By the end of last year, it was clear that AI had been a bubble. New models were just like the old ones. They were still clunky and error-prone, excelling only at meme-slop. AI companies would never recoup their spends on data centres, and their clients couldn’t make money from AI projects. In October and November a trillion dollars of value was wiped from US stocks (before they recovered).
Last month, though, investors started freaking out for precisely the opposite reason. AI was TOO good. The subscription software industry, in particular, was seeing an extinction event. The Jobs Apocalypse was back on.
What prompted this realisation was a set of plugins Anthropic released for its Claude model that promised to make redundant the specialised, industry-specific software used in law, finance, marketing and more. Software-as-a-Service was doomed. Almost a trillion dollars was wiped from the value of stocks (before they recovered).
And yet, to the average user, the experience of using LLMs is pretty much the same as it was 18 months ago. How can all this be true at once?
The explanation, commentators now think, has to do with a growing divergence, a bifurcation, in the way that AI is being used. The vast majority of us still ask chatbots for things like shopping and dining recommendations, drafting text and making videos. We sit and wait for the answer, sometimes copy-and-pasting the results into other apps.
Power Users, by contrast, harness teams of AI agents to execute large chunks of their workflows. They don’t ask questions, they set their agents goals – often with their own KPIs – and let them get on with it. They are not looking at a chat box; they work with command lines and automation layers, and their AI agents run in the background, connected to all the other tools they use.
And in contrast to what was reported a year ago about AI’s negligible effects on productivity, Power Users record big time savings and performance improvements.
This phenomenon, dubbed the “Agentic Divide”, has been the subject of influential papers from institutions like Anthropic Labs, MIT Sloan Business School and Stanford’s Digital Economy Lab. They prophesy the “death of the chatbot era”.
A big catalyst for the Agentic Divide has been Claude Code. This started as a terminal tool for coding, but has now evolved into an orchestrator, or conductor, of “Agent Teams”. Unlike standard chatbots, it allows a lead agent to spawn and manage multiple sub-agents that work simultaneously across different files and complex technical tasks.
It is this shift that defines the Power User. They’ve moved beyond chatting with an AI to essentially managing a small, autonomous software engineering firm.
Stanford’s Erik Brynjolfsson predicts: “By 2050, most people will command workforces larger than the biggest multinational corporations of today. But our ‘employees’ won’t be people sitting in cubicles or standing on factory floors. They will be fleets of AI agents.”
Don’t feel like you’re about to become a power user? Worries that you might get left on the wrong side of this digital divide are probably premature. The internet also used to be abstruse before it was democratised by graphical user interfaces and point-and-click functionality. As this happens with AI, perhaps the best thing to work on is your mindset. Can you, like a Power User, think less about chatting to your AI, and start commanding it?
Joe Smith is Founder of the AI consultancy 2Sigma Consultants. He studied AI at Imperial College Business School and is researching AI’s effects on cognition at Chulalongkorn University. He is author of The Optimized Marketer, a book on how to use AI to promote your business and yourself. Contact joe@2Sigmaconsultants.com.