What does document and invoice automation cost for an accounting or tax firm?
Short answer: Off-the-shelf capture tools list from €19 to €179 a month for 20–500 documents. Cloud extraction APIs charge about $0.01–$0.10 per page or document. Published agency quotes for a custom production pipeline run roughly $12,000–$50,000. A private, on-premise build keeps documents in your building. Start with a fixed-fee audit and a pilot on 100 of your own invoices.
Last reviewed: 2026-10-31. Third-party prices were read from the vendors' own pages on that date. They are list prices, net of VAT where stated, and in USD where the source gives USD.
- Accounting firms and tax advisers who receive receipts and invoices from clients in every format: paper, email attachments, phone photos, portal downloads.
- SME finance and AP teams who type invoice data into accounting software and chase approvals by email.
- Firms already running Paperless-ngx (or another DMS) who want the next step: documents that are read, checked and routed, not just archived.
- Firms that want Paperless-ngx without running it themselves. We set it up as part of a complete document workflow (see "Done for you" below).
Reading text off a scanned invoice is a solved problem. The cost sits in everything around it: opening each PDF, finding the vendor, total and due date, retyping, checking, chasing the approval and filing.
Published benchmarks give a sense of scale. In Ardent Partners' 2025 AP benchmark, the average cost to process one invoice was $9.40, against $2.78 for best-in-class organisations, and the average invoice took 9.2 days to process (WEX, reporting Ardent Partners).
Your own number matters more than any benchmark. Use this formula:
Annual cost = (documents per month × 12 × minutes per document ÷ 60 + rework cases × minutes per rework ÷ 60) × loaded hourly cost
Time 20 real invoices end to end, including approval and filing, before you trust any estimate.
In plain words: software reads each document, turns it into a fixed set of fields, checks those fields against your rules, then either prepares the next step or hands the document to a person, with the reason written down.
- Collect. Invoices arrive from a shared inbox, a scan folder, a cloud-scanning app or an upload. Structured e-invoices (XML) skip OCR entirely.
- Quality check. Resolution, sharpness and skew are checked first. Bad scans are flagged before anyone trusts them.
- Read. An AI OCR model reads the page, including layout, tables and reading order. It can run on your own hardware.
- Structure. The text is forced into a fixed field list: vendor, invoice number, dates, currency, net, tax, gross, PO number, line items, bank details. If a required field is missing, the pipeline stops instead of guessing.
- Check and decide. Your rules run as plain code. Does the maths add up? Is it a duplicate? Is it over the approval limit? Is it a new supplier, or have the bank details changed?
- Act and log. The system prepares the booking or payment, files the document in your DMS (for example Paperless-ngx with the right tags and permissions) and writes every step to an audit log.
Each document ends in one of three states: prepared (clean, under the limit), flagged (a hard rule failed) or needs review (new vendor, blurry scan, uncertain field). The AI never overrides a failed rule, never releases a meaningful payment on its own and never takes instructions from text inside a supplier's PDF.
This isn't a slide concept. Made By Agents has published two open-source reference builds of this design: doc-approval-demo and ocr-agent-pipeline. The build behind the second one is explained in How to Build a Local AI Agent for Invoice Processing.
| A. Off-the-shelf capture / AP software | B. Cloud OCR API + your own integration | C. Custom private build (on-premise or your chosen region) | |
|---|---|---|---|
| Best when | Standard invoices, standard approval flow, a supported accounting system | You have developers and want extraction as a building block | Your rules, document mix, data-location needs or volume don't fit a product |
| Where documents go | The vendor's cloud, under their terms and processor agreement | The API provider's cloud; you control everything around it | They can stay on your own hardware or in your chosen region |
| Fit to your rules | Whatever the product's settings allow | Whatever you build around the API | Your rules as code: limits, matching, vendor checks, routing per document type |
| Upfront effort | Low: configuration and onboarding | Medium: integration, rules, review screen | Higher: build, test set, integration, monitoring |
| Running cost shape | Subscription, often tiered by documents or users | Per page or per document | Hardware or hosting plus maintenance; the cost per document stops growing with volume once the hardware is paid for |
| Lock-in | Highest: workflows and data live in the product | Medium: the API can be swapped with some work | Lowest: you own the code and can swap models |
| Watch out for | Tier jumps as you grow, export options, edge cases | You still build the checks, approval queue and audit log | Someone must own it; model licences above revenue thresholds |
A hybrid is common: keep your accounting system and its standard features, parse e-invoices directly, and add a custom layer only where it pays. That usually means the messy scans, the special approval rules, or documents that must stay on-premise.
