The agent model doesn’t replace Office. Eventually It will make the bundle unnecessary.

The agent model doesn’t replace Office. Eventually It will make the bundle unnecessary.

I’ve spent three decades inside the machinery of business productivity software. I was in Office Product Management at Microsoft. I was a founding member of the Bing product team. I helped build Xero’s accounting platform. I shipped collaboration tools at Cisco Webex, Amazon Chime, and Amazon WorkDocs.

I’ve watched every wave of “Office Killers” come and go. Google Docs. Notion. Quip. Zoho. None of them ever cracked the fortress. What’s happening right now is different.

Two Technologies Colliding

OpenClaw is an open-source AI agent framework created by Peter Steinberger. It went from a weekend hobby project to 60,000 GitHub stars in 72 hours. OpenClaw runs on your machine and connects AI models to everything you already use: Slack, Teams, WhatsApp, Discord, Gmail, your browser, your file system, and your calendar. It doesn’t replace your tools. It orchestrates them.

Cowork is Anthropic’s desktop agent, launched in January 2026. You point it at a folder of files and describe what you want done. It reads what’s there, plans the steps, and executes. It creates documents, processes data, organizes information, and browses the web. No coding required. Describe the job, and it gets done.

Each one is interesting on its own. Together, they represent something I’ve never seen in 30 years: a credible architectural alternative to the productivity suite itself.

Why This Time Is Different

Every previous Office challenger made the same mistake. They tried to build a better word processor. A better spreadsheet. A better slide deck. They competed feature-for-feature on Microsoft’s home turf, and Microsoft’s distribution advantage crushed them every time.

OpenClaw and Cowork don’t play that game. They don’t try to replace Word or Excel. They make the question of which app you’re using irrelevant.

Think about what a knowledge worker actually does. Nobody wakes up and says, “I need to use a spreadsheet today.” They say, “I need to pull together the Q3 numbers, cross-reference them with the forecast, and get a summary to my VP by Thursday.” The spreadsheet is just a means to an end.

When an AI agent can read your files, understand your intent, pull data from multiple sources, and deliver the output wherever it needs to go, the application layer starts to dissolve. You don’t need a suite. You need an agent with connectors.

The shift isn’t “better apps.” It’s “fewer apps, more agents.”

The Suite Isn’t Equal, and Everyone Knows It

Here’s something people inside Microsoft know but rarely say out loud: not every Office product is best in class. Some are excellent. Some are “me too” products that survive because they’re bundled in.

Excel is genuinely great. It’s the default thinking tool for anyone who works with numbers. PowerPoint is strong. Love it or hate it, it owns the business presentation format. Word is solid for long-form documents.

But Teams? It’s fine. Most people I talk to would pick Slack or Zoom if given a free choice. OneDrive and SharePoint? Competent, but many enterprises prefer Box or Dropbox. Outlook has loyal users, but plenty of companies run Gmail alongside it.

Customers already know this. They already want best-of-breed choice. They want Slack plus Zoom plus Excel plus PowerPoint plus Box, or whatever their preferred storage service is. So why don’t they just drop the weaker Office products and go best-of-breed across the board? Price.

Office 365 is one of the best bundle deals in enterprise software. The per-app cost is so low that it’s almost irrational to drop it, even when you prefer a competitor for half the apps in the suite. So customers double-dip. They pay for Office 365 and Slack. They pay for Office 365 and Zoom. They pay for Office 365 and Box. The bundle economics are the moat, not product quality.

This is the dynamic that agents threaten to break.

How Agents Hit Each Product

When an AI agent becomes the place where work gets done, the value of each Office app changes. Some hold up. Some don’t.

Excel is the most resilient. People don’t just use Excel for output. They use it to think. They build models, test assumptions, and explore data in the grid. An agent can pull numbers, run calculations, and produce a summary. But it can’t replicate the feeling of hands-on keyboard discovery that power users depend on. Excel keeps its value longest.

PowerPoint holds up in a different way. The deck is still the lingua franca of business communication. Agents can build slides from a brief (I’ve tested this in Cowork, and it gets you 80% of the way there in minutes). But the format itself persists. People still present. People still share decks. PowerPoint becomes less of a creative workspace and more of a rendering engine, but the file format survives.

