Ndoc is the reason Rohit Patel is leaving Meta. After eight years building AI there, most recently as a director at Meta Superintelligence Labs (MSL), Patel has launched Ndoc, a startup he describes as an AI-native replacement for Microsoft Office and Google Workspace. TechCrunch reported the move as an exclusive on 27 October 2026.
The pitch is bold for a company of three people with no outside funding. Ndoc wants to pull the work that today sits across Excel, Word, PowerPoint, Tableau, Replit and copy-and-paste ChatGPT sessions into one shared document. It also wants to change the file itself, with an export format that opens in any web browser and that Patel intends to make open.
This guide explains what the product is, how its three-part layout works, why Patel built it to show its working, and how it compares with Microsoft 365 Copilot, Google Workspace with Gemini and OpenAI’s new ChatGPT Space. It ends with what Ndoc means for businesses that run on office software today, and what to test before trusting any AI-built report.
Table of contents
- What Ndoc Is and Why Rohit Patel Left Meta
- How It Works: Stage, Narrative and Model
- Why Ndoc Shows Its Work
- The .ndoc.html File and the Open-Format Bet
- Ndoc Versus Microsoft Office, Google Workspace and ChatGPT Space
- Built on Open-Source Models
- The Odds Facing a Three-Person Startup
- What Ndoc Means for Businesses
- Our Verdict
- Ndoc FAQs
- References
What Ndoc Is and Why Rohit Patel Left Meta
Ndoc is a document workspace in which an AI assistant builds the analysis, the write-up and the presentation together, and keeps the data and code that produced them in the same record. Patel calls it “an AI-native, open-format alternative to the office suite” on his personal website.
The company is very young. As of Tuesday 27 October, the company was running an enterprise pilot and opening an alpha version to a limited number of people who join its waiting list. Patel told TechCrunch he hopes to release a beta “in the coming weeks”, with a full launch planned for next year.
Eight years at Meta, ending at MSL
Patel’s own site says he “spent eight years building AI at Meta, most recently as a director at MSL (Meta Superintelligence Labs) working on our AI models.” Before Meta he worked with chief executives of private-equity-owned businesses. Speaker profiles list that earlier role as PwC Strategy&.
TechCrunch says his Meta work centred on reinforcement learning and on evaluating whether AI output can be trusted. His publication list backs that up. He is credited on Meta’s Llama 3 paper and on several evaluation benchmarks, including CRAG-MM, WearVQA, the Modality Maturity Index and GIM.
He is also known outside Meta for a long explainer, “Understanding LLMs from Scratch Using Middle School Math”, published in October 2024. His site says it became one of Medium’s most shared stories of that year. That habit of explaining the maths step by step shows up again in how Ndoc treats its own calculations.
The 27 October announcement
TechCrunch’s Rebecca Bellan broke the story at 14:59 UTC on 27 October 2026. Other outlets, including Mezha and KuCoin News, followed with rewrites of the same report. Patel’s site already links to the TechCrunch piece under “In the press”.
We have treated TechCrunch and Patel’s own website as the primary sources here. At least one follow-up article added claims that do not appear in either source, so we have left those out.
Part of a run of Meta departures
TechCrunch places Patel in “a handful of high-profile departures from Meta.” It names three: Chief Global Affairs Officer Joel Kaplan, AI researcher Andrew Tulloch and Emily Dalton Smith, who was Head of Product for AI-for-Work.
That last name is worth noting. Dalton Smith ran the product side of Meta’s push into AI at work, which is the market Ndoc now targets from the outside. TechCrunch’s report does not include a comment from Meta on Patel’s exit.
| Item | Ndoc, as reported on 27 October 2026 |
|---|---|
| Founder | Rohit Patel, former director at Meta Superintelligence Labs |
| Time at Meta | Eight years |
| Team size | Three people |
| Funding | None raised yet, by choice |
| Stage | Enterprise pilot plus a waiting-list alpha |
| Next steps | Beta in the coming weeks, full launch in 2027 |
| Pricing | Free to sign up, then pay for usage and tokens |
| Export file | A single .ndoc.html file that opens in a browser |
How It Works: Stage, Narrative and Model
Ndoc puts three connected views of the same piece of work into one layout. TechCrunch calls them tabs. Patel’s site calls them “surfaces” and says each one replaces a different tool that people currently glue together by hand.
