n8n Receipt OCR Automation: Email a Receipt, Get a P&L Row

A box of receipts. A shoebox, a phone camera roll, a folder of forwarded PDFs nobody's opened. Month-end arrives and someone keys them into a spreadsheet by hand, mistyping a total or skipping the tax line. The P&L is late and slightly wrong, every single month.

n8n receipt OCR automation ends the manual keying. You forward a receipt or invoice to an inbox, GPT-4o-mini Vision reads it, and the vendor, date, amount, tax, and category land as a row in your Google Sheets P&L. No third-party parsing SaaS, no shoebox, and a confidence gate that keeps the bad reads out of your books.

The guides ranking here lean on extra services and skip the gate. The Azure OCR receipt guide chains Azure Computer Vision plus a chat model and never mentions confidence scoring, email ingestion, or P&L reporting. The popular Telegram expense template is photo-via-Telegram with a Mistral agent, no Gmail polling, and no confidence routing to an Errors tab. Email-in, vision-direct, confidence-gated is the unwritten version.

What receipt OCR automation can do

Receipt entry is a clerical loop that a workflow erases:

• Watch an inbox for emails carrying a receipt or invoice attachment.
• Read the image directly with a vision model, no separate OCR step.
• Extract vendor, date, amount, tax, currency, and category as JSON.
• Score each extraction's confidence and route on it.
• Append high-confidence rows to the P&L, low-confidence to an Errors tab.
• Send a monthly category-breakdown digest of where the money went.
• Run entirely on your n8n instance plus an OpenAI account.

That's bookkeeping that updates itself between month-ends.

Why the confidence gate is non-negotiable

Here's the take worth defending: a receipt OCR flow without a confidence check is worse than no automation, because it logs wrong numbers with the same confidence as right ones. A human keying receipts at least squints at a blurry total. A naive workflow appends whatever the model guessed and you discover the error during a tax filing.

The fix is one extra field and one extra branch. Ask the model to return a confidence score with the data. An n8n-nodes-base.if v2 node reads it. High-confidence receipts flow to the main P&L tab; anything uncertain or unreadable routes to a separate Errors tab where a person spends ten seconds confirming or fixing it. The fast path stays fully automatic for the clean 90 percent, and the murky 10 percent gets a glance instead of silently corrupting the ledger.

Dropping the separate OCR service is the other quiet win. The two-step Azure-then-model pattern was necessary when vision models couldn't read images. GPT-4o-mini Vision reads the receipt and returns fields in one call, so the workflow loses a service, a credential, and a failure point.

The receipt OCR pipeline

Gmail poll (every 5 min, attachment present)
        │
   Extract first attachment (image/PDF)
        │
   GPT-4o-mini Vision → vendor, date, amount, tax, currency, category, confidence
        │
   Parse JSON (Code node)
        │
   IF confidence high? ── yes ──► append to P&L tab
        │ no
   Append to Errors tab (manual review)

One pass, with a gate that protects the books.

Step-by-step breakdown

1. Poll Gmail for attachments

A Gmail trigger checks every 5 minutes for new emails carrying an attachment. Poll on an interval, not every minute; tight polling burns OAuth quota for no benefit since receipts arrive in clusters, not by the second. Forwarding a receipt to one inbox is the lowest-friction capture there is.

2. Pull the attachment

A n8n-nodes-base.code v2 node grabs the first attachment, whether it's a photo, a screenshot, or a PDF, and prepares it for the model. One attachment per email keeps the mapping simple; multi-attachment emails are an edge case you can add later.

3. Read the receipt with Vision

The @n8n/n8n-nodes-langchain.openAi v2.1 node sends the image to GPT-4o-mini Vision with a structured prompt asking for vendor, date, amount, tax, currency, category, and a confidence score. Its response lands at $json.output[0].content[0].text. The model needs the modelId.__rl wrapper (value gpt-4o-mini, mode list) to resolve correctly; copying an older n8n-nodes-base.openAi config from a 2023 tutorial breaks the output path silently.

4. Parse the JSON

The model returns text, and the next node expects an object, so a n8n-nodes-base.code v2 node parses the JSON every time. Add this step even when it feels redundant. A missing parse is the most common reason these workflows fail, and the execution log makes the broken parse obvious while the failure itself doesn't announce.

5. Route by confidence and log

An n8n-nodes-base.if v2 node compares the confidence score against your floor. High-confidence extractions append a timestamped row to the main P&L tab via n8n-nodes-base.googleSheets v4.5; low-confidence ones land in an Errors tab. A separate monthly digest reads the ledger on the 1st and emails per-category totals, so trends surface without anyone building a report.

Implementation patterns

Pattern 1 — Vision-direct extraction. Skip the separate OCR service entirely. One vision call returns structured fields, which removes a credential and a whole class of layout-change failures.

Attachment → GPT-4o-mini Vision (image + structured prompt) → JSON fields

Pattern 2 — Confidence routing. The model's own confidence score drives the split between trusted and review.

Parse JSON → IF (confidence ≥ floor) → P&L tab | Errors tab

Pattern 3 — The month-end digest. A second n8n-nodes-base.scheduleTrigger v1.2 fires on the 1st, reads the ledger, totals by category, and emails a plain-text summary. The reporting comes free off the data the intake already captured.

n8n nodes you'll use most

Getting started

1. Set up a Google Sheet with a P&L tab and a separate Errors tab.
2. Add the Gmail trigger and point it at the inbox you'll forward receipts to.
3. Pull the first attachment in a Code node.
4. Send it to GPT-4o-mini Vision with a structured field-and-confidence prompt.
5. Parse the JSON, then route on confidence with an IF node.
6. Append to the right tab, and add a monthly digest on a Schedule trigger.
7. Or import a tested flow and start forwarding receipts now.

Where this fits

Receipt capture is the front door of a finance stack. Once expenses land in a structured ledger, the same data feeds reporting and reconciliation, which is why teams pair this with broader expense tracking automation. The confidence-routing pattern here is the same trustworthiness gate that shows up in data validation workflows anywhere a model's output enters a system of record.

A workflow that turns a forwarded receipt into a clean P&L row makes month-end a query instead of a chore. For the bigger picture, see how this connects to finance automation across the books, and if invoices flow out as well as in, pair it with the Automated Invoice Chase & Escalation template to keep receivables moving too.