Why did AI categorization sometimes fail?
Automatic transaction categorization relies on an AI model that analyzes bank operation descriptions. The problem? AI model providers have rate limits and outages. When the model is overloaded — it returns a 429 error and your transactions wait without categories.
This is frustrating because you see new transactions in the app, but without assigned categories. Spending charts don't show the full picture, and your household budget looks incomplete. One unavailable model blocked the entire categorization process.
Three models instead of one
Martia now uses a chain of three AI models: Gemini 3 Pro, Gemini 3 Flash, and Grok. If the first is overloaded — the system automatically switches to the next. No pause, no waiting. Additionally, transactions are processed in three batches simultaneously instead of one by one.
The result? Categorizing 200 transactions, which used to take several minutes and could stall, now finishes faster and doesn't fail even when one provider has issues. The progress bar shows how many transactions have been categorized — in real time. As of March 2026.
Domande frequenti
All three models deliver comparable quality for Polish transactions. Gemini 3 Pro is the default, but Flash and Grok categorize equally well — we tested on thousands of transactions from Polish banks.
No. The change works automatically for all users. Your next bank sync will already use the new three-model system.
Inizia a seguire le tue finanze con Martia
Collega il conto bancario e monitora automaticamente le spese, pianifica i budget e analizza le tendenze — tutto in un unico posto.
Inizia ora