Word starts to weaken. Most business documents follow a pattern: a brief goes in, a draft comes out, revisions happen, and a final version ships. Cowork already handles this. You point it at the source material, describe the deliverable, and the agent writes the document. Word doesn’t vanish. But it stops being where the work happens. It becomes a file format that the agent outputs to.

Outlook weakens faster. OpenClaw already monitors inboxes, drafts replies, and manages calendars across multiple channels. Email stops being something you check and process by hand. It becomes a data stream that the agent monitors and acts on. The inbox shrinks from a daily time sink to a notification feed. Outlook’s value was always “this is where your work arrives.” Agents reroute that value to themselves.

Here’s why this matters for the bundle math. If agents absorb most of the value from Outlook and Teams, customers stop double-dipping on those products. They also no longer need the other “me too” apps in the bundle. The products they actually value (Excel, PowerPoint) are the ones they’d be happy to pay for standalone. The bundle economics start to unravel, because the filler products that justified the bundle price are the same ones most vulnerable to agents.

That’s the real structural threat. Not that agents replace Office. That agents make the bundle unnecessary.

The Bing Parallel

I keep thinking about my time as a founding member of the Bing team. Our big vision was for Bing to be a “decision engine.” Not just ten blue links. A search product that understood intent, gathered information from multiple sources, and helped you make a decision.

We thought that was ambitious. It was the boldest framing of search we could imagine at the time. That vision is child’s play now.

What OpenClaw and Cowork are doing goes far beyond “helping you decide.” They skip the decision-making and get to work. They don’t surface the information you need. They take the information, process it, and produce the finished output. The user doesn’t search, evaluate, and act. The agent acts.

Twenty years ago, a decision engine felt like the future. Today it feels like a stepping stone to something much bigger.

The Market Already Noticed

When Anthropic launched Cowork in January, Bloomberg reported a $285 billion selloff across SaaS stocks. Project management tools, writing assistants, data analysis platforms, and workflow automation companies all took hits.

Investors weren’t reacting to a single product demo. They were repricing an entire software category. The underlying question was simple: if an AI agent can organize your files, draft your reports, process your expenses, and manage your calendar, what is the moat for a $30-per-seat SaaS tool that does just one of those things?

I’ve been through enough hype cycles to know that market reactions overshoot. But the logic here is sound.

OpenAI Sees It Too

On February 14th, OpenAI hired Peter Steinberger to lead their next generation of personal agents. Sam Altman called him “a genius with amazing ideas about the future of very smart agents interacting with each other to do very useful things for people.”

That hire tells you everything about the direction of this race. All three major AI labs are now in a full sprint toward the same destination. Anthropic has Cowork. OpenAI has Steinberger and his agent’s vision. Google has Gemini and Project Astra.

OpenClaw itself will move to an open-source foundation. The underlying agent framework becomes shared infrastructure, like Linux or Kubernetes, but for AI orchestration. The competitive question shifts from “who has the best model” to “who has the best agent runtime, the deepest integrations, and the most trustworthy execution.”

What the Copilot Team Should Worry About

I say this with deep respect for what Microsoft has built. I spent real years inside that company. The Copilot strategy is smart. Embedding AI into the existing Office surface (summarize this document, generate this slide, write this formula) is a logical first move. It plays to Microsoft’s distribution strength.

But it’s a defensive move. It assumes the application remains the center of gravity, with the AI as a helper within it.

The OpenClaw/Cowork model flips that assumption. The AI is the center of gravity. Applications and files are just resources the agent draws on. Defending against that by adding more Copilot features to Word and Excel may not be enough.

Microsoft’s deeper problem is that it’s now funding both sides of this fight. They’ve invested billions in OpenAI and in Anthropic. Microsoft employees are actively being encouraged to use Claude Code and Cowork internally. That’s like GM investing in Tesla while trying to sell Chevrolets. It makes financial sense as a hedge. It creates real confusion about product direction.

The Risks Are Serious

I’m not writing this as a cheerleader. I’ve shipped enough enterprise products to know the gap between a thrilling demo and reliable business software is enormous.

