Nobody Ever Got Fired for Buying Copilot
Somewhere right now, an employee who spent the weekend using a frontier AI model to write software, analyze research, build an agent, and restructure an entire project is arriving at work Monday morning.
He opens his laptop. "Good news," IT says. "We've approved AI."
Fantastic. "Microsoft Copilot."
Oh.
So we're doing AI with the Tonka set.
Here's the thing though — and this is where it gets interesting.
Microsoft isn't building a bad product.
Microsoft is attempting something much more ambitious: turning its existing enterprise software footprint into the distribution layer for enterprise AI.
Word. Excel. PowerPoint. Outlook. Teams. SharePoint. Azure. Entra. Windows.
Microsoft already owns the building.
The question is whether owning the building means you automatically own the intelligence layer sitting inside it.
And the evidence so far suggests the answer is: not necessarily.
The Amgen receipt is almost too perfect.
Microsoft publicly promoted Amgen's decision to deploy Microsoft 365 Copilot to 20,000 employees. This was a flagship deployment. A case study. The kind of thing that goes in the earnings call deck.
Thirteen months later, Bloomberg reported that Amgen employees were using ChatGPT instead.
That's extraordinary. Microsoft had literally used Amgen as a Copilot success story. The poster child walked out the back door and started using somebody else's product.
Microsoft: "Look! Amgen chose Copilot."
Amgen employees: "Cool. Anyway, open ChatGPT."
The anonymous pharma CIO is even uglier.
A pharmaceutical-company CIO deployed Copilot to 500 employees for six months, then canceled it. His explanation was essentially that Microsoft was asking roughly twice as much money for Office with Copilot, and he couldn't identify twice the value.
He compared Copilot's PowerPoint generation to "middle school presentations."
The meeting-summary feature — which he initially considered one of the better functions — was disabled after the legal department became concerned about storing transcripts.
That's almost my Tonka-set line written by somebody else.
Gartner found this wasn't isolated.
In its 2024 enterprise research, Gartner found only 6% of surveyed organizations had finished a Copilot pilot and were moving toward large-scale deployment. 72% reported employees having difficulty incorporating it into daily work. 57% said engagement fell off quickly.
Gartner also identified oversharing and security as a significant barrier. Subsequent findings were similarly restrained: employees could find value in Copilot, but tangible business impact remained elusive, while security remediation, training, and information-management cleanup required more effort than expected.
And here's the important nuance.
Copilot isn't useless.
A large randomized study across 56 companies and 6,000+ workers found measurable benefits: faster document creation, less time spent reading email. Real effects. Not imaginary.
But only about 40% of workers given access used it regularly during the six-month study.
So the indictment isn't: "Microsoft built an AI that literally does nothing."
It's: "Microsoft took a useful productivity assistant and marketed it as the enterprise AI layer."
Those are wildly different things.
The sales numbers tell their own story.
In late 2025, Microsoft reportedly lowered sales-growth expectations for some AI products after salespeople missed targets, reflecting resistance from companies to spending more on AI.
By early 2026, the Wall Street Journal was reporting broader Copilot problems: low adoption, confusing product branding, underused subscriptions, and users preferring competing products.
Microsoft had about 15 million paid Microsoft 365 Copilot seats against roughly 450 million commercial Microsoft 365 seats.
That ratio is fascinating.
Microsoft already owns the building. It controls Word, Excel, PowerPoint, Outlook, Teams, Windows, Entra, SharePoint, Azure.
And it's still having to fight to get people to use Copilot.
That says something.
And regulators are watching.
In May 2026, the UK's Competition and Markets Authority opened an investigation into Microsoft's business-software ecosystem. It specifically said it would examine whether bundling, interoperability limitations, and defaults restrict customer choice — including whether competing AI systems can integrate properly with Microsoft's software.
Copilot is explicitly in scope.
The CMA's language is almost tailor-made for this argument: businesses should be able to "mix and match software and AI services" from different suppliers rather than having Microsoft's existing position determine which AI they use.
Italy opened a separate investigation in June 2026 over Microsoft integrating Copilot and Designer into Microsoft 365 while increasing subscription prices, examining whether consumers were given sufficient ability to make an informed choice.
This isn't theoretical antitrust anxiety.
Microsoft already went through the same movie with Teams.
The European Commission challenged Microsoft over bundling Teams into Office/Microsoft 365. It eventually extracted legally binding commitments requiring Microsoft to keep offering suites without Teams and improve interoperability and data portability for competitors.
That's the historical predicate.
Microsoft learned something enormously valuable from Teams: you don't necessarily have to build the product everybody voluntarily chooses. You can make your product the path of least resistance inside an ecosystem everybody already has.
And enterprise IT makes that effect much stronger.
The CIO doesn't necessarily say: "We compared Claude, Gemini, ChatGPT, Copilot, and six agent frameworks and concluded Copilot has superior reasoning."
He says: "We're an M365 shop. Copilot is approved. It's integrated with Entra. Legal already understands Microsoft's DPA. Procurement already has the vendor. Security already knows the environment. Nobody is getting fired for enabling Microsoft."
And suddenly 30,000 employees are told: this is the AI you're allowed to use.
That's not technological selection.
That's procurement gravity.
And then you get the other phenomenon.
Employees have used ChatGPT, Claude, Gemini, Codex, and whatever else in their personal lives. They know what frontier AI actually feels like.
Bloomberg's reporting specifically described Microsoft struggling with corporate customers because employees wanted ChatGPT instead.
That difference matters enormously.
Because Microsoft's original advantage was supposed to be: sure, ChatGPT is cool, but Copilot knows your company. Outlook. Calendar. Teams. SharePoint. OneDrive. Word. Excel.
Fantastic proposition.
But if the intelligence sitting on top of that context feels constrained, inconsistent, or mediocre, you've essentially created:
The world's most extraordinarily integrated mediocre coworker.
And that's where the category error lives.
Microsoft has built an incredible enterprise plumbing system.
Identity. Permissions. Compliance. Data residency. Document access. Endpoint management. Audit. Governance.
That stuff isn't toys. Microsoft is terrifyingly good at it.
But companies are making a category error when they assume:
Best enterprise plumbing + AI button = best enterprise intelligence.
No.
Those are different layers.
Enterprise AI is in danger of repeating enterprise software history.
Where tool selection is driven less by capability than by integration, procurement, compliance, incumbent contracts, and switching costs.
The difference this time is that intelligence isn't CRM software.
Giving everybody the third-best expense-report application is annoying.
Giving everybody the third-best cognitive system may be a competitive disadvantage.
You're not just picking a tool.
You're picking the ceiling on how smart your organization can be.
And if you're picking based on who already has the vendor relationship rather than who has the best intelligence, you're making a procurement decision dressed up as a technology decision.
The interesting question isn't whether Copilot is good or bad.
The interesting question is whether the mechanism that determines which AI enters an organization has anything to do with which AI is actually best.
Microsoft's distribution advantage is real. Its plumbing is exceptional. Its enterprise footprint is genuinely formidable.
But distribution isn't capability.
Availability isn't adoption.
And a monopoly on the building doesn't automatically give you a monopoly on what happens inside the building.
The employees already know the difference.
That's why they're opening ChatGPT on company time.
And that's the part that should worry Microsoft — and every CIO who thinks they've solved their AI problem by checking a box on a procurement form.
Intelligence isn't software.
You can't just install it and forget it.
And the people who've tasted what frontier AI can do aren't going to pretend the Tonka set is the same thing.
They know.
They're just waiting for procurement to catch up.
Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital.
Comments