Document AI Reading business documents into fields, tables and labels, not just text.
Short answer. Document AI is the family of techniques that read business documents, such as invoices, forms, contracts and reports, and turn them into structured data: the right fields, tables and labels, not just a block of text. The authors of LayoutLM define it as techniques for automatically reading, understanding and analysing business documents. A typical system digitises the page with OCR, extracts fields and tables, then classifies and splits files by type. Cloud services such as Google’s Document AI do this in the provider’s cloud; open tools and local vision models can do it on an office’s own hardware.
01
The four jobs
Google’s overview groups document processing into these jobs, and the same split applies to open tools.
Job
What it does
Example in an office
Digitise
Runs OCR to get text and layout, with checks such as image quality and deskewing
A scanned GST invoice or a phone photo of a form
Extract
Finds fields, tables and key-value pairs, such as Name: Jill Smith
LayoutLM, a 2020 research model, learns from the text of a page together with where the text sits and what the page looks like. Its authors reported gains on form understanding (from 70.72 to 79.27), receipt understanding (94.02 to 95.24) and document image classification (93.07 to 94.42) on public benchmarks.
Donut takes a different route and reads the page image directly, without a separate OCR step. Google’s service is built on its Vertex AI products with generative AI, offers pretrained processors for common document types, and lets you create custom extractors or up-train pretrained ones. Benchmark gains come from public datasets, not your invoices.
03
Local or cloud
Option
Where documents go
Example
Cloud document service
To the provider’s cloud platform
Google Cloud Document AI, which Google describes as cloud-based document processing
Open tools on your own hardware
They stay on your machine
Tesseract for OCR; LayoutParser, an open-source library for deep learning on document images
Local vision-language model
It stays on your hardware
A vision model that reads the scan and returns fields
04
How to evaluate one on your documents
Collect 30 to 50 real documents of each type, including the worst scans.
Measure field-level accuracy, not only character accuracy: a correct page with a wrong tax amount is a failure.
Check what happens with blank, smudged or missing fields: does the tool say it is unsure or guess?
Set a human-review rule, for example every amount above a limit and every low-confidence field.
Reconcile totals and validate identifiers such as the GSTIN checksum automatically.
Review where the data goes, how long it is kept and who can see it, before sending a single real document.
05
Where ANVE stands
ANVE reads GST bills and scanned documents with its vision models, and answers questions over the office’s own documents with the source passage shown. Full document drafting with design templates is not on ANVE yet. Models run on the box unless an administrator adds a cloud provider, in which case their answers are labelled, and every model call is logged with whether it ran locally or in the cloud, as ANVE’s security page sets out.
06
Questions people ask
Is Document AI the same as OCR?
OCR is the first step. Document AI adds extraction of fields and tables, classification and splitting.
Does Document AI need training data?
Google’s service offers pretrained processors for common document types, custom extractors and up-training of pretrained processors. How many labelled examples you need depends on the method.
Can Document AI run offline?
Cloud services cannot. Open-source tools such as Tesseract and LayoutParser, and local vision models, can.
How accurate is it?
Published figures such as LayoutLM’s gains on form and receipt benchmarks come from public datasets. Accuracy on your invoices and forms can differ, so test a sample.
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.
