The Medium Is the Message

A strange contradiction is emerging in enterprise AI. The digital workplace incumbents had every advantage: hundreds of millions of users, dominant platforms, and the resources to move fast. In enterprise AI, those advantages are beginning to look like constraints and there's more than a hint that the two main incumbents might have moved in the wrong direction.

Dominant platforms now discover their advantages may be structural liabilities

Microsoft embedded Copilot into Office. Google folded Gemini into Workspace. Both decisions, of course, appeared sensible, if you already have hundreds of millions of daily users on your software apps. Adding AI to their existing tools requires nearly no behaviour change of users and from a distribution perspective, this was the inevitable move.

But does such a move miscalculate what this new technology is?

I mentioned before in this article that AI sits in an awkward position - not quite software, not quite intelligence. Somewhere uniquely in the middle. We see some vendors pitching it as intelligence (it's not) and others pitching it as a software tool (it's not).

The critical question for any organisation currently spending billions of dollars to place AI into its existing suite is can AI be meaningfully delivered inside applications designed for different purposes.

Early usage patterns suggest the mismatch is more significant than has been publicly acknowledged.

Microsoft and Google did what any company would do with hundreds of millions of existing users. They added AI to the place those users already were. The decision required no defence at the time. It was simply what their position made possible.

Value migrates from executing functions to processing novel information combinations

A tool executes a specific function within defined parameters and its value comes from operating reliably within constraints. This is how software has worked for decades, and it is how users have learned to think about the applications they use daily.

In contrast AI shifts with context. The same model produces different outputs depending on how it is prompted and what it encounters. Its value comes from reasoning about ambiguous problems, processing novel combinations of information, and identifying patterns that escape human attention. When you confine this to a feature inside an existing application, you are limiting the technology to a fraction of what it can do. You are potentially asking a reasoning engine to behave like a formatting assistant.

Software tools are designed to behave the same way each time you use them. That predictability is what makes them useful. AI behaves differently depending on what you ask it and what context surrounds the question. These are different kinds of technology, and it is not obvious that one fits neatly inside the other.

Gemini grows to 21.5 percent by defying its own distribution advantage

The distribution theory has started generating data. According to SimilarWeb's Global AI Tracker, Google's Gemini has grown from 5.7 percent to 21.5 percent of generative AI web traffic over twelve months. ChatGPT's share dropped from 86.7 percent to 64.5 percent in the same period. At first glance, this appears to vindicate Google's strategy.

But the data measures something specific: traffic to gemini.google.com, the standalone destination. Usage of Gemini inside Workspace, whether smart replies in Gmail or document drafting in Docs, does not appear in these numbers. The growth that is showing up comes from users seeking out Gemini directly, not from users being funnelled into AI through their existing productivity tools.

Microsoft's results sharpen this distinction. Copilot, despite being baked into Windows, Edge, and the Office suite, moved from 1.5 percent to 1.2 percent market share over the same twelve months. If distribution alone determined adoption, Copilot should be growing. It is not.

The contrast suggests that embedding AI into existing workflows does not automatically translate into engagement. Users seem more willing to visit AI as a destination than to invoke it as a feature within tools designed for other purposes. Pre-installation creates exposure. It does not guarantee use.

Google's AI growth is coming from gemini.google.com, the standalone destination. Microsoft's Copilot, embedded across Windows and Office, lost share over the same period. We are starting to see where users actually go when given the choice.

Twelve to eighteen months separate successful implementations from abandoned experiments

Gartner's 2025 analysis found that forty percent of agentic AI projects will be abandoned by 2027. Deloitte's tech trends report from the same year found that organisations with successful AI implementations did not simply add agents to existing workflows. They redesigned their workflows first, a process that took twelve to eighteen months. Deloitte's Bill Briggs puts it directly: 'If you just take your existing workflow and try to apply advanced AI to it, you're going to weaponise inefficiency.'

The pattern is worth noticing. The organisations that succeeded changed their processes to fit the technology. The organisations that struggled tried to fit the technology into their existing processes. The former approach is expensive and slow. The latter approach is fast and appears to be running into limits.

Microsoft and Google chose the fast approach for compelling reasons. Whether speed and distribution can compensate for a fundamental mismatch between what AI does and what these applications were designed to accommodate is what the coming years will determine.

OpenAI and Anthropic design for autonomous phases, not embedded assistance

OpenAI and Anthropic are building from different assumptions. Rather than embedding AI into existing applications, they are creating interfaces that assume autonomous execution from the start. Users trade the familiarity of traditional software for workflows where AI handles research, context synthesis, and reasoning as a separate phase before human judgement takes over. The applications are designed around what AI does well rather than around what existing software already does.

OpenAI and Anthropic are not trying to add AI to existing software. They are building new interfaces that assume AI needs room to reason before handing decisions back to humans. It is a different theory of what the technology is and how it should be used.

Intelligence as feature versus intelligence as foundation

The two approaches rest on incompatible assumptions about what AI is and how it should be deployed. One treats AI as an enhancement to existing software. The other treats AI as something that requires different interfaces, different workflows, and different expectations entirely.

Both have challenges, but the integration model may have solved the wrong problem. Getting AI to more people faster assumed that reach was the primary challenge. The data suggests the challenge might be more fundamental, whether AI can deliver its value when confined to interfaces designed for different purposes.

We are watching two different theories of AI play out in real time. One assumes intelligence can be packaged as a feature. The other assumes intelligence requires rethinking how software works entirely. Both approaches have committed significant resources based on these assumptions.

The curious thing is how long it might take for the data to definitively favour one approach over the other. In the meantime, organisations continue making architectural decisions based on theories that remain untested at scale.