AI Products
Kvitta
A Fortnox-approved AI bookkeeper for Swedish accounting — and the idea itself came out of an agentic research team I built to find businesses worth building.
Kvitta reads invoices from Fortnox or Visma and codes them to the right Swedish BAS account — an accounting agent trained on the 2026 BAS rules, with a human reviewing before anything is pushed back. It's the narrow, high-volume, unforgiving kind of task where reliable accuracy and where the data lives matter more than cleverness. Fortnox — Sweden's leading cloud-accounting platform — reviewed and approved it; I could sell it as a plugin on their marketplace tomorrow. It's fully working and live; I've simply not opened public registration while I decide whether to run it standalone or let Heywire absorb it. Built model-swappable, EU-data-residency-first.
Visit Kvitta- Role
- Builder (solo)
- Type
- AI bookkeeping · BAS 2026
- Integration
- Fortnox + Visma
- Stage
- Live · Fortnox-approved
- ~94%
- correct on a 50-invoice internal benchmark
- Fortnox
- reviewed and approved — sellable as a plugin
- Live
- working today; public registration held by choice
The idea came out of an agentic research team I built to find businesses worth building.
Kvitta didn’t start as a hunch. I built a small team of agents with distinct roles, set to challenge and question each other until they surfaced lucrative, buildable business ideas with a clear path to market. An in-app accounting assistant for Fortnox and Visma — the two platforms Swedish bookkeeping runs on — was one of the ideas that came back. So I built it. Using AI to find the business, not just to power it, is its own kind of product judgment.
Connect your ledger, an agent does the books, you review, then push back.
Connect Fortnox or Visma (or both). When invoices land, one button hands them to an accounting agent trained on the Swedish BAS rules for 2026; it fetches each invoice, does the bookkeeping, and you review before anything is pushed back to the platform. Connect your email and it scans the inbox for new invoices; photograph a receipt and it codes that too — all through the same flow. It’s the narrow, unforgiving kind of task — get it wrong and the books are wrong — where reliable accuracy beats a flashy interface, which is exactly what makes it a clean AI deployment target.
Model-swappable and EU-data-residency-first, on purpose.
Swedish financial data shouldn’t have to leave the EU, and no deployment should be married to one vendor’s model. Kvitta is built so the model is a swappable component and data residency is a first-class constraint — the cleanest version, in everything I’ve built, of the deploy-the-right-model-where-the-data-lives story this kind of work is really about.
Fortnox approved it — I could sell it as a plugin tomorrow.
I took Kvitta through Fortnox, Sweden’s leading cloud-accounting platform, and they reviewed and approved it: I can sell it on their marketplace as a plugin whenever I choose. It’s fully working and live — the only reason public registration isn’t open is a business decision, not an engineering one (whether to run it standalone or fold it into Heywire). A third-party bar cleared, not a self-graded one.
Accuracy: ~94% correct on my own 50-invoice test set. That’s an internal benchmark — my data, my scoring — not an independent evaluation, and it’s reported that way on purpose. The Fortnox approval is the external stamp; this is the accuracy measure behind it.
Builder — solo, end to end: the agentic research that found the idea, the BAS-trained accounting agent, the Fortnox and Visma integrations, the model abstraction, and the benchmark. A deliberately small surface that carries the EU and model-swap story better than a big one would.
Next
Blackbox