Rohit Patel spent eight years building AI at Meta and has left to start Ndoc, a company of three people with no funding. It wants to replace the mix of Excel, Word, PowerPoint and copy-pasting from ChatGPT with one workspace. An alpha is opening to a limited waiting list, and a full launch is planned for 2027.
The product is early, and it may stay small. The idea inside it is the part worth your attention.
A cell hides its working, and a script does not
A spreadsheet shows you the result of a calculation, not the reasoning. To see why a number is what it is, you click its cell and read the formula, then do the same for the next cell, across thousands of them. That is how an average that skipped five countries' rows, in a paper that politicians cited in 2010, stayed hidden for three years until a graduate student got the file and went through it.
AI that fills cells for you repeats that problem on a larger scale. It writes formulas faster than any person can read them, and each one looks as plausible as the last.
A script is easier to check, because the steps are written in order. One line picks the rows, another takes the average. Ndoc keeps its scripts in a Model tab, the written analysis with numbered citations in a Narrative tab, and the slides and dashboards in a Stage tab. A claim on a slide leads back to a paragraph, and the paragraph leads back to the line of code that produced it. Its chat will explain any line in plain words, so you do not need to know Python to ask which line holds your formula.
None of this is new in itself. Data scientists have long used notebooks that hold code, results and explanation in one document, and ChatGPT already writes Python for an analysis and lets you review the code. What changes is who gets the benefit. When the AI writes the script, a finance manager gets the working a data scientist always had, without writing any of it.
The job moves from building to vouching
Take a finance manager at a regional distributor who builds a monthly margin report. She collects the sales exports, removes the returns, works out the margin for each product line and puts the results into slides. Building it takes her most of a day. With an AI doing the build it takes minutes, but the day does not disappear. It becomes the checking.
Checking is faster than building, because she only has to read what the script does. She asks which line handles returns, reads that one line, and confirms the average covers every product line. I would estimate that at an hour or less. The skill her team needs changes with it: less time making the report, and more time able to vouch for it. Someone who cannot see the working cannot do that, however good the AI is.
The file is the other half
Patel also wants to change the file. Ndoc exports everything as one ".ndoc.html" file that opens in any browser, and colleagues can view the work without signing up. He plans to open-source the format so other developers can build on it.
That matters because files are what hold people in an office suite. Microsoft's old.doc format was closed, and other programs struggled to read it. The newer.docx is technically open but hard for rivals to support fully. An open alternative, OpenDocument, has been an international standard for about twenty years, and most offices still run on Microsoft's formats. An open format alone has never beaten habit, which is probably why Patel also built connectors that import data from the tools people already use.
So I would not expect three people to displace Office. I would expect the big labs to copy the habit. OpenAI launched its own shared workspace in September, and once buyers start asking to see the working behind an AI-made report, every office AI will offer it. Until then, ask for it yourself. When AI builds a report your team will rely on, ask for the script and a plain explanation of it, and name the person whose job is to read them.