cannot hand it the final decision, the numbers and the laws: a person checks those. Every task leaves a step for you: delete the promise, find the quote, check the price, calculate the sum, open the law article. If that step takes a minute, the task is worth handing over.
We wrote about how to check what ChatGPT already says about your business in the article “What ChatGPT Says About Your Business: a 10-Minute Check”. The routine is handed to AI, and the advertising is better checked by people you trust.
Want a free audit of your Google Ads account? Send us a request — we will look at your account and tell you where your advertising wastes money.