In automating finance workflows, measure time-to-resolve exceptions, completeness of source evidence, avoidance of duplicate events, follow-up messages per close item, reversals of accepted proposals, and the bandwidth spent explaining decisions to reviewers or auditors.

Replacing spreadsheets with dashboards helps, but the real gains come from optimizing the underlying workflow around the financial number rather than just moving tasks into a nicer interface.

Design should start with the exception path, defining what an operator needs when uncertain, and capturing exception records with fields such as event_id, amount, currency, evidence, reason, and state.

For each material decision, the finance workflow should document which source records were considered, which rule produced the proposal, who reviewed or changed it, and what happens if the same event arrives again, emphasizing idempotency and stable event identity.

Engineering gaps include handling retries, out-of-order webhooks, and transaction restatements; enforcing idempotency and maintaining an auditable event history prevents duplicates and confusion.

Keep automation proposals separate from accounting decisions by layering observed facts, proposed interpretations, and accepted decisions to enable model improvements without tampering with history.

The automation gap in finance is a workflow problem, not primarily a spreadsheet issue, because finance work hinges on handoffs, exceptions, approvals, and evidence checks that aren’t well captured by a single system.

Treat the exception queue as a core product surface, design retries from day one, and preserve the distinction between machine proposals and human-approved outcomes, focusing on provenance, ownership, and explainability of numbers.

A reviewer should be able to accept, edit, reject, or request more evidence, with every action adding an event to maintain an auditable history.

A routine supplier payment review shows how coordination across participants takes hours with minimal final accounting work, underscoring that automation rate alone is a weak metric.