Rebuilding Office with AI: what already works and what I keep asking myself

Office has been getting on my nerves for years. Not because it's bad, but because in many places I want to work differently than it lets me. So the idea of building my own suite has been around for a while. I even started on it, then it stayed a side project. It was simply too big.

Then came the wave of posts claiming that AI replaces every developer and that building software has become easy. I wanted to find out for myself. So I picked the project up again, this time with AI as the tool and in public, so anyone can follow along. Today there are eight apps: documents, spreadsheets, presentations, forms, reports, email, a content planner and a PDF tool. Around 456,000 lines of code in total, a little over 83,000 of them tests.

I'm not the only one with this idea. ArtCraft is making the rounds right now: a developer used AI to rebuild first Adobe's apps and then Word, Excel and PowerPoint. Whether that involved decompiling is hotly debated. He himself calls it a rebuild from public file format specifications. For Balane Office I can say it plainly: nothing was decompiled, nothing was copied. Open-source suites like LibreOffice serve as a reference for scope and behaviour, but none of their source code is used.

The idea: similar in spirit, not the same

PowerPoint didn't invent the presentation. Slides, projectors and talks existed long before. PowerPoint turned a familiar idea into software. That's how I see Balane Office: a suite that solves the same tasks as Office, but decides differently in the places where Office has bothered me for years.

The principles are quick to explain:

• Local first. Every file lives on your own device and can be fully edited offline. No mandatory account, no server of mine in between.
• Standalone apps, one way of working. Each app works on its own. Terms, shortcuts, home screen and command search are the same everywhere.
• Honest compatibility. For every format, a matrix lists what survives opening, editing and saving back. Anything untested doesn't count as supported.

Why I built it this way

The features that matter most to me are the ones I've often missed in Office.

The AI never changes anything silently. It makes proposals: each change on its own, with before and after, with the cells or paragraphs it relies on, and a short reason. I accept or reject. Accepted changes can be undone in one step. In spreadsheets it writes real formulas instead of finished values, so the result stays traceable.

The AI is replaceable. It stays off until you add your own key. Today that's OpenAI; later you should be able to plug in the provider of your choice. There is no server of mine that your data passes through. Why that matters to me is something I wrote about in my analysis of AI pricing (in German): whoever builds today should build so they can switch providers.

The apps are connected. A range from a spreadsheet can go into a report, a document or a presentation and be refreshed there. Anything I changed by hand in the target document is never silently overwritten.

Other tools can join in. Most apps come with an interface for AI assistants (MCP). An assistant like Claude can read and edit files through it, with versions and a change log.

Why it isn't finished software

This belongs on the table just as much:

• Platforms: Development and visual checks happen on macOS and in the browser. Windows, iOS and Android haven't been built or tested yet.
• Office files: DOCX, PPTX and ODT can be opened and saved, but with documented limits. LibreOffice opens the files without repair; the cross-check in Word, Excel and PowerPoint is still open. With XLSX, charts, images and macros, among other things, are lost when saving back.
• File format: The suite's own working formats may still change. Important data shouldn't live only in these apps.
• Scope: Email, the content planner and the PDF tool are clearly behind the other apps.

Anyone trying Balane Office is trying an experiment. That's exactly how the website puts it, too.

What already works surprisingly well

Even so, I was surprised how far some of the apps have come:

• Spreadsheets: around 230 functions in German and English, sorting, filters, pivot tables, conditional formatting, data validation and seven chart types.
• Reports: query steps for merging, grouping and unpivoting, relationships between tables, measures with time intelligence, drillthrough and a separate phone layout per page. That's closer to Power BI than to Excel.
• Forms: 21 question types, conditional logic, drafts, kiosk mode and analysis.
• Presentations: masters and layouts, transitions, animations and presenter view.
• Documents: its own editor with styles, footnotes, a table of contents, comments and track changes.

Every app is backed by automated tests, 390 test files in total. Without them I wouldn't know after any change whether everything still works.

The question that worries me more than the code

Software is already highly customisable. Menus, views, colours, shortcuts, in many programs even your own extensions. But underneath, it's almost always the same: the same file formats, the same structures. An Excel file from Hamburg opens in Munich because both sides speak the same format. Those formats were standardised over years: Office Open XML as ECMA-376 and later ISO/IEC 29500, the OpenDocument Format as ISO/IEC 26300.

If AI makes building software so cheap that everyone can build their own, software becomes more personal. That's a good thing at first. The question is what happens to everything underneath. If every tool does its own thing, you end up with a thousand private formats, and every file becomes an island.

I have to include myself here. Balane Office itself saves in its own formats: .bdoc, .bsheet, .bslides and so on. There are good reasons for that. The formats are simple, versioned and can evolve cleanly. For exchange, every app reads and writes the standards: DOCX, XLSX, PPTX and the OpenDocument formats. Between the apps, data moves through small, versioned contracts.

But that's exactly the problem: the standard slides into being an export format. It becomes a translation layer instead of the basis you work on. And translation loses things. My own tests show it: even with the full specification at hand, it's hard to save a file back so that everything survives.

I know the strongest counterargument: if AI can read any format and translate it into any other, maybe standards matter less than they used to. A converter is quick to build these days. There's something to that. But it doesn't help much if nobody defines anymore what a paragraph, a formula or a slide actually is. Then the AI translates between formats it invented itself, and nobody can check whether something got lost along the way.

Personal software is good. Personal file formats are a price you only notice when you exchange files.

Why knowing how to code still helps

The AI writes a spreadsheet faster than I can describe one. It doesn't tidy up, though. Left alone, it solves the same problem slightly differently every time. Before long there are several versions of the same thing, and nobody knows which one is meant to win.

What helps is surprisingly old-fashioned. There are fixed patterns in the sense of the book "Design Patterns" by Gamma, Helm, Johnson and Vlissides from 1994. They're defined once and reused everywhere. There are ten written architecture decisions, shared building blocks for all apps and an automated check that flags any dependency pointing the wrong way. The AI works within these rules. Setting them up means understanding why they exist.

And then there's the day the AI fails. A provider has an outage, prices go up, or the model just doesn't get a problem. Then I have to be able to read and change the code myself. That only works if it has stayed readable.

To be honest: writing code myself takes me much longer today than it used to. You lose practice when you spend most of your time reading and reviewing instead of typing. But I can still read code, find bugs and judge a structure. And that's exactly the skill that matters now.

AI took over the typing. Not the deciding.

What comes next

The apps run as a demo in the browser at balane.online. If there's enough interest, they'll go to the stores for Windows, Mac, iOS and Android. Until then I'll keep reporting here in the Balane Office category, including what doesn't work.

There's one answer I'm most curious about: which feature have you been missing in Office for years?

Sources

• PetaPixel: Someone Rebuilt Free, Open-Source Versions of Photoshop, Premiere, and Lightroom Using AI, 7 October 2026
• Windows Central: "Begun, the clone wars have": Microsoft 365 apps have been copied by the AI dev who cloned the Adobe suite
• Design Week: ArtCraft’s Adobe Suite copy was made with AI
• Ecma International: ECMA-376 Office Open XML File Formats
• OASIS: OpenDocument Format (ISO/IEC 26300)
• Erich Gamma, Richard Helm, Ralph Johnson, John Vlissides: Design Patterns. Elements of Reusable Object-Oriented Software, Addison-Wesley 1994