Nanonets
by Nanonets
Getting-started guide
Feed it the same document a hundred times and it learns your fields — the first twenty are the work.
Current prices: Free, from $499/mo · See all plans
Last reviewed 2026-09-09 · for Nanonets
You upload documents — or forward them from an email inbox — and Nanonets reads them, showing the fields it found beside the page it read them from. You check each field, fix what it got wrong, and approve. What you approve feeds back in, so the next batch arrives needing fewer corrections than the last.
Nanonets pulls structured data out of documents that arrive as files — supplier invoices, purchase orders, bank statements, delivery notes, identity documents — and hands the result to the software you keep records in, so nobody retypes it.
It is built for the back office rather than the desk. The people it is designed for handle the same kind of document hundreds of times a month and want that handling to stop being somebody's afternoon.
Beyond reading, it will act on what it read: route anything it is unsure about to a named person, then push the approved result into accounting software like QuickBooks or Xero, or a larger business system like SAP or NetSuite.
If you have one contract to understand, this is the wrong shape of tool — a document chat tool answers you in a minute. Nanonets only pays back on repetition.
Every account currently starts with free credits (the platform's unit of spend — each thing you run costs some) and no card, and Nanonets says credits never expire. That is enough to put real documents through before you decide anything.
The prices are published, which is unusual in this category: each step has a listed rate, with a standard AI step costing more than a simple one and a data-extraction run more again. There is one self-serve paid plan; above that the buttons read Talk to sales and Book a demo.
Bring a stack, not a file. Nanonets' own documentation describes a new extraction starting at 0% confidence and rising only as you correct and approve documents, with 90% named as the point it is ready to run unattended. One document cannot move that number.
You do not need accounting or ERP software connected to try it. The free plan currently includes importing from an email address and from cloud folders such as Google Drive, Dropbox, OneDrive and SharePoint, and results can come back out as a spreadsheet.
The free plan is currently capped at three users with community support, so it is a working trial rather than a small-team licence.
The stake here is different from a chatbot's. You are not typing questions — you are uploading whole documents, and the documents this tool exists to read are the sensitive ones: supplier bank details, payment amounts, staff expenses, and, if you use its pre-built identity-document extraction, driving licences. One upload can contain more personal data about other people than a year of chat.
One asymmetry beginners miss: under GDPR (European privacy law governing what companies may do with your data), if you upload an invoice containing a supplier's employee's name and bank details, you are usually the party answerable for it, not the tool. Nanonets publishes access and erasure rights for its own users; it does not carry your obligation for the people inside the documents.
You collect the last twenty invoices from a single supplier, upload them together, and correct every wrong field before approving any of them.
Why this one: A one-file test measures nothing, because the mechanism the tool is built on — confidence rising only as you correct and approve — has not run yet.
Tweak: Pick the supplier whose invoices are ugliest. Clean ones flatter it.
You time how long checking and fixing the extracted fields takes across ten documents, then divide by ten for a per-document figure.
Why this one: Nanonets is unusual in publishing what each run costs up front, so the machine side of the bill is knowable on day one. The minutes a person spends fixing fields before approval are the half nobody publishes, and early on they are the larger number.
Tweak: Take the same measurement again after fifty approvals. The gap between the two readings is the whole return.
You upload a phone photograph of a crumpled receipt alongside a clean scan of that same receipt, then compare the two results field by field.
Why this one: Nanonets reports a confidence score for the extraction, not for the image it started from, so a bad photograph and a hard document look identical in the output. A controlled pair is the only way to see which one you actually have.
Tweak: Then try a layout you have never sent it.
Instead of uploading a file by hand, you forward a real invoice email — attachment and all — to the address Nanonets assigns your workflow.
Why this one: Email and cloud-folder intake are both on the free plan, and they are the only test that matches how documents actually reach a back office — as attachments nobody downloaded.
Tweak: Forward one email with two attachments, and one with the invoice pasted into the body instead. The second is the case that breaks.
You raise a field's confidence requirement until documents start failing review, then follow where the review request goes.
Why this one: Nanonets raises what it is unsure about in Slack, Teams or email rather than inside its own screen, so the person doing the checking may never open the tool. Deciding who that person is before rollout costs less than discovering it after.
Want the full picture first? See the Nanonets overview for pricing, the free tier and related tools.
