Why doesn't a company just build its own Office?

On LinkedIn I wrote: “Microsoft paid for my Office.” That was a provocation, and it was meant as one. Microsoft has put billions into OpenAI and Anthropic. With Anthropic's Claude, I built myself an office suite. It cost me my subscription and about 10 hours of work.

Which raises a question nobody would have asked seriously two years ago: why does a company pay per-seat licences every year when a rebuild costs less than a weekly subscription?

A LinkedIn post can ask that question. It can't answer it. That's what I'm trying to do here, as fairly as I can. I have a stake in the answer, because I'm building Balane Office. So I'm putting my numbers on the table and testing the objections to my own argument as carefully as the case for it.

What has changed: code is cheap

From 7 to 16 October 2026 I worked on Balane Office with Claude Code. My actual working time was about 10 hours, spread over those days and alongside my regular work. The result is eight apps: documents, spreadsheets, presentations, forms, reports, email, a content planner and a PDF tool. Around 456,000 lines of code, a good 83,000 of them tests. It's a first version, not finished software. But it's far more than I would have thought possible in that time.

Claude Code keeps its logs locally on my machine. I went through them:

The most interesting number isn't the price. It's the ratio. For every token the AI wrote, it read more than a thousand: existing code, rules, building blocks. That wasn't an accident, it was the trick. At the start I defined standard building blocks: the editor frame, menus, saving, command search, settings. Every new app is assembled from them instead of reinventing everything. The AI mostly assembled and adapted. It invented very little.

That's the real shift. Software used to be expensive because someone had to write every line. That part now costs almost nothing. What this means for a company, though, depends on a different question: what exactly is it paying for with its Office licence?

What the licence really buys

Since 1 July 2026, Microsoft 365 Business Standard costs 14 dollars per user per month in the US on an annual commitment, up from 12.50. For a company of 50 people that's a little over 8,000 dollars a year. Microsoft 365 Copilot costs extra on top, per seat. At launch Microsoft charged 30 dollars a month for it, in addition to the licence.

What does that money pay for? If you only look at Word, Excel and PowerPoint, you see programs that can now be rebuilt. But the licence contains more:

• Maintenance. Updates for new operating systems, new browsers, new devices. Year after year.
• Security. Office files are a popular attack vector. Someone has to find and close holes before others exploit them.
• Compatibility. The file a customer sends you looks the same on your screen as on theirs.
• Operations. Mail servers, storage, sign-in, permissions, backups, Teams.
• Accountability. When something goes wrong, there's a contract and someone to call.
• Familiarity. Every new hire can use Word. That saves training, and it's real value.

None of this gets cheaper because code gets cheaper. So the question becomes more precise. Not “can a company rebuild Office?” but “can it carry these six things itself, and is it worth it?”

That's what the three objections I expect most to my post are really about.

Objection 1: “We can't maintain it.”

This is the strongest objection, and it's justified.

For decades, studies have shown that most of the cost of software comes after it ships, not while it's built. Estimates vary widely depending on the kind of software and the method. A commonly cited range is that half to three quarters of the total cost goes to maintenance and further development. There's no single reliable recent study on this. But the direction isn't in dispute: the launch is the cheap part.

What does maintenance mean for an office suite in practice?

• Apple, Microsoft and Google change their operating systems every year. Something always breaks.
• Programs that open files from outside are an entry point. Every hole has to be closed quickly.
• Users find bugs no test ever covered. Someone has to fix them.
• And someone has to understand the code when the person who built it has moved on.

Does AI help with that? Yes, a lot. Finding bugs, writing tests, migrating to a new version: all of that is much faster with AI than it used to be. But my experience is also this: AI writes fast, it doesn't tidy up. Left to itself, it solves the same problem differently every time. Without fixed rules, tests and someone who can read and judge the code, you grow a mess that eventually nobody can maintain, not even an AI.

So maintenance gets cheaper, but it doesn't go away. And it needs people in-house who understand what the AI is doing. For many small and mid-sized companies, that's exactly the bottleneck: not the money for the licence, but the people to take responsibility.

My verdict: the objection holds. “For now,” I wrote in the post. I stand by that, but the “for now” is about cost. The responsibility stays.

Objection 2: “Our customers send Word files.”

In the post I answered: the format is an open standard. That's true, but it's only half the story.

