Our AI approved the invoice. But nobody could tell us why.

Описание и применение

AI‑агент автоматически одобрял счета, но не записывал причины решений, что затрудняло аудит. После изменений каждая сделка фиксируется с деталями, а проверка ограничений вынесена в отдельный сервис.

Задача

The AI agent approved supplier invoices automatically but did not record the reasoning behind each decision, leaving finance unable to explain why an invoice was approved.

Результат

The system now provides traceability for each approved invoice, enabling finance to audit decisions, while maintaining high throughput for most invoices.

Рабочий процесс

• 1. Invoice arrives. 2. Agent matches invoice to purchase order and decides approve/reject. 3. Agent writes a record of the decision, including inputs, rule version, model version, confidence, and reason. 4. Agent proposes the decision to a separate service. 5. The service checks hard limits (amount threshold, new supplier, exact match). 6. If limits are met, the proposal is posted to the ledger; otherwise it is sent to a human for review. 7. Access to data is limited to approved queries and actions. 8. Matching rules and prompts are versioned and reviewed like code.

Доказательства

We logged the outcome, not the decision

Источник

Now, whenever the agent approves or rejects something, it writes down the inputs it used, the rule version, the model version, its confidence, and a plain‑language reason

Источник

The agent no longer posts anything itself. Instead, it proposes, and a separate, boring, well-tested service checks that proposal against hard limits before anything touches the ledger

Источник