Meta AI director Rohit Patel leaves to take on Microsoft Office and Google Workspace

Rohit Patel spent eight years building AI at Meta, most recently as a director at Meta Superintelligence Labs. He's leaving to launch Ndoc, pitched as an AI-native alternative to Microsoft Office and Google Workspace. He joins a run of high-profile exits from Meta that includes Joel Kaplan, Andrew Tulloch and Emily Dalton Smith.

Two things are worth tracking here. One is the talent picture at Meta's AI operation, which keeps losing senior people. The other is the product question underneath every enterprise AI pitch: should AI live inside the tools we already use, or replace them?

We think Patel is asking the right question, even if the odds are long. Most enterprise AI today is attached to existing software. It fills cells, drafts slides, summarizes documents. Patel calls that putting lights on the bulls that till the soil when you should be building a tractor. It's a sharp line, and it lands, because a lot of AI at work right now really is the old workflow with an autocomplete.

What stands out in Ndoc is the emphasis on checking the machine's work. Each document pairs a presentation layer, a narrative layer with citations, and a model layer that holds the data and Python scripts behind the numbers. If AI is going to produce the decks executives use to make decisions, auditability isn't a nice-to-have. It's the whole game. A finance lead should be able to ask which line holds a formula and get a real answer. That's useful AI, and we'd like to see the big suites copy it.

The open-format idea matters too. Patel plans to open-source an exportable file that runs in any browser, and says that ambition is part of why he left. Lock-in has defined office software for decades. If a new format breaks that, good. If it simply becomes a new walled garden, nothing changes.

Now the hard part. This is a team of three with no funding, going up against OpenAI, Anthropic, Google, Meta and others, all of whom are wiring their models into Workspace and Microsoft 365. OpenAI's Space puts a shared workspace directly inside ChatGPT. Habits are sticky, and the incumbents own the distribution. Patel's own early users told him they didn't care much about lower cost from open-source models; they cared about having everything end to end in one place. That's a product insight, and also a warning that the moat is experience, which is easier for incumbents to imitate than it sounds.

I'm also a little wary of the speed-first philosophy. Moving fast on a suboptimal plan is fine for a startup. For a tool that generates the numbers companies act on, errors compound quietly. The verification layer has to be as fast as the generation layer, or the promise of transparency becomes decoration.

On Meta: one departure is a personnel note. A string of them, across policy, research and product, starts to look like a pattern. We'd treat it as a signal to watch rather than a verdict.

Ndoc is running an enterprise pilot and an alpha for a limited waitlist. Sign-up is free, with charges for usage and tokens, a beta is planned in the coming weeks and a full launch next year. Patel says it's built and fine-tuned on a mix of open-source models and includes connectors to import data from existing tools.