I want to build products, not do admin

Admin was never the problem

I run a small Swiss software company. The work I like is building: designing systems, writing code, shipping product.

Admin comes with running a company, and I never expected otherwise. It has to be done, and it has to be done correctly. What annoyed me was doing it the dumb way: invoices scattered across an email inbox, a spreadsheet with rules only I knew, a bank statement in one window, an invoice PDF in another, and me retyping numbers between them by hand.

The thing that got to me was not the hours. It was that I could see, precisely, that a machine should be doing it.

It was never hard. It was just endless.

Bookkeeping for a small company is not intellectually difficult. It is a long chain of small, obvious decisions. This is a supplier invoice. It belongs in this account under the Kontenrahmen KMU. The VAT on it is 8.1%. It matches this line on the PostFinance statement. Book it, file it, move on.

Each individual link in that chain takes ten seconds and a person who knows what they are doing. There are just thousands of them a year, and no software I could buy would take the whole chain, only individual links, handing the rest back to me.

I tried buying the answer first

I want to be clear that building this was not the first idea. It was roughly the fifth.

I tried the Swiss accounting packages. They are decent at being ledgers and they were never designed to read a crumpled receipt and decide what it is. I tried a Treuhänder, which works beautifully and costs real money and still needs me to hand over an organised pile. I tried the international all-in-ones, which understand accounting in general and Swiss accounting not at all. And in accounting, 80% right is not 80% useful, it is a correction job.

And underneath all of it, the thing that genuinely annoyed me: every one of those little systems was its own subscription. Accounting in one tool, time tracking in another, the CRM in a third, tickets in a fourth, passwords in a text file I am not proud of, each with its own invoice, its own login, its own idea of what a customer is. Eight people should not need eleven subscriptions to run one small company.

The week it became buildable

What changed is narrower than “AI got good”. Two specific things crossed a line, recently.

First, vision models got reliable enough to read a real receipt (folded, thermal-printed, photographed badly on a phone) and not just transcribe it but understand what it is. Second, the cost of that reasoning fell far enough that spending it on a CHF 14.80 receipt stopped being absurd.

Those two facts together are the whole company. Before them, a product like this would have been a demo. After them, it is an engineering problem, and engineering problems I know how to work through.

So I built it for myself first

The first version had one job: read the invoices in my inbox and propose the booking. When it started getting them right more often than I did, I knew what I was looking at.

One rule stood from day one: do not change how accounting works, automate how it is administered. Underneath Flitz there is a real double-entry ledger on the Kontenrahmen KMU, real fiscal periods, real MWST rules. The AI works on top of that base system: it reads, proposes and books, and the framework decides what correct means. This is not asking an AI for something plausible; a booking is either right or it is wrong.

Everything since has come from the same place: something I actually needed in a real week of running the company. Bank reconciliation, because matching statements by eye is where the mistakes happen. VAT reporting, because the quarterly MWST return is a hard deadline. Time tracking, then projects, then the CRM, then the intranet, then the password vault, each one replacing another subscription on the pile.

Why Switzerland first, and why that is not a limitation

Swiss accounting is specific. OR Art. 957ff, the Kontenrahmen KMU, MWST rates, cantonal variation, payroll with AHV and ALV and BVG. A generic global product will always be approximately right here, and approximately right is the one thing accounting cannot be.

That specificity is not a constraint on the business. It is the moat. A market small enough that the giants will not do the work properly, demanding enough that doing it properly is genuinely valued, and disciplined enough that if a product survives here it can survive anywhere in Europe. Germany and Austria have the same shape of problem and more than ten times the businesses.

Why the whole desk, not one drawer

The received wisdom is to build one tool and do it well. I think that advice was written for companies selling to companies big enough to have departments.

A Swiss KMU with eight people does not have an accounting problem. It has an admin problem, and accounting is the loudest part of it. Splitting that into nine products with nine bills and nine integrations is how the industry made its own life easier, not the customer’s. So Flitz is one system: your invoices, your bank, your hours, your team, your customers, your documents, your compliance, in one place, where the AI can actually see across all of it.

Where the AI has to live

Every Swiss business owner I have shown this to asks the same question within about four minutes: where does my data go?

That question deserves a place as an answer, not a policy document. Which is why the next thing we are building is our own AI, running on Swiss infrastructure, so that the answer is “here”. Not a clause in a data-processing agreement, but a location, in Switzerland, that I can name.

What I want this to be

I would like Flitz to be the last piece of software a Swiss KMU has to think about. Not the most exciting one. The one that quietly does the work that nobody went into business to do.

And selfishly: the less time this company spends on admin, the more of it I get to spend building.