The idea is that a reader can move from the headline chart, to the explanation, to the raw numbers and code, without leaving the document or asking someone to send the spreadsheet.
The Stage tab
Stage is where findings are presented. It can hold slide decks, dashboards, simulations or small interactive apps. Patel’s site gives the example of “scenario sliders” that a reader can drag to rerun a forecast.
In Patel’s framing, Stage replaces PowerPoint, business intelligence tools such as Tableau, and the standalone web apps that some teams now build to share a model.
The Narrative tab
Narrative is the written account. TechCrunch describes it as holding analyses, numbered citations, generated figures and “the why behind the numbers.” Patel’s site adds that it supports LaTeX for maths, and that numbers in the text link to the Python run that computed them.
This is the part that replaces Word or Google Docs. The difference is that a figure in the report is not typed in by hand. It is tied back to the calculation that produced it.
The Model tab
Model holds the underlying datasets and Python scripts. Patel’s site says they are shown “as a flow diagram of what feeds what”, so a reader can see which script reads which table and which output feeds which chart.
This is the replacement for Excel and for data-science notebooks. It is also the part that makes the product’s checking story possible, because the logic is visible rather than hidden inside cell formulas.
Sharing with people who have not signed up
TechCrunch reports that the unified workspace is one “all colleagues can access, even if they’re not signed up for Ndoc.” That matters for adoption. A tool that forces every reader to create an account spreads slowly. A document that opens for anyone spreads at the speed of a shared link.
The table below sets out what each surface holds, what it is meant to replace and the open format Patel says it is stored in.
| Surface | What it holds | Tools it replaces |
|---|---|---|
| Stage | Slides, dashboards, simulations, mini apps | PowerPoint, BI tools, standalone apps |
| Narrative | Analysis, citations, figures, LaTeX | Word, Google Docs |
| Model | Datasets and Python scripts as a flow diagram | Excel, notebooks |
| Storage | Markdown, React, Parquet, Python, JSON | Closed binary office files |
Why Ndoc Shows Its Work
The most distinctive thing about Ndoc is not the three tabs. It is the insistence that a reader can check what the AI did. That comes straight from Patel’s background in evaluation.
From model evaluation to document evaluation
At Meta, Patel worked on reinforcement learning, in which an AI system learns by trial and error to maximise a reward, and on judging whether model output is trustworthy. His benchmarks measure exactly that kind of gap. GIM uses 820 expert-written problems, and the Modality Maturity Index uses 893 questions with human-written marking rubrics.
TechCrunch sums up his view this way: if AI is going to build the spreadsheets and slide decks that companies use to make decisions, someone should be able to check its work. Ndoc is his attempt to build that check into the file.
“Nothing hidden behind the curtain”
In Ndoc, users can open the Python scripts used to create a document. They do not need to read Python to do so. “Even if you don’t know any Python, you can ask [the AI chat] to explain, ‘Hey, what does this script do? Which line has my formula?’ And you can verify that formula,” Patel told TechCrunch. “There’s nothing hidden behind the curtain here.”
That is a sharp contrast with the usual AI-in-a-spreadsheet experience, where a model fills cells and the user has to trust that the numbers are right or rebuild them to find out.
Memory at four levels
Users can upload a dataset, ask the assistant to analyse it and have it write a narrative about the findings. They can also give the document its own context, including audience, tone and memory.
