After eight years developing AI at Meta, Rohit Patel is leaving the company to challenge something far more familiar than frontier AI models: the office software millions of people use every day.

Patel, formerly a director at Meta Superintelligence Labs, has launched ndoc, a three-person startup building an AI-native alternative to traditional productivity suites. His ambition is not simply to make documents, spreadsheets or presentations easier to create. He wants to eliminate the fragmented workflows that force businesses to move information between those tools, often spending more time checking and updating outputs than producing the original analysis.

The company is already conducting an enterprise pilot and opening limited alpha access, although it has yet to disclose paying customers or commercial revenue. Patel has also chosen not to raise outside funding until he sees stronger evidence that businesses want to keep using the product.

The office workflow AI still hasn’t fixed

Imagine a finance team preparing its quarterly board presentation. Revenue figures begin in Excel, financial analysis moves into Python, the executive summary is drafted in Word, and the final numbers are transferred into PowerPoint.

AI can now assist with nearly every stage of that process, but the underlying workflow remains fragmented. When the finance team changes a revenue assumption, someone still needs to update the calculations, revise the written explanation and ensure the presentation reflects the latest figures.

For Patel, this is the problem traditional productivity software has failed to address.

In his essay explaining why he left Meta, he describes how AI has accelerated individual tasks without necessarily improving the way those tasks connect. His own working habits had already shifted away from conventional office applications toward Markdown, Python and interactive dashboards, where the underlying calculations could remain connected to the final output.

He believes office software should work that way by default, particularly as AI becomes responsible for producing more of the information companies use to make decisions.

How ndoc brings data, analysis and presentations together

ndoc approaches the problem by combining three components within a single document: Model, Narrative and Stage.

Model contains the underlying datasets, calculations and Python scripts. Narrative holds the written analysis, supporting evidence, citations and explanations, while Stage turns those findings into interactive dashboards, presentations, simulations or lightweight applications.

The purpose is to keep the finished work connected to the information that produced it. In ndoc’s intended workflow, a change to the underlying model can update the relevant analysis and presentation, reducing the need to manually reconcile figures across different applications.

That could be particularly useful for executives who work with constantly changing information. A CFO reviewing financial forecasts could inspect the calculations behind a projected result, while a sales leader could adjust conversion assumptions and explore how those changes affect expected revenue.

The more consequential feature is traceability. Rather than accepting an AI-generated chart at face value, users can inspect the Python scripts responsible for the result or ask the built-in assistant to explain how a particular figure was calculated.

For companies making financial, operational or investment decisions, the ability to understand where a number came from may be more valuable than generating another polished presentation. However, ndoc has not yet published independent evidence establishing the accuracy or efficiency gains of its approach.

Why Patel left Meta to build the product independently

Patel’s interest in verifiable AI reflects his previous work at Meta, where he contributed to research involving Llama models, reinforcement learning, model evaluation and multimodal AI.

Much of that work focused on understanding model capabilities and measuring how reliably AI systems perform. With ndoc, he is applying a similar concern to everyday business work, where an incorrect figure can affect decisions regardless of how convincing the surrounding report looks.

Patel has said he did not leave Meta with a fully developed product concept. He initially wanted the freedom to build independently and explored different ideas before focusing on the office productivity problem.

That independence also matters to his longer-term vision for the document itself.

ndoc supports exporting work into a self-contained .ndoc.html file designed to open in a browser, including offline, while Patel has discussed plans for an open format that developers can use outside the company’s platform.

He believes this approach could make business documents more portable and reduce dependence on proprietary software ecosystems. In discussing his decision to work independently, Patel also pointed to the difficulty of coordinating a new open format across a large organisation, where decisions can require agreement from many different teams.

Can three people challenge Microsoft and Google?

The opportunity is substantial, but ndoc is entering one of enterprise software’s most established markets.

Microsoft continues to integrate Copilot across Microsoft 365, while Google is bringing Gemini deeper into Workspace. Both companies already serve organisations whose documents, spreadsheets, permissions, integrations and internal processes are built around their products.

Their biggest advantage may not be AI capability alone. They already occupy the environments where employees work, making it easier to introduce new functionality without asking businesses to change their existing systems.

ndoc is taking the opposite approach by questioning whether those separate applications should remain the default way to organise work.

The competitive challenge extends beyond Microsoft and Google. Products such as Hex and Deepnote already connect computational analysis with interactive outputs, meaning ndoc must demonstrate why its broader document model offers a meaningful improvement rather than simply a different interface.

Patel is trying to lower the adoption barrier through connectors that bring information from existing tools into ndoc. This would allow teams to experiment with connected documents without immediately replacing the systems they depend on.

The startup also uses a mixture of fine-tuned open-source models, which Patel believes can reduce operating costs. No independent cost comparison has been published, although he told TechCrunch that early users appeared more interested in the connected workflow than in the underlying AI economics.

The real test begins with enterprise adoption

For now, ndoc remains an early-stage product with a limited alpha, an enterprise pilot and a small founding team. Users can sign up for free, with usage-based and token-based pricing planned as the product develops.

Patel is targeting a beta release in the coming weeks and a broader commercial launch in 2027, but the company has not disclosed the identity of its enterprise pilot customer, customer retention, recurring revenue or measurable productivity improvements.

Those details will become increasingly important as ndoc moves beyond experimentation.

Enterprise buyers will want more than an impressive demonstration of connected documents. They will need confidence in calculation accuracy, data security, permissions, collaboration, compatibility with existing systems and the reliability of AI-generated outputs.

For Patel, the immediate challenge is therefore not replacing Microsoft Office or Google Workspace across an entire organisation. It is demonstrating that a connected document can solve a persistent and expensive problem well enough for businesses to change an established workflow.

If ndoc can help an executive update a financial model, verify the analysis and share an accurate board presentation without rebuilding the same work across three applications, it will have a compelling reason for customers to pay attention.

That is the opportunity Patel left Meta to pursue, and the enterprise pilot will provide an early indication of whether the product can deliver on it.