At a fourteen-person software company in Stockholm, the free card plan ends when the sixth employee needs a card. The next invoice brings per-user fees and 2.45% on every foreign-currency purchase, and a software company buys mostly software, in USD and EUR. On 1.5 million kronor a year, that is roughly 37,000 kr.
Under the Mittkurs web app, on cloud infrastructure in Stockholm, sits a ledger of balances, one per currency a company holds. A charge is paid from the balance in its own currency, so a USD subscription never converts at checkout. Money converts only at top-up, at the European Central Bank rate plus 0.4%. At month-end, coded receipts go out as one journal per currency.
Every card charge needs a document behind it, and reading that document is the job Mittkurs gives its models. A charge reads like GITHUB.COM; its receipt may be a photographed taxi slip. Each document goes through two models: one finds the fields, another reads them with a confidence score. A lookup table and the company’s coding history code the recurring half of charges with no model; a classifier suggests codes for the rest.
A receipt field the Mittkurs reader is less than 0.85 sure of stays blank for the person closing the month. An account code counts only once the finance lead accepts or changes it; nothing is posted until the close is confirmed. Each confirmation leaves a pair: a crop of the document line, identifiers removed, and the confirmed code. We fine-tune on the pairs pooled across companies, because a GitHub invoice looks the same in Helsinki as in Malmö.
GPU demand at Mittkurs peaks in the eight working days around each month-end close, when most receipts arrive, every page read on a GPU. Onboarding adds a second burst: a new company’s twelve months of statements and receipts, coded in one run. Our card issuer and card network agreements forbid sending card data to a model API outside our control, so the models are served and fine-tuned on cloud GPU capacity we control in Stockholm.
Mittkurs reads receipts in three steps today: text recognition, a field-finding model and a field-reading model. In Q1 2027 the three steps merge into a single vision-language model with open weights, fine-tuned on the confirmed pairs. In Q2, embedding search gives each new document the code of the nearest merchant among its company’s confirmed lines. By Q3 2027 month-end bursts run on reserved GPU capacity in Stockholm.
Mittkurs Oy is a Finnish osakeyhtiö in Helsinki, founded in July 2025, with its cards issued through an e-money institution licensed in Finland. The first accounts opened in September 2026; no company has yet closed a full year on them.