| Option | Published price | Source |
|---|---|---|
| Off-the-shelf capture (DACH example) | From €19/month (20 documents) to from €179/month (500 documents), plus VAT | GetMyInvoices pricing |
| Off-the-shelf parsing (international example) | $49/month for 100 pages to $499/month for 10,000 pages | Parseur pricing |
| Cloud API, invoice-specific | $0.01 per page for the first million pages a month (Textract Analyze Expense, US West example) | AWS Textract pricing |
| Cloud API, invoice parser | $0.10 per document at the entry tier | Google Document AI pricing |
| Custom pilot (one document type) | $6,000–$12,000, 2–4 weeks | Stackzeno |
| Custom production (1–2 document types, integration) | $12,000–$35,000, 4–10 weeks, plus $200–$1,500/month to run | Stackzeno |
| Custom production workflow (agency band) | $15,000–$50,000, usually 4–16 weeks | KUMO |
| First document pipeline (UK fixed-price band) | £4,000–£15,000 | Augustova |
| Made By Agents | Fixed fee, quoted after a free Discovery Call; the audit fee is credited if implementation starts within 90 days | /audit |
The agency bands above are other firms' published ranges, not Made By Agents quotes. Made By Agents doesn't publish prices: every engagement is a fixed fee, quoted after a free Discovery Call once the scope is clear.
- Number of document types and layouts. Invoices, receipts, credit notes and delivery notes each need their own fields and rules.
- Scan quality. Clean digital PDFs are easy. Faxes, photocopies and phone photos of crumpled receipts are where errors come from.
- Where the data must end up. An accounting system with an API or an import file is cheaper to connect than one that needs workarounds.
- Your approval rules. Limits, four-eyes approval, PO matching and supplier checks are the real work, and the real value.
- Data location. On-premise processing adds hardware and maintenance but removes the processor question.
- Volume. Per-page pricing is cheap at low volume. Own hardware gets more attractive as volume grows.
- E-invoice share. The more invoices arrive as XRechnung or ZUGFeRD, the less OCR you need.
| Phase | Typical length | What happens | What you get |
|---|---|---|---|
| 0. Discovery Call | 20 minutes, free | Are there real processes, decision makers and access? If an audit is the wrong step, we say so. | An honest next step |
| 1. Deep AI Audit | Most audits take 1–2 weeks; fixed fee | Map the invoice process as it really runs, rank what to automate, decide build vs buy, write the spec for the first system | A roadmap and a spec; fee credited if implementation starts within 90 days |
| 2. Pilot | Scope and length set in the audit | One document type, one intake channel, a test set of about 100 of your own invoices, shadow mode (the system proposes, your team decides) | Your real accuracy, your real time per document, a go/no-go against criteria agreed up front |
| 3. Production | Rolled out in steps | Background queue, approval queue with a real button, integration with accounting, live credentials last, monitoring | A system your team uses daily, with an audit trail from document to booking |
| 4. Extend | Ongoing | More document types, an e-invoice fork, more intake channels | More of the process covered |
Before the audit, check what you already have:
- A rough monthly document volume.
- One person who owns the invoice process.
- About 100 real sample invoices, including messy ones.
- Written approval rules and limits (or ones you could write in a day).
- Knowledge of which system the data must end up in, and whether it can be connected.
- A decision on where documents may be processed: on-premise, EU cloud or any cloud.
An unticked box isn't a blocker. It's a work item for the audit.
Invoices contain personal data: names, addresses, bank details. That makes the processing setup a compliance question, not just a technical one. This is general information, not legal or tax advice.
- Cloud services are processors. If documents go to an external OCR or AI service, you need a written processor agreement under Art. 28 GDPR, plus appropriate security measures under Art. 32 GDPR.
- Transfers outside the EU follow the rules in Art. 44 GDPR and the articles after it. That matters for US-hosted APIs.
- On-premise is possible. Open-source OCR models can run on your own hardware, so the answer to "where did this document go?" is "nowhere". Check the licences: the Surya code is Apache 2.0, but its model weights are free only for research, personal use and startups under $5M funding or revenue (Surya on GitHub). Chandra's threshold is $2M (Chandra on GitHub).
- Retention still applies. Under § 147 (3) AO, German booking vouchers such as invoices are kept for eight years and books and records for ten. Your archive (for example Paperless-ngx, open source under GPL-3.0, on GitHub) has to hold them for that long, whatever the automation does.
