I spend my days helping companies ship software — and my evenings doing my own bookkeeping. So we built an AI to take it off my plate. Here's what running a Swiss SME's back office on AI agents actually taught me.
The tax nobody puts on the roadmap
Ah, the back office. The part of running a company nobody warns you about when you're busy being excited about the product.
I run The Mejlerö Company — fractional CTO work, mostly. My calendar is sprint reviews, architecture calls, and the occasional "why is production on fire" conversation. And then, somewhere around the 25th of the month, a different version of me shows up: the one squinting at PDF invoices, copying numbers into spreadsheets, matching bank lines, and trying to remember which account a Hetzner bill goes to under the Swiss Kontenrahmen KMU.
It's not hard work. That's the insulting part. It's just busywork — the kind that eats an evening and adds nothing. Every hour I spent on admin was an hour I wasn't building. So I did the thing CTOs do when a process annoys them enough: I treated it as an engineering problem.
Why off-the-shelf didn't cut it
The honest first move is to buy, not build. I tried. The problem is that a Swiss SME's back office isn't one problem — it's ten of them, and they all want their own subscription.
Accounting in one tool. Invoicing in another. A separate thing for time tracking. Another for the team chat. CRM over here, a password vault over there. Each one is fine. Together they're a tab graveyard, and none of them really understand Swiss specifics — double-entry under the OR, MWST at 8.1%, the Kontenrahmen KMU, quarterly VAT filing the way the ESTV actually wants it.
What I wanted didn't exist: something that could read an invoice the way my accountant does, book it correctly, and only bother me when it genuinely wasn't sure. So we built it. We called it Flitz.
A form with a database is not an AI product
Here's where it got interesting as a technical problem.
The lazy version of "AI accounting" is a form with a chatbot bolted on. The useful version is a pipeline of agents, each with a narrow job, handing work to the next one — and a human only in the loop when it counts:
Parse — an agent reads the invoice (PDF, photo, scan) and pulls out structured data: vendor, amounts, VAT, IBAN, QR reference.
Book — a second agent turns that into a balanced double-entry journal entry using the right accounts.
Audit — a third agent checks the first two, structurally and semantically, before anything touches the books.
Review — anything below a confidence threshold gets flagged for a human instead of guessed at.
That structure — small specialised agents, not one giant prompt — is the part I'd tell any CTO to steal, whatever you're building.
What building it taught me about AI agents
A few lessons landed hard enough that I now bring them into client work:
Treat an AI agent like a talented, overconfident junior. It will produce great work and the occasional confidently-wrong disaster. The engineering isn't the model — it's the review layer around it. We auto-book above 85% confidence and escalate below it. That single number does more for trust than any amount of prompt-tuning.
Guardrails are the product. Retries at the API layer, retries at the queue layer, and a full audit log of every decision. When an AI touches money, "it usually works" is not a feature. Being able to answer "why did it do that?" six months later is.
Boring infrastructure wins. Queues, idempotency, transactions that actually roll back, structured logging. The glamorous part is the AI; the part that makes it trustworthy is the plumbing every senior engineer already knows how to build.
Ship the human-in-the-loop, then earn the autonomy. We didn't start by automating everything. We automated the confident 90%, surfaced the rest, and let the system earn trust one corner at a time.
None of that is unique to accounting. It's how I'd approach any agentic system that has to be right, not just impressive in a demo.
Where it's at
Flitz quietly runs the back office now. Invoices come in by email, get read, booked, and audited. Bank statements reconcile themselves. VAT reports come out in the format the tax office expects. And it grew past accounting into the rest of the tab graveyard — chat, time tracking, CRM, a password vault, and more — all in one place, built to Swiss standards for Swiss SMEs.
The funny part? The most valuable thing I got out of it wasn't the saved evenings (though those are nice). It was a working, opinionated answer to a question every founder and CTO is asking right now: what does it actually take to put an AI agent in charge of something that matters?
Turns out it takes exactly the discipline good engineering always took. The AI just raised the stakes.
If you want to see it, it's at flitz.ai. And if you're a CTO staring at your own version of the 25th-of-the-month problem — whether it's accounting or anything else you suspect an agent could own — that's the kind of thing I help companies figure out.
Switzerland.