“It maintains a memory across your person, that particular document, your team, and your company,” Patel said. In practice, that should mean a quarterly report written in Ndoc can pick up the team’s house style and the company’s standard definitions without being told each time.
What inspection does not guarantee
Visible code is a big step forward, but it is not the same as correct code. A script can be readable and still use the wrong date range, drop rows silently or join two tables on the wrong key. An AI that explains a formula can also explain a wrong formula confidently.
So the value of this approach depends on people actually looking. The design makes checking cheap. It does not make it automatic, and businesses that adopt it should still decide who reviews an AI-built analysis before it reaches a board pack.
The .ndoc.html File and the Open-Format Bet
Patel’s bigger ambition is to reinvent the file. TechCrunch reports that any Ndoc document can be exported as a single “.ndoc.html” file that opens in any browser.
One self-contained file
A self-contained HTML file is a clever choice. Every computer already has a program that can open it, and it can carry text, charts, interactive controls and data in one place. Nobody needs Ndoc installed to read one.
Patel’s site says each part of a document is stored “in an open format AI reads and writes fluently (Markdown, React, Parquet, Python, JSON).” Those are formats that today’s LLMs already read and write well, which is the point. The AI is working in its native languages rather than in a proprietary spreadsheet format.
Apache 2.0, stated but not yet visible
Here the two sources differ slightly. TechCrunch says Patel “plans to open-source” the format so other developers can build apps that work with it. His website goes further and says the .ndoc.html spec “is open under Apache 2.0, so your work outlives any tool, ndoc included.”
We could not find a published Ndoc specification on 27 October. Patel’s public GitHub account shows no repository for it, although it does include a fork of the Milkdown Markdown editor created on 16 October. Until the spec is published, treat the Apache 2.0 licence as a stated intention.
Why open formats matter
Lock-in to office formats is old and expensive. The two big open document standards, OpenDocument (ISO/IEC 26300) and Office Open XML (ISO/IEC 29500), exist largely because governments and businesses wanted their records to outlive any one vendor. Even so, most organisations still depend on Microsoft or Google to open their own files.
TechCrunch notes that Patel wants developers to work with the format “without being locked into Ndoc the way so many have been locked into Microsoft or Google for years.” An open Ndoc format would let a rival build an editor, a viewer or an archive tool, which is a strong promise for a startup to make about its own product.
Why he left to do it
The open format is partly why Patel went independent. At a company of Meta’s size, he told TechCrunch, agreeing an open-source format would need the input of “hundreds of people and opinions on how the format should look.”
He wanted to move faster than that. “Humans can only move at human pace, whereas … I can move at the pace of what AI can do,” he said.
Ndoc Versus Microsoft Office, Google Workspace and ChatGPT Space
Ndoc is entering a market where the biggest AI labs and both office giants are already moving. TechCrunch lists OpenAI, Anthropic, Meta, SpaceXAI and Google as having enterprise AI tools that connect to Google Workspace, Microsoft 365 and other work apps.
Adding AI to the incumbents
Microsoft and Google have taken the obvious route: put AI inside the tools people already use. Microsoft sells Microsoft 365 Copilot as an add-on to a qualifying Microsoft 365 licence, at a list price of $30 per user per month for enterprise plans and $21 for the Copilot Business plan aimed at organisations with up to 300 users.
Google has bundled Gemini into Workspace instead. Its US pricing page lists Business Standard at $14 per user per month, with the AI assistant in Gmail, Docs and Meet included. Business Plus is $22. For a business already paying for either suite, the question about Ndoc is whether it saves enough work to justify another tool.
The cost gap between the incumbents’ AI options is large. The chart below compares their published US list prices per user per month.
For a 50-person team on annual terms, that is $18,000 a year for the enterprise Copilot add-on (50 × $30 × 12), $12,600 for Copilot Business (50 × $21 × 12) and $8,400 for the whole of Workspace Business Standard (50 × $14 × 12). The Copilot figures sit on top of the base Microsoft 365 licence. Ndoc has not published usage prices yet, so it cannot go on the chart.
