Automatically classifying and routing incoming documents with AI in n8n (Drive, Dropbox)

A shared "To process" folder that receives invoices, contracts, resumes and ID documents in a jumble is a classic in almost every small business. Someone drops the file, nobody renames it, and three weeks later you're digging by hand to find "the contract client X signed back in June." A n8n pipeline that reads each document on arrival, identifies what it is, and files it in the right place with a consistent name eliminates this filing work — without changing anyone's upload habits.

Why automate classification instead of sorting by hand

Manual document sorting suffers from three structural flaws:

• It depends on individual discipline: as long as everyone renames and files their documents correctly, the folder stays clean. It only takes one person in a hurry for the system to degrade;
• It doesn't scale: ten documents a week sort themselves without effort, two hundred become a project of their own;
• It leaves no usable trail: without a log of who filed what and when, retrieving the history of a specific document remains a manual search through the file tree.

An LLM plugged in after a storage trigger solves all three at once: it reads the content (not just the filename), classifies according to a schema defined once, and logs every decision.

Step 1 — Detect the arrival of a new document

n8n's Google Drive Trigger node watches a specific folder (it's essential to specify the folder ID rather than leave it on "My Drive" entirely, or the workflow fires on every existing file in your account). Set to a one-minute interval, it detects file creations and updates in that folder. For use on a shared team Drive, remember to enable the "Use Shared Drive" option on both the credential and the node — it's the most common oversight, and it shows up as 404 errors on files that are otherwise perfectly visible in the interface.

For a team that already centralizes files on Dropbox, the Dropbox Trigger node covers the same need with equivalent polling logic on a watched folder — the rest of the pipeline (extraction, classification, filing) stays identical regardless of the source storage.

Step 2 — Extract the content and classify with AI

Once the file is detected, the Extract from File node retrieves the text for native PDFs and Office documents. For a scan or a photo of a document, a multimodal (vision) LLM processes the image directly with no separate OCR step — the same approach detailed in our article on extracting invoice data from PDFs with AI.

The extracted content then flows into a Basic LLM Chain paired with a Structured Output Parser, with a schema that forces a usable answer rather than free text:

{
  "category": "invoice | contract | resume | id_document | purchase_order | other",
  "confidence": 0.0,
  "suggested_name": "2026-07-invoice-vendor-x.pdf",
  "summary": "One-sentence summary of the document's content"
}

The system prompt should explicitly enumerate the possible categories — exactly as with email classification in the Inbox AI Pack — to stop the model from inventing near-identical but non-matching labels ("invoice" one day, "bill" the next), which would be impossible to aggregate later in a Switch node or a report.

This principle of automatic document classification isn't new: as far back as 2014, a study by Kang, Kumar et al. (ICPR 2014) showed that a convolutional neural network could classify document images with performance clearly superior to the rule-based, hand-engineered feature approaches of the time. The difference today: a general-purpose LLM reaches a comparable level zero-shot, with no dedicated training corpus or specialized vision pipeline to maintain — which puts this kind of automation within reach of a small business in an afternoon rather than a data team over several weeks.

Step 3 — Route and rename automatically

The structured category field feeds a Switch node that routes the document to one of its branches. On each branch, a Google Drive node (Move operation) moves the file into the matching subfolder (/Invoices, /Contracts, /HR/Resumes, /KYC…), and an Update operation applies the model's suggested name — typically YYYY-MM-category-identifier.ext, a format that naturally sorts by date in any file browser with no extra effort.

For sensitive documents (ID documents, contracts), it's worth routing them to a restricted-access folder rather than a broadly shared one — the Switch can easily have a dedicated output that also applies a permission change via the Drive API.

Step 4 — Log decisions and handle ambiguous cases

Every filing decision gets written to a Supabase table (classified_documents: original name, category, confidence, final path, timestamp) via the Supabase node. It's the same audit-trail logic described in our n8n-Supabase connection guide, and it plugs directly into the architecture of the Compliance & Audit Pack for organizations that need to demonstrate full document traceability.

The confidence field acts as a safeguard: below a defined threshold (0.7, say), an IF node diverts the document to a Slack notification instead of automatic filing — exactly the pattern described in our article on human approval with the Wait node and Slack buttons. An unreadable scan or a hybrid document (an invoice that also contains contract terms) ends up in front of a human reviewer instead of being silently misfiled.

Pitfalls to avoid

• Processing a bulk import with no pacing: if you point this pipeline at a folder already holding several hundred files, the trigger will send all of them to the model's API at once. A Loop Over Items with a Wait between batches avoids 429 errors — see our article on rate limits of AI APIs in n8n;
• Ignoring duplicates: two uploads of the same file (renamed slightly differently) produce two separate entries. Checking by content hash rather than filename before classifying reduces this risk;
• Letting errors disappear silently: a timeout on extraction or the LLM call should never make an important document vanish without notice — a dedicated Error Workflow, as described in our n8n error-handling guide, covers this case.

How much does it cost

With a cost-efficient LLM (GPT-4o mini or equivalent), classifying a document of a few pages costs on the order of a few cents in tokens, vision OCR on a scan included. For a flow of a hundred documents a month, the bill stays in the range of a few euros — easily offset by the manual sorting time saved, not to mention the reliability of finding a properly filed document instead of one lost in a catch-all folder.

Conclusion

A document classification pipeline comes down to four building blocks: detection on arrival (Google Drive or Dropbox Trigger), AI extraction and classification with structured output, automatic filing and renaming, and logging with a confidence safeguard for ambiguous cases. It's exactly the document-ingestion architecture of the RAG Assistant Pack, reusable as-is: if your classified documents then need to be queryable ("find me the notice-period clause in contract X"), the RAG pipeline picks up right where this one leaves off.