How AI Categorization Works

GigBooks uses a multi-step process to categorize your transactions into CRA-compliant expense categories.

The categorization pipeline

Step 1: Known merchants

First, GigBooks checks if it already knows the merchant. This includes:

• Your personal corrections — if you've recategorized "STARBUCKS" as a client meal before, it remembers
• Global patterns — common merchants like "ADOBE", "OPENAI", "ESSO" have pre-set categories

Step 2: AI categorization

For merchants that aren't recognized, GigBooks sends the merchant name and amount (nothing else) to Anthropic AI. The AI assigns a CRA T2125 category and a confidence level:

• High confidence (>90%) — the AI is very sure. Example: "ADOBE CREATIVE" → Software Subscriptions
• Medium confidence (60-90%) — likely correct but worth reviewing. Example: "LCBO" → could be business entertainment or personal
• Low confidence (<60%) — needs your input. Example: "2876543 ONTARIO INC" → could be anything

Step 3: Your review

Transactions with medium or low confidence are flagged for review. You'll see Q&A cards asking things like:

• "Is this fuel purchase for a business vehicle?"
• "Was this restaurant visit a client meal?"

What makes it smart

• Merchant memory — every correction you make trains the system for YOUR transactions
• Context awareness — the AI understands Canadian business expense patterns
• Conservative approach — when unsure, it asks rather than guessing