The problem
Following up on unpaid invoices takes more than spotting an overdue date. Someone has to check the customer’s history, understand disputes and choose an appropriate response.
What the assistant does
It reads invoice and customer records, identifies what needs attention, and drafts the next action with the relevant evidence. A person can edit and approve each draft. A client version was delivered; this public example uses twelve sample invoices and customer records.
Follow an invoice through the workflow
Start with “What is the story with the disputed one?” The example shows why a correction can be more appropriate than another payment reminder. The demo records approvals without sending email.
2Ask the Assistant
Up to 8 questions per visit in this use case.
3Your Decisions
Drafts you approve show here. In a deployment, approval is the step that sends the email.
- No decisions yet.
Why the review step matters
The assistant can gather context and prepare a response, while the owner controls what is sent. Drafts remain editable, and the receipt identifies the records and tools used. This demonstrates the workflow; no measured time-saving claim is made here.
Implementation notes Data, deployment and checks
/api/ops-bot. A Claude model plans through the Anthropic API with read-only tools over the records; a rules path over the same tools keeps the page answering if the model call fails.