How Finn works
The model layer
Which AI models power Finn, and how the right one is picked for each job.
Finn isn't tied to one AI model or one AI company. She uses the best model for each job, and changes model when a better one arrives.
The models behind Finn
- Conversation and reasoning: Claude Sonnet 5.5 runs Finn's chat by default.
- Enterprise tier: Claude Opus 5.5, with deeper reasoning, for the most complex books.
- Quick replies: a small, fast model handles simple openers like "hi" and "thanks", and hands anything real to the full model.
- Reading documents: a fast vision model reads receipts, statements and exports at volume.
- Live voice: a real-time voice model for spoken conversations and video calls.
- Voice notes: text-to-speech for Finn's replies on WhatsApp and Telegram.
Model-agnostic by design
- One switch. Every model is named in one place, so moving to a new model is a one-line change.
- Failover across providers. Anthropic, Google and OpenAI are all wired in. If one provider has an outage, the work moves to another without you noticing.
- The agent survives an outage. If the main chat model is down, a second provider runs the same agent, with the same tools and the same approval steps.
- Cost tracked per company. Every model call is metered, so model choices are made on evidence, not guesswork.
Evals: how a new model earns its place
Before a new model goes live, it runs through Finn's own evaluation suites:
- Behaviour evals: real accounting conversations, scored on whether Finn did the right thing.
- Trust evals: does she stay honest about figures, never claim work she didn't do, and ask before acting?
- Document evals: real and synthetic bank statements and exports, checked line by line.
- Migration evals: exports from other accounting software, mapped onto the books.
A model that wins on the benchmark but loses on the evals doesn't ship.