What do business owners ask an AI that knows their books? The obvious answer would be tax questions: what can I deduct, when is the deadline, how much do I owe. The data tells a different story. People who have an AI with access to their books rarely ask it anything. They give it work.
This analysis is based on more than 50,000 messages from about 3,000 self-employed people and small businesses in Germany to Norman's AI, from January to 21 November 2026. Norman's AI is an AI bookkeeper: it reads receipts, books payments, writes invoices and prepares tax returns, in the app's chat or through Claude and ChatGPT. We did not read any message. Topics are counted with keyword rules inside the database, and every figure on this page stands for at least 40 businesses.
Three in four messages are tasks
Only about one text message in four is a question, meaning it ends with a question mark or starts with a question word such as "how", "what" or "when". The other three in four are instructions: "Book this as a business meal", "Create the invoice for March", "Match the receipts from October".
Then there are the receipts themselves. Almost one message in four (23 percent) is not text at all but a photo or PDF, usually a receipt or invoice for the AI to read and book. Business owners do not use their AI like a search engine for tax questions, but like a colleague you hand a receipt to.
What the AI does in response points the same way. Most often it looks up bookings, and among the changes three moves dominate: correcting or completing a booking, creating a new booking and linking a receipt to its payment. It is the small work of bookkeeping that nobody could take off people's hands before.
The most common topic is the bank account, not tax
Almost four in ten businesses talk to their AI about the bank and payments: a transfer that is not matched, a direct debit they do not recognize, a payment that should belong to an invoice. Next come booking itself (34 percent of businesses), receipts (30 percent) and the question about their own money: profit, revenue, "how much did I make this month" (29 percent).
The first real tax on the list is VAT, which almost three in ten businesses raise. It is the tax that comes back every month or quarter. Deducting expenses concerns just under one business in five. The big annual returns, income tax, trade tax and corporate tax, each come up for fewer than one in ten. Deadlines are rarely a topic either.
People who talk to their AI talk about running the business, not about tax law. Tax is the result of clean books, and most conversations are about the clean books.
More than half arrive after hours
53 percent of all messages to the AI arrive outside business hours, meaning before 8am, after 6pm or at the weekend. More than one in five (22 percent) fall on Saturday and Sunday alone.
This matches our analysis of after-hours bookkeeping: the self-employed do every second bookkeeping hour in the evening, at night or at the weekend. For the AI this means its most important working hours are the ones when no tax advisor picks up the phone.
VAT comes up more often before the 10th
The month has a rhythm, and conversations with the AI follow it. In the first ten days of the month almost one message in ten mentions VAT (9.5 percent); after that, only about one in fourteen (7.1 percent). Before the 10th, when the advance VAT return is due, the topic comes up a third more often.
This matches our analysis of after-hours bookkeeping, where the 10th of the month sets the pace for advance VAT returns. For business owners VAT is not an annual topic but a monthly one, and the questions come as the deadline gets closer.
What it means
For a long time the expectation of a bookkeeping AI was a better tax guide: an answer machine for questions you would otherwise ask a tax advisor. Usage says something else. Business owners do not want answers about their bookkeeping, they want the bookkeeping done: the receipt booked, the payment matched, the invoice written. They ask about tax once the books are right.
One limit of the data, stated plainly: people who use Norman have already chosen an AI for their bookkeeping. Whether business owners without that experience would use their AI in a similar way, our data cannot say.
This analysis will be updated quarterly at this address; the changelog at the bottom of the page records changes. All analyses in this section: Norman Report.
Methodology
The basis is more than 50,000 messages from about 3,000 businesses to Norman's AI in the app's chat, 1 January to 21 November 2026. Internal and test accounts are excluded. We did not read any message and exported no message text: all counts run inside the database, and only totals are returned.
Topics are detected with keyword rules in German, English and Russian (for example "Umsatzsteuer", "UStVA" or "VAT" for the VAT topic). A message can count for several topics. Topic shares refer to the businesses that wrote at least one text message, and count a business at most once per topic. Keyword rules do not catch every wording, so the shares are lower bounds. A message counts as a question if it ends with a question mark or starts with a question word. Business hours are Monday to Friday, 8am to 6pm, German time. The AI's actions count the tools it ran in response to a message.
The monthly rhythm compares the share of messages with VAT keywords in each part of the month. We do not publish absolute user numbers. No published figure is based on fewer than 40 businesses. Use through Claude, ChatGPT and other AI tools outside the app is not included in the topics, because those conversations sit with the respective providers. How we work with data in general is documented in Methodology & data basis.
About this data
Norman is an AI bookkeeper for self-employed people and small businesses in Germany; the anonymized usage data analyzed here comes from it. Charts and figures on this page may be used freely with the credit "Norman". Press inquiries and data requests: press@norman.finance.