The artificial intelligence industry has spent years feathering the accelerator of frontier model development in a fierce contest to dominate the technology. Now, it is considering something that until recently seemed almost unthinkable: slamming on the brakes.
The prospects of the frontier labs are closely linked to the fortunes of the US stock market, and they buttress vast spending on infrastructure. But for businesses still working out how to harness AI’s illusive productivity gains, a slowdown is not a bad thing. It will give them time to close the yawning gap between AI’s accelerating capabilities and their productive use of them.
Still, calls to pace the frontier grow louder. Anthropic chief executive Dario Amodei made the case in an essay in which he said capabilities are advancing faster than our ability to understand, align and control them. His proposal was to deliberately slow the improvement enough to give safety research and independent evaluation time to catch up.
Is it practical?
There remains considerable disagreement over whether this is practical, desirable or possible. But that’s beside the point for businesses, since most of them are nowhere close to exhausting the AI capabilities they already possess.
According to the Stanford 2026 AI Index, 88 per cent of organisations surveyed reported using AI in at least one business function last year. But breadth of adoption is not the same thing as depth. The research shows that the use of AI agents remained in the single digits across almost every business function.
Relatively few have fundamentally redesigned how work gets done around it. Research from the Federal Reserve Bank of St Louis found that generative AI accounted for only 5.7 per cent of total working hours by August last year. This tells us that companies already have powerful AI, and are simply failing to make the most of it. They don’t need another massive leap in model capability to claim the productivity gains that are already on the table.
A METR randomised study found that experienced open-source developers working on highly familiar repositories actually took 19 per cent longer when allowed to use early-2025 AI tools. The reason is that the bottleneck is increasingly organisational.
Companies have to connect models to proprietary data. They have to integrate AI into existing systems. And they need to redesign workflows rather than simply insert a chatbot into them. They also need appropriate governance, security and human oversight, while employees have to learn how to delegate effectively to AI. Managers, meanwhile, need to rethink roles, performance measures and decision rights.
Those changes happen at a comparatively glacial pace when compared with the speed at which AI labs have been pushing the frontier.
More capabilities
OpenAI’s enterprise research finds that the heaviest AI adopters use substantially more capabilities across more tasks than the median employee. These intensive users report much larger productivity benefits. Seventy-five per cent of surveyed workers also reported being able to perform tasks using AI that they previously could not do.
So the AI productivity frontier inside most companies sits far behind the technological one. This is not unusual. General-purpose technologies typically arrive before organisations learn how to exploit them. Electricity existed before factories were redesigned around distributed electric motors, and computers came along long before companies reorganised processes around digital information.
AI is likely to follow the same pattern, which only strengthens the case for hitting the brakes. That will not stall the AI economy, which is so crucial to public equity markets, because it will give companies the badly needed time to extract value from the marvellous capabilities already at their fingertips.
That means moving beyond experimentation towards workflow redesign. Moving from individual productivity tools towards end-to-end processes. Connecting AI to enterprise data and systems. Using agents where they genuinely add value. Building governance that allows experimentation without unacceptable risk. And, perhaps most importantly, developing the organisational capabilities required to use AI well.
Executives should continue watching the frontier, of course. Better models will create new opportunities, and a substantial slowdown is far from guaranteed. But they should not confuse a slower technological frontier with a slower business opportunity.
Even if frontier development paused dramatically tomorrow, companies would still have an enormous backlog of economically valuable opportunities across customer service, software development, finance, procurement, marketing and many other functions.
So just because the frontier slows, it doesn’t mean the business opportunity does. The AI race, at least for enterprise users, needs to shift from simply building better AI to building better businesses around it.
Leave a Reply