ChatGPT Space
The more direct threat is OpenAI. At DevDay at the end of September, OpenAI launched ChatGPT Space, a shared workspace inside ChatGPT where colleagues and ChatGPT work on the same pages and files. It is available on the Pro, Business and Enterprise plans, and on mobile it is read-only for now. We covered the early leak in OpenAI Spaces: Smart, Surprising Bet on Human-Agent Teamwork and the launch in our DevDay keynote report.
Patel accepts that Space could be a threat. He also thinks the labs have mostly got the wrong idea about what AI should do to office work.
“Get a tractor”
“Companies are basically attaching AI to spreadsheets and it fills your cells for you,” Patel told TechCrunch. “It just sort of kills my soul. This is like if you just figured out electricity, and then you decided the best use case is to put lights on my bulls that till my soil. No. Get a tractor.”
It is a good line, and it describes Ndoc’s bet precisely. The incumbents are improving the bulls. Ndoc is trying to build the tractor, a new kind of document designed around what AI does well, rather than an AI helper living inside a decades-old document model.
| Factor | Ndoc | Microsoft 365 Copilot | Workspace with Gemini | ChatGPT Space |
|---|---|---|---|---|
| Approach | New AI-native document | AI added to Office apps | AI added to Google apps | Shared pages inside ChatGPT |
| File format | .ndoc.html, open spec promised | Office Open XML | Google cloud files | ChatGPT pages |
| Models | Fine-tuned open-source mix | Microsoft-hosted frontier models | Gemini | OpenAI models |
| Code visible | Yes, Python in the Model tab | Partly | Partly | Partly |
| Status | Pilot and alpha | Generally available | Generally available | Launched September 2026 |
Built on Open-Source Models
Ndoc does not rely on a single frontier model from OpenAI, Anthropic or Google. Patel says it is built and fine-tuned on a mixture of open-source models.
Built from scratch to cost less
Because the platform was built from scratch, Patel says it costs considerably less to run than alternatives built on frontier AI models. For a startup charging by usage and tokens, a lower cost of serving each request is a real advantage. It leaves room to undercut on price or to spend more compute per document.
The open-weight route also fits his Meta background. Meta has been one of the biggest backers of open-weight models, a strategy we looked at in Meta’s open model and its Chinese open-weight rivals.
What early users actually cared about
The lower running cost was not what won over Ndoc’s first testers. “I thought that’s what people would be excited about, but people don’t care about that,” Patel said. “What people are excited about is that they love seeing all their end-to-end in one place.”
That is a useful signal. Office users do not buy model architecture. They buy fewer copy-and-paste steps between a dataset, a chart and a slide.
Connectors for existing data
Patel knows that habits are hard to change, which is why the AI labs have tried to bring their models to where people already work. His answer is connectors that let customers import existing data from the tools they use today. A team can bring a spreadsheet or a database export into Ndoc rather than rebuilding it.
The Odds Facing a Three-Person Startup
TechCrunch is blunt about the scale of the challenge. Trying to become the place where enterprise work happens is “a huge bet for a founder with a team of three and no funding,” especially when the big AI labs want the same territory.
No funding, by choice
Patel has not raised money and says that is deliberate. “I didn’t want to take money until I had some proof that people love the product,” he said. The enterprise pilot and the alpha are his way of getting that proof before talking to investors.
His bar for the product is personal. “My bar was that I’m not going to release anything I don’t love using,” he said. He added that he has “a graveyard of tools” that failed that test.
Speed over a perfect plan
Patel’s operating philosophy is to move fast. “Execution is extremely cheap … so go execute on a suboptimal plan and worry about it later,” he said. “By the time you make the best plan, the execution will be done by somebody else.”
That approach suits a small team using AI to write much of its own code. It is riskier for a product that asks businesses to trust it with financial models and board reports, where reliability counts for more than release speed.
