Every receipt filed, tagged and ready for your accountant
Setup: A weekend (~5-8 h)Running cost: $4-8/mo(software included)what it costs to run ↓Time saved: saves ~7–12 h/week for the owner at 20 runs per day — basisAfter setup: You run it— corrects the tags and amounts it guessed wrongData lives: Your server— Receipts and extracted amounts stay on your server; the supplier mailbox sits with your provider.
Built for small businesses in general — not a sector list.
Paste it into your AI assistant — ChatGPT, Claude, Gemini, whichever you use. It first asks what tools you already have, then rebuilds this blueprint to fit them.
any AI platformno signupblueprint travels inside
Before you start
A scanner or phone app that drops files into a watched folder without extra clicks
A dedicated mailbox for supplier invoices that Paperless-ngx can fetch over IMAP
Disk for years of scans plus a real backup — this becomes your tax record
An accountant who accepts a folder of PDFs and a CSV instead of the shoebox
What breaks firstFAILS SILENTLY
OCR on faint thermal till receipts — the total or the date comes out wrong. Wrong figures get filed and look fine until the accountant queries them; spot-check ten each month.
Version history
v1.0 · 2026-07-05 — First release.
Every supplier invoice and fuel receipt lands in one searchable archive, OCR'd and tagged, ready for your accountant — self-hosted, software cost zero.
One machine, one Docker Compose file. paperless-ngx watches a mail account and a folder; everything dropped in comes out OCR'd, dated, tagged by supplier and document type, and full-text searchable. Your accountant gets a filtered export instead of a shoebox.
Paste it into your AI assistant — ChatGPT, Claude, Gemini, whichever you use. It first asks what tools you already have, then rebuilds this blueprint to fit them.
any AI platformno signupblueprint travels inside
The problem
It's the 4th of the month. Your accountant wants Q2 expenses. Half your supplier invoices are PDF attachments buried in email, a quarter are paper receipts in a drawer, and the rest are screenshots on your phone. You spend an evening forwarding, scanning, and renaming files — and you'll do the exact same thing next quarter. For a 3–15 person business this is typically 2–5 hours a month of pure drag, plus the ones you lose, which cost real deductible expenses.
Tool choices — and why
paperless-ngx
(GPL-3.0, 40.7k stars, release April 2026). The core. Chosen because it bundles the whole intake chain — mail consumer, OCR, classifier, archive, API — in one actively maintained package with a large community. The built-in classifier is not an LLM: it learns from your manual corrections, which is exactly right for a stable set of 20–50 recurring suppliers. Alternatives like building your own OCR pipeline mean owning table extraction, retries, and monitoring yourself — a net negative for a small team.
OCRmyPDF
(MPL-2.0, 34.1k stars, release April 2026). The OCR engine paperless-ngx uses under the hood (Tesseract-based). Listed separately because it's also useful standalone for batch back-OCR of your existing PDF archive before migration.
Optional LLM extraction — Ollama
(MIT, 175k stars, release June 2026). If you later want line-item extraction (amounts, VAT, invoice numbers as structured data), run a small local model via Ollama against paperless-ngx's API. Start without it; the searchable archive alone kills most of the pain. A note on vetting: the popular paperless-ai companion project failed our checks — its README carries an explicit "not maintained" notice and the last release is roughly 8 months old — so we don't recommend it, whatever the star count says.
Monthly cost
Software: €0, all self-hosted open source. Hardware: €0 if you have any spare mini-PC or NAS that runs Docker; otherwise a small European VPS runs roughly €4–8/month (verify current pricing — it moves). Your time: expect a half day for setup and a couple of weeks of casually correcting tags while the classifier learns. Compare: commercial document-intake SaaS for this typically starts around €15–50/month per user.
Setup outline
1. Create a dedicated mail address (invoices@yourdomain) and tell suppliers to use it. This one habit does half the work.
2. Deploy paperless-ngx with the official Docker Compose file on your box or VPS.
3. Configure the mail rule: consume attachments from invoices@, delete-after-import off until you trust it.
4. Set up a consume folder synced from your phone's scan app (any app that saves PDF to a synced folder works).
5. Define document types (invoice, receipt, contract) and correspondents for your top suppliers; correct tags for two weeks so the classifier learns.
6. Monthly: filter by tag and date range, export, hand to accountant — or import the CSV of metadata into your bookkeeping tool.
Pitfalls — the real ones
Don't expose it to the internet
The paperless-ngx team says it plainly: documents are stored unencrypted, so run it on your LAN or behind a VPN, never on an open port. Your invoices are exactly the data an attacker wants.
Backups are not optional
This becomes your system of record within a month. Snapshot the data volume off-machine, and test a restore once.
OCR is not bookkeeping
Tesseract-class OCR reads clean print well and handwriting badly. Treat extracted totals as search aids, not accounting truth — the human posting entries still checks the figure.
EU note:
structured e-invoicing (Peppol and national mandates, already mandatory in Belgium for B2B since January 2026) will progressively shrink the PDF-invoice problem. This pipeline stays useful for receipts, foreign suppliers, and the long tail — but don't over-invest in extraction sophistication the regulation will make redundant.
Verified repos
paperless-ngx — GPL-3.0, 40.7k stars, active (April 2026) OCRmyPDF — MPL-2.0, 34.1k stars, active (April 2026) Ollama — MIT, 175k stars, active (June 2026)
Get the next blueprint as it publishes — free, practical, verified tooling only.
Subscribe · Want this audited or extended for your business? DM @RXed_EU
What this costs to run
Priced as Extract fields from a document — one OCR'd page; output is a JSON record of 8-20 fields. one run per received document.
Adjust the volume to yours; the bill is a range because the assumptions are ranges.