Multi-agent Bank Statement Intake: Raw Transactions to Categorized Financial Report

Trigger condition

Steps

1. 1

   Parse raw transaction lines into structured entries: extract date (YYYY-MM-DD), description, and amount per line. Handle negative amounts (withdrawals) and positive (deposits). Flag unparseable lines for user clarification

   tool: code_execution · output: Array of {date, description, amount, type: income|spending}

2. 2

   Categorize each transaction using multi-pass classification: known merchant/pattern map, keyword matching, amount heuristics, and LLM classification for ambiguous entries. Categories: eating_out, food, transport, shopping, subscriptions, debt, errands, income, misc

   tool: llm · output: Array of {date, description, amount, category} with confidence level

3. 3

   Compute aggregate metrics: total income, spending per category, net balance change. Identify patterns (recurring charges, temporal clusters, category concentration) and anomalies (unusual amounts, new merchants, category spikes exceeding 2 standard deviations from mean)

   tool: code_execution · output: {summary, categories, patterns: [], anomalies: []}

4. 4

   Synthesize into structured report: summary table, category breakdown, flagged patterns with explanations, anomaly alerts, and prioritized actionable recommendations. Include month-over-month comparison when historical data exists

   tool: llm · output: Human-readable financial analysis with JSON backing data

Expected output

Tools required