The incumbents’ advantages
Microsoft and Google own the files, the identity systems, the compliance paperwork and the procurement relationships. Most large organisations will not move their records to a new format without security reviews, data-residency answers and a long track record.
Ndoc’s best route in is therefore probably not a full replacement. It is the analytical report: the piece of work that already starts in a spreadsheet, passes through a notebook and ends in a slide deck, where the glue work is most painful.
What Ndoc Means for Businesses
For most organisations, Ndoc is something to watch and perhaps to test, not something to switch to this quarter. Here is how to approach it, and AI-built documents in general.
Start with low-stakes reporting
If you join the Ndoc alpha, pick a recurring internal report with known answers, such as a monthly sales summary. Rebuild it there and compare the results line by line with your existing version. That tells you quickly whether the analysis is right and whether the Model tab makes checking easier. Our data visualisation services team uses the same approach when moving reports between tools.
Ask where the data goes
Before uploading client or employee data to any new AI tool, ask where it is hosted, how long it is kept, whether it is used for training and which subprocessors are involved. TechCrunch’s report does not cover these points for Ndoc, so get the answers in writing. UK organisations will need them for their UK GDPR records.
Test the export
The .ndoc.html file is Ndoc’s biggest promise. Export a test document, open it on a machine with no account and check that the charts, interactive controls and data all still work. If the spec is published under Apache 2.0, check that too.
Get more from the suite you already pay for
Most businesses will get faster gains from using the AI already in their Microsoft 365 or Google Workspace licences well. If you are weighing Copilot, our Copilot readiness assessment checklist covers the permissions and data clean-up to do first. For a wider plan, see our AI strategy service.
Our Verdict
Ndoc is one of the more thoughtful attacks on the office suite we have seen. It is not another chatbot in a sidebar. It treats the document itself as the thing to redesign, keeps the AI’s working visible and promises an open file that outlives the company.
The weaknesses are just as clear. A three-person, unfunded team is taking on Microsoft, Google and OpenAI at once. The open specification is not yet public, prices are not published and the product is an alpha. Those are normal for a launch week, but they mean Ndoc is a test-and-watch tool for now.
Rohit Patel’s departure is also a sign of the times. Senior AI people at the biggest labs now believe a tiny team with good models can build a serious product quickly. Whether Ndoc proves that will become clearer when the beta arrives.
Ndoc FAQs
What is Ndoc?
Ndoc is an AI-native document workspace founded by former Meta AI director Rohit Patel. It combines presentation (Stage), written analysis (Narrative) and data plus Python code (Model) in one document, aimed at replacing tools such as Excel, Word and PowerPoint.
Why did Rohit Patel leave Meta?
Patel wanted to build an open document format and move quickly. He told TechCrunch that agreeing an open-source format at Meta would involve hundreds of people, and that he could move at the pace of AI on his own.
Is Ndoc available now?
Ndoc is in an enterprise pilot and a limited alpha for people on its waiting list. Patel expects a beta in the coming weeks and a full launch in 2027.
How much does Ndoc cost?
Signing up is free. Customers then pay for usage and tokens. The company had not published its usage prices when it launched on 27 October 2026.
Is the .ndoc.html format open source?
Patel’s website says the spec is open under Apache 2.0, while TechCrunch says he plans to open-source it. We found no published specification on launch day, so treat it as a stated plan.
Should my business switch to Ndoc?
Not yet. Test it on a low-stakes internal report, check data handling and the export, and keep using your existing suite for core work until Ndoc reaches a full release.
References
Meta AI director Rohit Patel leaves to take on Microsoft Office and Google Workspace (TechCrunch)
Rohit Patel: personal website and publications
Understanding LLMs from Scratch Using Middle School Math
OpenAI launches ChatGPT Space for team collaboration (IT Brief UK)
Microsoft 365 Copilot for business
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