The format behind DOCX, XLSX and PPTX is called Office Open XML. It's standardised as ECMA-376 and as ISO/IEC 29500, and it's publicly documented. Anyone may read and write it. It's not a secret.

The problem is something else. The standard has two variants. The strict variant was meant for new documents. The transitional variant also carries legacy baggage from older Office versions so that old files keep working. Word still writes the transitional variant today. In practice you don't just have to master the standard, you also have to master what Word actually writes.

And even then: reading a file correctly doesn't mean displaying it identically. The standard describes what's in a document. Where exactly a line breaks, when a page ends, how a missing font is substituted: that's up to the program. That's why a Word document sometimes looks different in other programs even though everything was read correctly. LibreOffice has been working on this for decades, and it's still not perfect.

I see this every day with Balane Office. Simple documents make the round trip well. With XLSX, charts and macros get lost, for example. That's why there's a format matrix showing, feature by feature, what works and what doesn't.

For a company, this means:

• Outgoing documents aren't a problem. Quotes, invoices and reports usually go out as PDF anyway.
• Internal documents aren't a problem, as long as everyone uses the same program.
• Incoming documents from customers are the problem. Especially when they need to be edited and sent back, with tracked changes, comments and complex layouts.

My verdict: the objection partly holds. The format is open. Opening it faithfully is work, but not a secret. How much it matters depends on how much your company works with other people's Office files. For a law firm sending contract drafts back and forth, it's decisive. For a tradesman who sends quotes as PDF, it barely matters.

Objection 3: “The risk is too high.”

This objection is almost always true. The only question is: which risk?

The risk of building your own is easy to see. The software could have bugs, lose data, be vulnerable. The person who built it could leave. Those are real risks.

The risk of dependency is harder to see, because it feels so normal. It has three parts.

Price. If you depend on someone, you pay what they ask. Microsoft raised Microsoft 365 prices on 1 July 2026, by 12 percent for Business Standard. That's not an accusation. Every company sets prices the market will bear. But if you can't switch, you have no leverage.

Law. In June 2025, the legal director of Microsoft France testified under oath before the French Senate that he could not guarantee French customers' data would not be handed to US authorities if they demanded it. Part of the background is the CLOUD Act, which lets US authorities access data held by US companies regardless of where it's stored. Microsoft stresses that it reviews such requests, rejects unfounded ones, and that no such case has occurred so far.

Politics. In May 2025 the Associated Press reported that, after US sanctions, the chief prosecutor of the International Criminal Court had lost access to his Microsoft email account. Microsoft denies having cut off services to the court. What exactly happened is still not fully clear. In October 2025 the court confirmed it is moving from Microsoft Office to openDesk, an open solution from Germany.

For a small business in rural Bavaria this may sound far away. But that's not the point. A risk you can't control isn't smaller just because it rarely happens.

My verdict: the objection holds, but in both directions. If you only assess the risk of building and set the risk of dependency to zero, your math is wrong.

Who is already switching, and how

Some organisations are moving away from Microsoft Office. What's interesting is how they're doing it.

None of the three built their own office suite. All of them are using existing open software that many maintain together. That's the most important distinction in this whole debate: becoming independent doesn't mean building everything yourself. It means owning the software, or at least being able to leave.

The numbers from Schleswig-Holstein are worth a second look, too. The migration costs money, and not a little. But on the state's own figures, it pays for itself in under a year. Even if those numbers are optimistic, the order of magnitude shows that licence costs are a real lever.

The dependency just moves

Now for the uncomfortable part, and it concerns my own argument.

I built my office suite with Claude, an AI from the US. If you build your own with AI, you free yourself from the software vendor and tie yourself to the AI provider. The big ones are in the US. And Microsoft holds stakes in both major providers, OpenAI and Anthropic. So the dependency doesn't disappear. It moves one level down.

Then there's price. My subscription costs under 50 dollars a week. The same tokens would have cost about 1,500 dollars at list price, thirty times as much. Whether a subscription at this price lasts isn't my call. In an earlier piece (in German) I calculated what the same work costs with different providers: 218 dollars with DeepSeek, 7,926 dollars with OpenAI. Nobody knows where those prices are heading. 26 people debated it, and there was no consensus.

There is one important difference, though. If Microsoft raises its prices, you pay more, or you can't work. If an AI provider raises its prices, your homegrown software keeps running. The code is yours. Only further development gets more expensive. That's a different kind of dependency: it affects your future, not your operations.

