Nanonets.
The quiet engine of the back office.

01 / Present TenseThe company you have already worked with, possibly without noticing

Somewhere in the bowels of a Fortune 500 finance department, at 11:42 on a Tuesday morning, an invoice for $47,302.18 lands in an inbox. Nobody opens it. A piece of software does. Within four seconds the line items have been read, matched against a purchase order, run through a fraud check, routed for approval, and posted into SAP. The person who would have done that work by hand is now in a meeting about something more interesting.

That piece of software is Nanonets. It is also, increasingly, every piece of software like it.

The company is headquartered on Market Street in San Francisco, employs about 130 people, and has raised roughly $42 million across three rounds. Its customer list reads less like a sales pipeline and more like a reading of the S&P 500. It does one thing - reads documents and acts on what they say - and it does it with a level of accuracy that would have seemed slightly mad ten years ago.

02 / The ProblemAn $8 trillion paper trail

The world runs on documents. Invoices, claims, purchase orders, bills of lading, tax forms, lease agreements, KYC packets, loan files. They arrive as PDFs, scans, photographs, faxes that somehow still exist, and email attachments named invoice_FINAL_v3 (1).pdf. Almost none of them are clean. Almost all of them must be turned into rows in a database before anything useful can happen.

For thirty years, the answer was a person with a keyboard. For twenty, it was offshore data entry. For ten, it was rules-based OCR that broke every time a vendor changed their letterhead.

The cost of this hairball, depending on whose analyst report you read, is somewhere between annoying and ruinous. Gartner has put manual document handling in the tens of billions a year for finance functions alone. Nanonets noticed this in 2017 - which is to say, slightly before everyone else did.

03 / The Founders' BetTwo repeat founders, a YC ticket, and an unfashionable idea

Sarthak Jain and Prathamesh Juvatkar were not first-time founders. They had already built, sold, and exited a content-aggregation startup called Cubeit, which Myntra picked up in 2016. They could have done anything next. They picked the part of enterprise software that nobody in San Francisco at the time considered cool: making OCR work properly.

It was a contrarian call, in a quiet way. The dinner-party AI in 2017 was image classification and game-playing agents. Reading invoices was, charitably, not the headline. Y Combinator backed them anyway, in the Summer 2017 batch.

04 / The ProductA no-code platform with deep-learning teeth

The Nanonets pitch, stripped of marketing, is this: upload a few hundred examples of a messy document, and the platform trains a custom model that reads the rest. No labeling pipeline. No data scientist. No six-month integration. The model is then wrapped in an API and a UI, plugged into whatever ERP or accounting system the customer happens to live in, and pointed at a fire-hose of incoming files.

Under the hood, the company runs a stack that would make most ML teams envious - PyTorch and TensorFlow models trained on millions of documents, transformer architectures from Hugging Face, vector embeddings, a Kubernetes-on-AWS deployment with the usual Grafana and Prometheus accompaniment. None of which the customer ever sees. They see a dashboard and a number that says how many invoices got processed last week.

The pre-trained APIs cover the boring classics: invoice OCR, receipt OCR, passport OCR, ID cards, bills of lading, purchase orders. The custom training flow handles everything else - loan files, claims forms, energy bills written in eight languages, the document your customer made up specifically to ruin your Tuesday. On top of all of it sits a workflow engine that routes, approves, validates, reconciles, and posts.

05 / The ProofNumbers, dragged into the light

Enterprise sales has a problem with proof. Everyone has a deck. Almost nobody has a control group. Nanonets has the advantage of selling into finance teams, which are - by professional disposition - allergic to vibes and partial to spreadsheets. Case studies on the company's site report 90-95% reductions in manual effort, 50% cost savings, and the kind of payback periods that procurement teams approve without a second meeting.

The customer roster is the other half of the argument. The company has not been chatty about names, in the way of B2B startups with serious enterprise contracts, but case studies cover alternative lenders, insurance carriers, large logistics operators, property managers, and at least one accounting firm that you have heard of. Partnerships with SAP, Salesforce, Microsoft Dynamics, QuickBooks and most of the major ERPs mean the platform slots into the system of record without a six-month integration project.

06 / The MissionLess paperwork, more work

The official mission statement is the kind of thing every AI company writes - unlock value, free knowledge workers, automate the boring stuff. The actual mission, as observed in product decisions, is sharper. Nanonets is building autonomous agents that handle entire back-office workflows without a human in the loop, escalating only when they cannot confidently act.

This is a more interesting bet than it sounds. It assumes that the future of enterprise software is not a better dashboard for people, but a smaller dashboard, because the agent is doing the work the people used to do. The dashboard mostly shows what the agent did and why.

07 / Why It Matters TomorrowThe agent that goes to work for you

Two industries are about to be loudly rearranged by agentic AI. The first - software engineering - already is. The second - the global back office - is just starting. Accounts payable, receivables, claims, KYC, underwriting, procurement, compliance. These are some of the largest cost centers in the modern economy, and they have spent the last forty years being not-quite-automated by waves of ERPs, RPAs, and consultants.

Nanonets is making the bet that this round is different, because the agents can actually read. The model does not need a brittle rule for every vendor; it understands what an invoice is. It does not need an exception queue every time the layout changes; it generalizes. That sounds incremental. It is not. It is the difference between automating 60% of the workload and 95%.

Back to the invoice at 11:42 on Tuesday. It is now Tuesday afternoon. The line items have been booked, the supplier has been paid, the ledger ties out. The person who used to spend three hours on that document has spent zero. Whether they think of that as a relief or a threat will depend on which year you ask them. Whether Nanonets becomes the company that quietly automates a measurable chunk of the world's clerical work - that is the question the next ten years answer.

The company is, for now, very busy not talking about it. Which, in San Francisco, is usually a good sign.