The invoice triage build: four hours a week back in twenty days

A UK professional services firm reached out in February. The brief was narrow and the frustration was obvious.

Their bookkeeping lived in a shared accounts@ mailbox. Suppliers emailed PDFs. Clients forwarded receipts. A part-time bookkeeper spent roughly four hours every week opening each message, figuring out whose invoice it was, classifying it by expense category, renaming the file and filing it into one of two Google Drives.

She didn’t hate the work. She hated that the work never ended.

What we built

A single n8n workflow runs on the box every five minutes.

1. Gmail polling pulls any unread email tagged Inbox and Invoice/Receipt (they already used labels, which made scoping clean).
2. Attachment extraction pulls PDF or image attachments into a temporary bucket.
3. Claude (Sonnet) classifies each attachment. The prompt includes the last thirty vendors they’ve paid, with their canonical names and typical expense category. Claude returns vendor, amount, VAT, date, and expense category in a strict JSON shape. We validate the shape before moving on.
4. The file gets renamed YYYY-MM-DD_VENDOR_AMOUNT.pdf and dropped into the right Drive folder. A row lands in Airtable with the audit trail (original email, Claude’s confidence score, the actual reasoning).
5. Edge cases escalate. If Claude’s confidence is below 0.8, or the VAT number doesn’t match UK format, the email gets a needs-review label. The bookkeeper sees it in the normal filter she already uses.

Numbers

• Week 1-3: build, test, iterate on the classification prompt with real invoices.
• Week 4: soft launch. Bookkeeper checks every single classification for five days.
• Month 2 onwards: spot-checks once a week.

After the first month: 96% classification accuracy across 400+ invoices. The 4% that escalated were genuinely ambiguous (missing VAT, unclear vendor, a supplier’s new trading name). Exactly the cases a human should see.

Four hours a week back. The bookkeeper moved to higher-margin reconciliation work. The firm didn’t need to replace her — they just stopped paying for the bad part of her role.

What we’d do differently

Two things.

One: we’d ship the dashboard sooner. The Airtable audit log was an afterthought. The moment the client actually looked at it — seeing every classification decision with Claude’s reasoning — trust in the system jumped. Visibility isn’t a “nice to have”. It’s the thing that turns “this weird AI” into “a system we understand”.

Two: we’d build the prompt with THEIR invoices, not synthetic ones, from day one. The first iteration used a generic “classify this invoice” prompt. Accuracy was 82%. The moment we swapped in their actual last-thirty-vendors list, accuracy jumped to 96% and stayed there.

If you’re thinking about this

The hard bit isn’t the automation. It’s the classification taxonomy. Before you write any code, you need to know the fifty most common vendors you pay and what they should be categorised as. If that list doesn’t exist, make it first.

If it does exist, we can wire this up in three weeks. Get in touch and we’ll tell you whether it’s worth it for your specific mailbox.