How the AI decides, how to correct it, and how learning works.
For each transaction we send Claude (Anthropic's LLM) the vendor name, amount, date, memo, your business type, and any rules the AI has already learned from your past corrections. Claude returns a category and a confidence score from 0–100. We never send your name, account numbers, or any personally identifiable info.
It's how sure Claude is about the category. Green (≥90) means we auto-post the transaction without bothering you. Amber (60–89) means we'll ask you about it in tonight's 6 PM SMS. Red (<60) means we're basically guessing — those always go to SMS for confirmation.
Click the transaction → click the category badge → pick the correct category. The fix is instant. More importantly: the AI writes a learned rule for that vendor, so it never gets the same vendor wrong again. One correction = permanent fix.
Every time you override a category, we save a row to your private learned-rules table: vendor pattern + correct category + which transaction it came from. Next time that vendor (or a normalized variation) appears, the rule fires before we even call Claude. This is why your categorization accuracy goes up week-over-week.
No. Your rules are scoped to your account only. The AI doesn't use one user's corrections to influence another user's books. This is a deliberate design choice — your business is unique, and a "Stripe" charge means something different for an e-commerce store than for a SaaS company.
It's the percentage of your transactions this month that got auto-posted (no SMS needed). A score of 90 means 9 out of every 10 transactions cleared with zero involvement from you. New users typically start at 50–60% and reach 85–95% by month two.
Yes — under AI Learning in the sidebar. You'll see every rule, where it came from, and how many transactions it's applied to. You can edit or delete any rule from there, which is useful when you change suppliers or your business model shifts.
Both. Income detection is actually easier — most income comes from a small number of payment processors (Stripe, Square, PayPal, ACH from your client list). We auto-categorize income by source and let you split a deposit across multiple invoices if needed.
You can split any transaction into multiple line items: click the transaction → "Split" → enter the amounts and categories. Splits also get learned — if you always split your AT&T bill 70/30 between Business Phone and Personal, the AI will start doing it for you.
Pick the "Personal — Owner Draw" category. The AI learns from this just like any other correction. If you accidentally swiped a personal card on a business purchase, mark it as Owner Contribution and the AI will track the contra-account for your equity.
Email the founder. Real reply, within 24 hours, no ticket queue.
brayden@claritybooksai.com