Workflow Teardowns

Most broken AI projects fail at allocation, not model choice. The team put judgment in the model and data entry in email. Flip the split and the system gets easier to audit and cheaper to run.

Databases own truth

Records, statuses, approvals, audit logs, configuration. If it must be true tomorrow, it lives in a database with constraints. Not in a spreadsheet. Not in a chat export.

Code owns rules

Routing, validation, retries, permissions, scheduling. If the rule is if status is X and amount exceeds Y then route to Z, that is code with tests, not a paragraph in a system prompt.

AI owns interpretation

Reading messy documents, classifying intent, drafting text, suggesting matches. Useful, probabilistic, always bounded. Output is a proposal until code or a human accepts it.

Humans own accountability

Final release, policy exceptions, edge cases the model has never seen. The UI makes those actions explicit: one button, one signature, one log entry.

Example: sample receipt

Receipt arrives by email. Code stores the attachment and creates a pending row. AI extracts vendor, date, line items into JSON. Code validates totals and matches PO numbers. Exceptions go to a reviewer queue. Human approves or rejects. Code updates status and notifies AP. At no point does an agent decide to pay the invoice.