And you can make it smaller. Keep the AI provider swappable. Use open models you can run yourself or in Europe. And above all: build the code so that a human can understand it. Then the AI is a tool you can replace, not a supplier without whom nothing works.

When building your own pays off

So the honest answer to the question in the title is: in most cases, a company shouldn't build its entire office suite itself. But it should ask the question again, because the math has changed. There are three paths, and most companies will mix them.

AI changes the right-hand column most. The edges of an office suite, the parts that fit your company specifically, used to be too expensive for a custom solution. A form that writes straight into your order system. A report generated from your data every Monday. A template that checks your quotes. Until now you cobbled those together with Office macros or bought expensive industry software. Today you can build them, in days rather than months.

Before you build, five questions help:

1. Who understands the code if the AI is unavailable? If nobody, don't build.
2. Who will look after it in three years? A name, not a department.
3. How many outside files do you need to edit? The more, the more you should buy.
4. What format is your data in? Open formats keep every path open.
5. What happens if AI gets ten times more expensive? If the answer hurts, plan differently.

What my own experiment shows, and what it doesn't

Balane Office is a first version. It shows that one person with AI can build an office suite in a few days that you can touch, test and criticise. It shows that standard building blocks and firm rules make the difference between code that grows and code that sprawls.

What it doesn't show: that this software survives five years of daily use in a company. That it opens every customer's Word file faithfully. That it's secure enough for confidential data. Those are exactly the things you buy with a licence. And exactly the things I still have to prove.

That's why the post is pointed and this article isn't. Both are meant honestly.

Conclusion

The excuse that your own office suite is too expensive no longer holds. Building costs almost nothing today.

The other reasons still hold, and they deserve to be taken seriously: maintenance needs people, compatibility needs work, and every option carries risk, including the comfortable one. If you weigh that honestly, you'll rarely end up building everything yourself. But you'll often end up with a mix nobody would have considered two years ago: an open core, your own tools around the edges, and data in formats that belong to you.

Microsoft didn't really pay for my Office. But the tools that let you break away from Microsoft now come, among others, from companies Microsoft has invested in. What your company makes of that is now a real decision, not a question of budget.

And this isn't just about Microsoft. Every piece of software billed per seat now has to face the same question: are we paying for something we could build ourselves by now, or for something we genuinely don't want to carry ourselves?

What's still keeping your company with Microsoft or any other vendor? I'd love to hear your answers on LinkedIn.

Related

• Is AI eating its own makers? I rebuilt Office
• Make vs. Buy: Custom Development or Standard Software?
• AI: more expensive or cheaper? 26 positions, no consensus (in German)

Sources

• My own analysis: Claude Code logs for Balane Office, 7 to 16 October 2026. 5.59 bn tokens read from cache, 63 m written to cache, 4.1 m output.
• Anthropic: Pricing, Claude Opus 5.5 list prices, retrieved 16 October 2026
• The Motley Fool: 203 Billion Reasons Why Microsoft Is a Buy in 2026, on the roughly 13 billion dollar investment in OpenAI
• Nasdaq: Microsoft and Nvidia deals push Anthropic toward 350 bln valuation, November 2025
• Office Watch: Microsoft Announces Global Microsoft 365 Price Increases
• derStandard: Microsofts KI-Copilot kostet 30 Dollar im Monat (in German)
• Université catholique de Louvain: LINGI2252 Introduction, overview of published estimates of the maintenance share
• ISBSG: Lessons Learned from the ISBSG Maintenance & Support Dataset
• Library of Congress: Office Open XML, ISO/IEC 29500, on the strict and transitional variants
• Microsoft: Office's support for ISO/IEC 29500 Strict
• heise online: Not sovereign: Microsoft cannot guarantee the security of EU data, July 2025
• The Boston Globe / AP: Trump's sanctions on ICC prosecutor have halted tribunal's work, 15 May 2025
• heise online: Microsoft denies mail blockade at the International Criminal Court
• TechSpot: International Criminal Court replacing Microsoft Office, October 2025
• Borncity: LibreOffice in Schleswig-Holstein: Fast 80 Prozent der Arbeitsplätze umgestellt (in German)
• heise online: From Word and Excel to LibreOffice: Danish ministry says goodbye to Microsoft, June 2025