- E-invoicing is coming anyway. Since January 2025 every German business must be able to receive e-invoices. Issuing becomes mandatory from January 2027 above €800,000 turnover and from January 2028 for all others (EDICOM). Build the XML path next to the OCR path.
Made By Agents doesn't sell a bare Paperless-ngx install. We set Paperless-ngx up as one part of a complete document workflow:
- Intake: shared inboxes, scan folders and cloud-scanning apps feed one queue.
- Archive and access: Paperless-ngx with tags, storage paths and permissions for your team and your tax adviser.
- Reading and checking: AI OCR, field extraction and your rules (totals, duplicates, limits, new suppliers).
- Routing: prepared bookings, an approval queue, and a review queue for anything uncertain.
- Operations: two-factor authentication, encrypted backups, monitoring and a documented handover.
The scope (which documents, which systems, where it runs) is set in the Deep AI Audit. You can then run it yourself or keep us on as your AI Partner.
Made By Agents helped Canova GmbH, a real estate business, set up Paperless-ngx and document workflows for managing their business.
- Documents: real estate transactions (buying and selling), rentals, bank financing, permits and other paperwork with authorities, and contracts with tenants.
- Users: a team of about 3–5 people. Their tax accountants have access too, with separate access rules, so each side sees only what it needs.
- Intake: everything is scanned with a ScanSnap document scanner straight into Paperless-ngx, where it is text-recognised (OCR), made searchable and filed.
"Made By Agents has been an invaluable partner in our digital transformation. They helped us identify and implement the exact AI tools needed to automate processes that were previously paper-heavy and prone to human error. Every technical question I had was met with high-level competence."
— Christoph F., Managing Director, Canova GmbH
- You handle a few dozen standard invoices a month and your accounting software already reads them well.
- Most of your suppliers will send e-invoices within the year.
- Nobody can own the process or the system after go-live.
In those cases we'll recommend a standard tool, or waiting. That's part of the audit's build-vs-buy step.
- Paperless know-how. Our Paperless-ngx guides cover tax-adviser and bookkeeping workflows, employee permissions, Nextcloud integration, two-factor authentication, installation on a VPS and encrypted backups.
- Local AI OCR. We've compared the main open-source OCR models and built a local invoice pipeline that runs on a laptop.
- You work with the founder. Tobias Wupperfeld has more than 5 years of AI development experience and runs every Discovery Call and audit himself. He is also Head of AI at JAN3, where he leads the A1 Lab division. Its first app, Echos, is a free, open-source transcription app that runs entirely on the device and was built in roughly two weeks (JAN3 announcement). The same privacy-first thinking goes into document pipelines.
How much does AI invoice processing cost for a small accounting firm?
Off-the-shelf tools list from €19 to €179 a month for 20–500 documents. Cloud APIs charge per page (about $0.01 on AWS Textract). Published agency quotes for a custom pipeline sit around $12,000–$50,000.
Do we need a custom build, or is a standard tool enough?
Often a standard tool is enough. Custom pays off when your rules, document mix, data location or volume don't fit a product. The Deep AI Audit decides this.
Can invoices be processed without leaving our office or the EU?
Yes. Open-source OCR models such as Surya and Chandra run on your own hardware; check their licence thresholds. Cloud services need a processor agreement under Art. 28 GDPR.
How long does a rollout take?
The audit usually takes 1–2 weeks. The pilot is scoped in the audit and measured on about 100 of your invoices. Production follows in steps.
Does this replace Paperless-ngx or our DMS?
No. Paperless-ngx stays the archive. The automation reads, checks and routes documents before they are filed.
How accurate is AI invoice extraction?
Only a test on your own documents can tell. We measure field-by-field accuracy on your invoices and send uncertain documents to a person.
What about e-invoices (XRechnung, ZUGFeRD)?
German businesses must receive them since 2025 and issue them from 2027 (above €800,000 turnover) or 2028. They skip OCR and are parsed directly.
Can you set up Paperless-ngx for us?
Yes, as part of a complete document workflow, not as a standalone install. The scope is set in the Deep AI Audit.
Book a free 20-minute Discovery Call. We look at your document process and tell you honestly what comes next: a Deep AI Audit, a pilot, an off-the-shelf tool, or nothing yet.
Book a Discovery Call → · About the Deep AI Audit →
Last reviewed: 2026-10-31 by Tobias Wupperfeld.