TL;DR
This is a distribution move, not a features move. Microsoft already owns the enterprise contract. The Anthropic deal just fills out the bundle.
Companies dropped Slack for Teams not because Teams was better, but because it was already paid for. Copilot Cowork is the same play for AI agents.
Claude Cowork runs on your machine. Copilot Cowork runs in Microsoft's cloud. Employees prefer local. IT prefers visibility. IT writes the policy.
Thin UI wrappers get squeezed. Regulated infrastructure (payments, messaging, compliance) doesn't disappear. Know which one you are.
Your company's data is not on the internet. Getting it into these agents safely is the work worth doing.
The bundle doesn't win on features. It wins on friction.
Microsoft is already in the building.
Not "entering." Already there. Already on your invoice. Already approved by IT. Already the thing your team groans about but still opens every morning.
So when Microsoft adds Anthropic's models to Copilot, the CFO conversation doesn't start with "which AI is better." It starts with "why are we paying for two things that do the same thing?"
We've seen this before. Companies were using Slack. Then COVID hit, people over-hired, and three years later the CFO started asking why they were paying for Slack when Teams was sitting right there in their Office 365 contract.
The answer was... it wasn't.
Copilot Cowork doesn't have to be the best agentic tool to win most of the market. It just has to be good enough, already paid for, and blessed by IT. That's a pretty short checklist.
Cloud-first vs local-first... and why employees are going to hate this
Here's the actual product difference worth knowing.
Claude Cowork runs locally on your device. It can create spreadsheets, presentations, and work across the tools on your machine. Employees who've used it tend to like it. You're working with your files, on your computer.
Copilot Cowork does the same category of tasks, but everything runs in the cloud, in OneDrive and Microsoft's ecosystem. Jared Spataro, who leads Microsoft's AI-at-work efforts, told Reuters: "We work only in a cloud environment... so you know exactly what information it has access to." Most enterprise clients apparently prefer that.
I believe the enterprise preference is real.
I also believe that most employees are going to hate hunting for AI-generated files inside OneDrive. (Personally, I hate when coworkers send me anything in OneDrive. Finding things in there is its own full-time job.)
So the enterprise wins visibility. The employee loses a little control. That tradeoff plays out all the time. Enterprise almost always wins.
Why Microsoft is dating Anthropic while still married to OpenAI
This deal is also about investor optics.
OpenAI accounts for nearly 45% of Microsoft's cloud business contract backlog. That's a dependency that makes investors nervous. Adding Anthropic models to M365 Copilot is Microsoft saying: we're the platform, not just OpenAI's best customer.
That's CFO logic. Diversify the supply chain. Don't let one vendor own your story.
The practical outcome for the rest of us: expect model menus inside the tools your clients already pay for. Pick the one that's best for the task, swap it out when a better one shows up. Microsoft becomes the interface layer and doesn't care which model wins.
"SaaS is dead" is a headline. The actual casualties are lumpy.
Some of this panic is real. Most of it isn't.
You're not vibe coding your way to a new global payments stack. You're not rebuilding Twilio this weekend. Any SaaS that lives inside a regulated relationship (payments, communications, healthcare data) is fine. The relationships those companies maintain with federal regulators don't get automated away with a prompt.
What gets squeezed: thin UI wrappers. Tools whose entire value proposition is "we made a nice screen on top of data you already own." When the agent can pull that data and build the screen on demand... the subscription is a harder sell.
The honest version of this: SaaS companies don't die, but they have to fight for their value in a way they've never had to before. More API. Less "we control the experience." Make your data useful to every tool in the ecosystem, or watch your customers route around you.
Somebody still owns the thing after it ships
I was talking to a CTO this week who laid out the tier structure they're thinking about for employee-built tools:
Tier 1: One person builds something. Solves their problem. Nobody else needs to know it exists.
Tier 2: A few people need it. Lock it down to those people, make sure it doesn't break.
Tier 3: A whole department is using it. Now you need real access control, security review, and someone who's accountable when it breaks on a Friday.
Here's what that CTO didn't say but I'm saying: the person who built it is usually still the one holding the pager when Tier 3 happens. They either become the accidental IT department, or it gets handed to a consultant to actually stabilize.
Agentic engineering is faster. It doesn't delete the maintenance reality. It just moves the bill to later, when the tool has users who depend on it.
If you want to gut-check what's actually blocking your business before you add more tools to the stack, the small business strategy diagnosis quiz was built exactly for that. Separate the real bottleneck from the noise.
What this means if you're watching from the outside
If you're in the Microsoft ecosystem: Copilot Cowork is coming whether you feel ready or not. Learn what it touches, get your data clean, and decide in advance who owns what it builds.
If you build SaaS: Your moat isn't the UI anymore. It's proprietary data, compliance depth, or integrations that are genuinely hard to replicate. If you can't name that in one sentence, the bundle is coming for your renewal conversation.
If you're a consultant or IT person: Your value is not "I know which model is best." It's "I can get a company's specific data into these agents safely, with the right access controls, without turning every employee into an accidental admin."
Generic internet knowledge is what Claude and ChatGPT are already trained on. Your company's data is not on the internet. The businesses that figure out how to get their real operational knowledge into these agents are going to run circles around the ones still arguing about model cards.
That's the edge. Go build it.
For a shortlist of what AI tools are actually worth using right now (and what to skip), use the AI tools checklist.