Bank statement extraction and reconciliation is solid ground for an AI agent.

The extraction schema is fixed (date, amount, description, account number) and the deduplication rules are explicitly defined with numeric tolerances. This is structurally identical across all 120 files, which is ideal for automation.

Success criteria are concrete: a populated CSV, deduplicated cross-account transfers by defined rules, and a flagged list for manual review. There is no subjective judgment required to know when the task is complete.

The agent needs access to the 120 PDFs and a capable OCR pipeline; PDF quality and formatting inconsistencies across 20 accounts can degrade extraction accuracy. Assuming file access is granted and a reliable OCR tool is available, this is workable but not guaranteed to be clean.

Extraction errors or missed duplicates could cause downstream accounting mistakes, but the task explicitly routes suspicious entries to manual review, which acts as a meaningful safety net. The output is an input to human review, not a final financial record.

The matching and flagging logic is fully rule-based and requires no intuition or contextual business knowledge. A human is still needed to review flagged entries, but the agent's role requires no judgment calls.