AI-Enabled AR Invoice Creation Case Study
From purchase order to paid invoice, with a human sign-off at every step.
The result: with AI reading each order and drafting the invoice, invoices now post to QuickBooks in seconds instead of 10 to 20 minutes of manual keying, saving roughly 16 to 33 finance-team hours for every 100 invoices, with zero re-typing and a signed-off audit trail on every step.
The customer
A mid-sized US business-services company that bills its clients through QuickBooks Online. A small finance team turns incoming purchase orders into invoices, sends them, follows up on replies and chases payment.
The problem
Every invoice was built by hand. Purchase orders arrived as PDF email attachments, and someone re-typed each one into QuickBooks, then tracked progress in a spreadsheet.
- 10 to 20 minutes of manual keying per invoice, with typos in amounts and quantities.
- No reliable record of who changed what, or who approved an invoice before it went out.
- Client replies (approvals, queries, disputes) sat in a shared inbox and were easy to miss.
- Payment reminders depended on someone remembering to send them.
Challenges faced
Purchase orders come in many layouts, so fields had to be read reliably without a fixed template.
Customer names on orders rarely match the accounting record exactly, so matching needed a confidence score and a human check.
Finance controls required separation of duties: the person who prepares an invoice must never be the one who approves it.
Client data could not be sent to an outside AI service. Extraction and reply reading had to run on rules first, with only a locally hosted model as a fallback.
Logging into the accounting web interface times out and asks for re-verification, so the solution had to work entirely through the official QuickBooks API, never by clicking through the website.
Some clients operate in other locales (for example Spanish number and date formats), so values had to be normalised before posting.
The solution
An AI-enabled AR invoice pipeline that reads the order, drafts the invoice, sends it, reads the client's reply and records the payment, with four approval gates. A “Maker” triggers each gate and a different “Checker” approves it. Every action lands in an append-only audit log.
Capture: a monitored mailbox picks up purchase-order PDFs and AI-assisted extraction pulls out the PO number, customer, dates, line items and terms.
- Gate 1
Gate 1, customer match: the order is matched to the QuickBooks customer with a confidence score, then confirmed by a Checker.
- Gate 2
Gate 2, invoice review: line items can be corrected with an internal note before the invoice is created in QuickBooks over the API, with the original PO attached.
Dispatch and follow-up: the invoice is emailed, replies are read and classified by AI (approval, query or dispute), and reminders go out on a 24/48/72-hour schedule (maximum three) or on demand.
- Gates 3 & 4
Gates 3 and 4: a Checker decides how to answer the client and confirms payment, which is recorded back in QuickBooks.
Results
Invoices are created in QuickBooks over the API in seconds, replacing 10 to 20 minutes of manual keying each.
Nothing is re-typed into the ledger: line items, notes and the source PO are written by the system.
Every gate decision is stamped with who approved it and when, both in the app and on the QuickBooks record.
Inside the app: from order to payment
1. Billing pipeline dashboard
The home screen. KPI tiles show invoices created, value billed, runs awaiting a gate, credits issued and amounts collected. Below, every run is listed with its customer, invoice number, amount, who handled it and its current status.
2. Gate 1: customer match
A sales order PDF has been read automatically (fields highlighted on the source document). The customer name on the order does not exactly match QuickBooks, so the app flags it and suggests the closest records with a confidence score. A person approves, rejects or corrects the match.
3. Gate 2: human review of the drafted invoice
The invoice has been created in QuickBooks and is shown next to the source order for a side-by-side check. Customer payment history appears at the top. Because the signed-in user started this run, a different Checker or Admin must approve it (four-eyes control).
4. Dispatch and payment reminders
The invoice email has been sent and delivery confirmed. Two payment reminders have gone out on schedule, and the app is watching the inbox for the client's reply. The gate decisions are written back to the QuickBooks customer notes.
5. Client response: corrections filled in automatically
The client replied by email asking to change line 3 to a quantity of 4. The app read the reply, classified it as a corrections request and opened Gate 3 for a Checker. In the Corrected Line Items panel the new values are already filled in: the old quantity (1) and amount ($3,100.00) are struck through, the new ones (4 and $12,400.00) are highlighted, and the new total ($42,500.00) is recalculated. The Checker can still adjust any quantity or amount before clicking “Accept with Edits & Re-invoice”, which re-issues the corrected invoice in QuickBooks. Nobody re-types the invoice.
Get In Touch