The Bifurcation of Intelligence

Why an “AI Ready” strategy might just be a BlackBerry moment in an iPhone world.

I know, I know. I asked you in our last deep-dive Briefing to “bear with me” as I wrote (once again) about AI. And now I am back at it. 🙃 But it’s hard to argue that the whole AI thing isn’t deeply important and something we should all be thinking about… right?

The other day I received an email promoting an “AI Fluency” course. Nothing wrong with that per se. But it made me wonder – while many of us are still trying to figure out this whole AI thing (by learning to “prompt” our chat interfaces) and businesses invest heavily in rolling out chatbots like Microsoft’s Copilot, the spear tip of the market has long moved on. The real power users have stopped using AI as “Google on steroids” and started using complex AI agents to take over ever-larger chunks of their work – and they are doing so autonomously.

Which leads to a weird bifurcation: On one side, business leadership is congratulating themselves on making their companies “AI-ready” by buying 10,000 seats of Microsoft Copilot so employees can summarize emails. On the other side, a completely different set of users has discovered that agentic coding tools (like Claude Code CLI) can, with a bit of tweaking, be extremely useful for tasks that have nothing to do with software engineering.

Martin Alderson recently pointed out this widening gap, noting that he is seeing finance directors and marketers – people who are decisively not engineers – running Python scripts in terminal windows to automate massive workflows. They aren’t chatting with a bot, but deploying the AI version of a whole data science team.

Now, the problem is that these tools tend not to be sanctioned by corporate IT departments. You generally can’t run a command-line interface or execute arbitrary Python code on a locked-down enterprise laptop. So, this “real” AI work is happening either in smaller, nimble companies or by employees who are actively circumventing the rules.

There is a historical rhyme here – we are in the “BlackBerry vs. iPhone” era of AI. Corporate IT loved the BlackBerry (yesteryear’s version of Microsoft Copilot) because it was secure, controlled, and fundamentally limited. The users, however, want the “iPhone” (agentic tools such as Claude Code/Cowork or ChatGPT Codex) because it actually allows them to do the things they need to do (in a rather magical way). And just like in 2008, the “shadow” usage is where the actual productivity revolution is happening.

The dichotomy between large-scale enterprise use of AI and what individual users can do with “tip of the spear” tools is vast – and it’s becoming a structural risk. To state it bluntly: The “Chat” interface is a dead end for complex work.

Alderson uses the example of a finance director trying to modernize a complex financial model. In the “sanctioned AI” world, they are stuck in Excel, asking Copilot to help with formulas. It’s slow, it breaks, and it’s still just a spreadsheet. In the “rogue AI” world, that same director uses an agent to convert those 30 sheets of Excel logic into a Python script. Suddenly, they aren’t just doing “better Excel” – they are running Monte Carlo simulations, pulling in live external data via APIs, and building web dashboards. They have jumped the species barrier from “clerk” to “engineer,” simply because they had access to a tool that could write and execute code.

The result is that the companies with the most resources (enterprises) are becoming the least capable of leveraging AI. While the small startup team is building an automated machine that runs circles around the competition, the enterprise team is stuck asking a chatbot to summarize a PDF.

The end result is two distinct classes of knowledge workers: There are the Consumers, who will stay within the guardrails, use the sanctioned tools, and see a marginal (10–20%) bump in productivity. Consumers draft emails faster and find documents easier. Bravo.

And then there are the Builders. They might not have “developer” in their job title, but they are using agentic tools to build their own infrastructure, automate entire processes, and bypass the limitations of their official software stack. Builders are seeing productivity gains of 10x or 100x.

The danger for leaders is assuming that buying the “Consumer” tools means you have solved the AI problem. You haven’t. You’ve just given your people a slightly better typewriter, while your competitors moved on to the networked laser printer.

@Pascal

Musical Coda:

(Because sometimes you have to break the house rules to actually build something new.)