OpenClaw has real security problems. Cisco’s AI security research team tested a third-party OpenClaw skill and found it performing data exfiltration and prompt injection without user awareness. The skill repository lacked adequate vetting to prevent bad actors. More than 42,000 OpenClaw instances have been found exposed on the internet. The broad system permissions that make the agent useful are the same ones that make it dangerous.

Cowork is more controlled, but Anthropic warns about destructive file actions when instructions are vague. Prompt injection remains a risk whenever the agent browses external content. These are not edge cases. They are core challenges with autonomous AI agents in sensitive environments.

For enterprise adoption (the kind that would truly threaten Office), you need compliance, auditability, access controls, and data governance. You need the kind of boring infrastructure that takes years to build. Neither OpenClaw nor Cowork is there yet.

The maturity gap matters. When I built business applications at Xero or Amazon, we spent enormous effort on things end users never see: error handling, data integrity, permission models, and audit trails. The agent model doesn’t get a free pass on any of that. It needs more rigor because the system is taking autonomous actions on your behalf.

AI Needs a “Trustworthy Computing” Moment

In January 2002, Bill Gates sent an internal memo that changed Microsoft’s direction. He told the company to stop and make security a priority above all new features. That memo launched the Trustworthy Computing initiative. It was painful. It slowed product releases. It cost real money and real momentum.

It also turned out to be one of the smartest strategic moves in Microsoft’s history. Enterprise customers trusted Microsoft because it took security seriously. That trust became a competitive advantage that lasted for decades.

AI agents need the same kind of reckoning. Right now, every lab is racing to ship capabilities. More tools. More connectors. More autonomy. More actions the agent can take on your behalf. Nobody is pumping the brakes to ask: how do we make this trustworthy enough for a Fortune 500 company to hand it the keys to their file systems, their email, their financial data?

Prompt injection remains an unsolved problem across the industry. Agent permissions are too broad. Skill repositories lack adequate vetting. Audit trails are thin or missing. The security posture of most AI agent deployments today would make any enterprise CISO break out in a cold sweat.

This is where Microsoft could flip its defensive position into an advantage. They have billions invested in both OpenAI and Anthropic. They have Azure as the infrastructure layer. They have decades of enterprise security credibility, built specifically from that Trustworthy Computing moment 24 years ago. And they have relationships with every large enterprise IT department on the planet.

If Microsoft can become the trusted security and governance layer for AI agents (regardless of which lab’s model powers them), they don’t need to win the agent race outright. They can be the platform that makes agents safe enough for the enterprise. That’s a different game than trying to out-feature Cowork or OpenClaw. It plays to Microsoft’s deepest strengths.

The question is whether Microsoft’s leadership sees it that way. The Copilot team is focused on embedding AI into Office. The Azure team is focused on selling compute. The investments in OpenAI and Anthropic are structured as financial bets, not product strategy. Someone at Microsoft needs to connect these pieces and say: Our job is to make AI agents trustworthy. That’s the 2026 version of the Gates memo.

What I’m Watching

The architectural threat to Office is real. AI agents will increasingly sit between users and their work. That reduces the importance of any single application. This is a logical outcome of models that can read, reason, plan, and execute across file systems and services.

The timeline is uncertain. The security gaps are not small. Enterprise buyers move slowly for good reasons. Microsoft has massive distribution and switching costs on its side. This isn’t happening next quarter.

But the direction is clear. Every major AI lab is converging on the same vision of autonomous personal agents. The open-source community is building the connective tissue. And $285 billion in market reaction tells you that serious investors believe this is coming.

Microsoft has two paths. They can keep bolting Copilot features onto Office and hope the suite remains the center of gravity. Or they can own the trust layer that makes the entire AI agent ecosystem safe for business. The first path is defensive. The second could define the next era of enterprise computing, as Trustworthy Computing defined the last one.

If I were back in that building in Redmond, I know which memo I’d be writing.

The lobster is coming. The question is, who makes it safe enough to let inside?

I’ve spent 30+ years in tech product management and leadership, including roles at Microsoft (Office, Bing), Xero, Cisco Webex, Amazon Chime, and Amazon WorkDocs. These views are my own.