Can AI sort and tag incoming documents? Yes, for categories you define clearly, with an unsure bin for a person.
Short answer. Yes, for categories you define clearly. A language model can read each incoming document and assign a type such as invoice, notice, contract or ID proof, and the structured-output features of Ollama and vLLM can force the answer to be one label from your list. Image-based models classify scanned pages directly: on the RVL-CDIP benchmark of 400,000 scanned documents in 16 classes, the DiT-L model reached 92.69 per cent accuracy. Real mail is messier than a benchmark, categories overlap, and a wrong label can hide a document. So include an unsure option, route those documents to a person, and measure accuracy on your own last month of mail before you switch it on.
01
Three ways to sort
Method
How it works
Good for
Watch for
Language model on extracted text
Prompt with your categories and a one-line definition of each. Use structured output so the answer must be one allowed label.
Typed or OCR-read documents, and categories that change.
Needs OCR first. Mislabels when two categories overlap.
Embeddings and labelled examples
Convert each document to an embedding and compare it with examples of each category.
Offices with many already-labelled documents.
Needs good examples for every category.
Image classifier on the page
A model classifies the picture of the page, for instance DiT, a document image transformer.
Scans whose layout is distinctive, such as invoices against letters.
Needs training on your own classes.
02
What a benchmark number means
RVL-CDIP is a standard test for document image classification: 400,000 greyscale images in 16 classes, with 25,000 images per class. The DiT-L model reached 92.69 per cent accuracy on it. That is a useful sign that page-level classification can work, but your categories will not match the benchmark’s 16, your scans will be of different quality, and a category such as correspondence will contain items that belong in three of your own folders. Treat the figure as an upper hint, not a forecast.
03
What a person must still check
The unsure bin, every day at first. It shows where your categories are unclear.
A random sample of confident labels. Wrong but confident is the dangerous case.
Look-alike categories, such as a tax notice and a tax invoice, where the wrong tag sends a document to the wrong desk.
Sensitive categories, such as complaints or legal notices. Tag them, but do not route them automatically until you have months of evidence.
That tags themselves do not reveal sensitive information to people who see only the tag.
04
How an office would set this up
Start with five to ten categories, each with a one-line definition and three real examples.
Take last month’s mail, label it by hand, and run the model on it. Count the errors by category.
Add an other option and an unsure option, and route both to a person.
Constrain the output to a fixed list of labels, so the model cannot reply with anything else.
Go live as tagging only. Start automatic routing with low-risk categories.
Review the error counts monthly and retest when you change the model.
05
What ANVE does today
ANVE does not list document sorting or tagging as a function. Its vision models read scanned documents with OCR, and its documents function answers questions over the office’s own files and shows the source passage.
Both run today in ANVE’s development build. No sealed unit exists yet.
06
Questions people ask
How many categories is too many?
Start with five to ten. Every extra category needs a definition, examples and testing, and more categories mean more look-alikes.
Can it sort Hindi or mixed-language letters?
It can if the model reads the language and the OCR handles the script. Llama 3.3 and Qwen3 list Hindi among their languages. Test with a month of your real mail.
What about instructions hidden in an incoming document?
OWASP describes indirect prompt injection, where content from files or websites changes a model’s behaviour. If the model can only output one label from a fixed list, a hidden instruction has little to act on.
Can it route documents to the right person automatically?
Yes, once tagging is accurate on your own mail. Start with low-risk categories, keep a human check on the rest, and log every routing decision.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
