- Custom n8n node
- Documents & AI
- AI-agent tool
Document AI
Google Document AI in n8n: OCR, invoice parsing and document splitting, plus full processor lifecycle control.
How it works in practice
A mixed scan, sorted into readable invoices
How we apply it
In our AI invoice-processing robot for a Bulgarian accounting firm, a custom-trained splitter and reader handle scanned batches of around 200 mixed documents; 97% of documents needed no correction over six months in production.
This node wraps the Google Cloud Document AI API. You send a scanned or digital document (n8n binary data, a Cloud Storage URI or inline JSON) to an OCR, form, invoice or custom extractor or splitter processor and get text, entities and tables back. It goes beyond single calls: batch jobs over a storage prefix, and the whole processor lifecycle (create, enable, train, evaluate, deploy, undeploy, set default version) as workflow steps. EU and other regional endpoints can be chosen for data residency.
Stock n8n has no Document AI node, and the hard part is rarely the process call itself. It is keeping the processor enabled and the right version deployed. Process Document can enable a disabled processor and deploy an undeployed version, wait until it is ready and retry, which takes the usual "not enabled" and "not deployed" failures out of unattended runs. A helper operation turns splitter output into page ranges such as 1-3,5,7-9 for the PDF node.
In practice: supplier invoices and statements that arrive as mixed scans are cut into single documents with a splitter, then read line by line, and idle versions are undeployed on a schedule so we stop paying for hosting. It is used in 17 of our production workflows.
Limits: you need a Google Cloud project with Document AI enabled and a service account, and Google bills per page and for deployed versions. Datasets and human-review queues exist in the API but are not exposed in the node.
For your technical team: it installs from npm as
What it means for you
Nobody splits or types invoices by hand
Month-end peaks no longer need extra people
Fewer typing errors reach the books
More nodes we built
- Banking & FinanceBulgarian BanksEvery bank's statement, read the same wayNo more retyping of statement linesAI-agent tool
- Government & RegistriesBulgarian Registry AgencyA new customer, checked against the official registerNobody copies register data by hand from a browser tabAI-agent tool
- Data & SpreadsheetsExcelA finance template, filled in without touching its lookNo copy-paste into spreadsheetsAI-agent tool
- ERP & AccountingMicroinvestPayments and invoices posted straight into the booksNobody re-keys entries one by oneAI-agent tool
- SecurityPassboltSigning in to a portal without a password in sightPasswords live in one place, not scattered across automationsAI-agent tool
- Documents & AIPDFOne big scan, cut into one file per documentNo manual splitting in a PDF viewerAI-agent tool
- Web AutomationPlaywright browserStatements collected from portals with no integrationNo more clicking through portals by handAI-agent tool
- Web AutomationRobot Framework runnerAn old Windows program, operated by a robotThe old software stays, with no forced replacementAI-agent tool
- Files & StorageSeaweedFS storageDocuments kept safely in storage you ownNo lock-in to a cloud vendorAI-agent tool
- Microsoft 365SharePointDocuments filed in SharePoint, where people lookNo manual filing or renamingAI-agent tool
- Files & StorageSMB network shareFiles exchanged with old systems through a shared folderOlder systems keep working as they areAI-agent tool
Want Document AI in your n8n?
Tell us what the workflow has to do and what it connects to. We will say whether this node covers it as it stands, what we would add, and when a different approach is the better one.
