How to automate invoice chasing with AI

What did you do that nobody billed?

Start upstream of the debt.

Harvest holds every tracked hour and every expense, and the question is which of them ever reached a document.

Prompt: “Open Harvest. List every time entry and expense from the last two closed months that is not attached to an invoice, grouped by client and project, with hours, the billable rate on the project, and the value at that rate. Then list any project where tracked hours exceed the budget field. Put it in a sheet called ‘Unbilled review’ and total it per client. Do not create anything yet.”

The total at the bottom of that sheet is money you already earned and never asked for.

Raise the invoices, one approval at a time

A companion drafts each invoice from the unbilled entries and shows you the table before anything is created.

Prompt: “For the three clients on the ‘Unbilled review’ sheet that I have ticked, draft one Harvest invoice each covering the unbilled entries, using each project’s own rate, with line items grouped by task and the period in the subject. Show me each draft as a table before you create it. Do not send any of them. I will decide who gets an invoice by email and who gets it inside their portal.”

Harvest and Xero writes both stop at an approval card, so a mistyped line item never reaches a customer as a PDF.

Which payments already landed?

Mercury reads accounts, balances, statements and transactions.

That is what closes the gap between unpaid in Xero and actually unpaid.

Prompt: “Pull every credit into our Mercury operating account over the last 90 days. Match each one to an open Xero invoice by amount, reference, and payer name. Give me three lists: confident matches with both ids, credits you cannot attribute at all, and Xero invoices still open with no matching credit. For the confident matches only, prepare the Xero payment records and show me the batch before applying it.”

The unattributed list is the one to read yourself. A €14,200 transfer covering three invoices minus a bank fee comes back to you rather than guessed at.

Rank the chase list by money, not by age

Sorting by days overdue puts a small dormant invoice at the top. Sorting by amount times days overdue puts the money there.

Prompt: “Take the still-open Xero invoices. For each, add days overdue, the amount, the contact from Xero, and the date of the last email in my Gmail that mentions that invoice number or the client’s billing address. Sort by amount times days overdue. Then draft a short chaser per row in Gmail as a draft, threaded onto the original invoice email where one exists, and leave every one unsent.”

Every row carries what decides the tone: how old, how big, and whether anyone has mentioned it since. Ordinary operations work, three systems into one ranked list.

How do you make it the Tuesday sweep?

Routines run a saved pass on a schedule. The whole loop above is worth repeating weekly.

Prompt: “Every Tuesday at 8am: refresh the unbilled sheet from Harvest, re-run the Mercury-to-Xero match, and post one Slack message to #finance: total unbilled, total open, what got paid since last Tuesday, and anything crossing 45 days for the first time. Leave the drafts in Gmail and never create an invoice without asking me.”

The routine keeps the ledger honest by itself. Anything that touches a customer or a document waits for you.

Frequently asked questions