Over the last month I have had the same conversation many times over, with executives and leaders in business, in sport, in education, in government. The questions are always about tools. Which one to use. Which model is best. Whether they are already behind. What almost none of them have done is spend real time with any of it. They are using it for a small fraction of what it can do, and so they do not know what it is for.
At Cannes Lions last week I had a version of that conversation on a terrace with someone who runs a large media business. He had a tool in his pocket that can write working software, run research across a dozen sources and act on what it finds, hold a brief in its memory for weeks. He was using it as a glorified search engine, something to shape an email or tidy up a slide deck. He was happy with it. By any fair measure he was barely using it at all.
I used to think the gap I was seeing was about access. That these people needed the better model, the newer release, the thing the labs had just shipped, and then they would pull ahead. I have stopped believing that. The gap has almost nothing to do with the model. The last month has made that easier to see than it has ever been, because the thing we have all been treating as one story came apart in public.
Here is how I have started to think about it, and it has changed how I answer them.
For about three years it made sense to talk about AI as a single thing moving in one direction. The models got better. You could get hold of them the day they launched. A while later you could do a bit more with them than before. Those three things moved together closely enough that treating them as one story did no harm.
They are not moving together any more. There are three separate clocks now, and they are running at different speeds.
The first clock is capability: what the best models can actually do. It is still moving fast. The second is access: whether you, specifically, are allowed to get the best model. That clock has started, for the first time, to run backwards. The third is deployment: whether your organisation can take the intelligence it already has and turn it into work that actually gets done. That clock has barely moved in three years.
Almost everyone I meet is watching the first clock. A few have started to worry about the second. Hardly anyone is watching the third, and the third is the only one that decides what happens inside their business.
It is worth looking at what the three clocks did this month, because they separated in plain sight.
The capability clock kept moving, and the surprise was who moved it. The 2026 Stanford AI Index found that the performance gap between the leading American and Chinese models has effectively closed, with the top US model ahead by under three per cent by the spring. Open weights models, the kind you can download and run on your own hardware, are now part of the same conversation as the most advanced models rather than a cheaper substitute for them. A model called GLM 5.2, released by a Chinese lab in the middle of June under a licence that lets anyone use it freely, went straight to the top of the independent rankings for open models and scored level with one of OpenAI’s flagship models on a test built to measure how well a model gets real work done rather than how it answers a single question. It costs a small fraction of the Western models to run, and once you have the weights, nobody can take them back.
The access clock did something it has never done. Anthropic launched its two most capable models on the ninth of June. Three days later the US Commerce Department issued an export control order that cut off access to both for any foreign national, including the company’s own employees, which in practice forced a global shutdown. It was the first time export controls had been used against a model rather than against chips. Two weeks after that, OpenAI released its newest family of models to around twenty organisations, chosen with the government’s involvement, and said openly that it did not think this kind of approval step should become normal. The most capable systems on the planet stopped arriving as products you could buy. They started arriving as capability that someone else decides whether to grant you.
So the top of the market is pulling away and partly disappearing behind the government, while at the same time the floor has fallen out, because anyone can now run a model almost as good as the best for almost nothing. If this were only a story about capability and access, it would be the moment everyone rushes to the cheap option. That is not happening, and the reason is the third clock.
A model on its own does almost nothing. It is a brain in a jar. It has no memory of your business, no hands to reach into your tools, no sense of which of its own answers to trust or which job should go to which model. The thing that gives it those is called a harness. A harness is the layer wrapped around the model: the memory that holds your context, the connections into your real systems, the routing that sends a simple task to a cheap model and keeps the hard one for the expensive model, the checks that catch it when it goes wrong. It is also what lets one setup work across more than one model, so you can plug a different brain in underneath without rebuilding everything on top. The model is the brain. The harness is the body and the nervous system. The product is the two together, and the model is the part you can swap.
This is why a model being a fraction of the price and almost as good does not trigger a stampede. Switching model is not swapping one part for another. It means rebuilding the harness around it, and most companies never built one in the first place.
When a capable team does make the move, the work is real, and it is not the work you would expect. Lindy is a small company in San Francisco that builds AI assistants for other businesses. In June its founder moved the entire product off Anthropic’s models and onto DeepSeek, a cheaper open weights alternative, because the monthly model bill had grown larger than the company’s payroll. He saved millions. He also said the migration turned out to be something like a hundred times more work than the team had expected, because of how much infrastructure and internal tooling they had to build to make the cheaper model behave. The model itself was a single setting they could change. The system around it was months of engineering.
I know the shape of this because I live in a small version of it. I run a handful of agents on a modest machine at home, and I have rebuilt the scaffolding around them more times than I would like to admit. The model underneath has changed more than once, and each time it mattered far less than I expected. The work went into the part you cannot download: what the agents remember, what they are allowed to touch, how they pass tasks between them, when they stop and ask me before acting. That is the actual job. It is also exactly where most organisations are stuck.
The wider evidence says the same thing in colder terms. Forrester puts most agent failures down to ambiguity, poor coordination and unpredictable behaviour rather than the model being too weak. Industry trackers keep finding that the large majority of pilots never reach production at all. They do not fail because the intelligence is not good enough. They fail in the last stretch, in the system that was meant to carry the intelligence into the work and was never built. Even the labs now name their models to fit this. For years a new model meant a bigger version number, as if progress were a single line going up. OpenAI’s newest release is not one model with a higher number. It is a set of models split by job, an everyday one and a heavier one, which is the company accepting that a model is something you point at the right task rather than a finished product on its own.
If that is right, these leaders are not behind for the reason they believe. They do not lack access to the newest model. Most of them could not use the most powerful model in the world if it were handed to them tomorrow, because they have nowhere to put it. They are behind because the distance between what they could already do with cheap, ordinary, widely available intelligence and what they have actually deployed has grown, and it is widening from both ends. The top is sprinting away and pulling the ladder up behind it. The floor has dropped until capable intelligence costs almost nothing. And they are standing where they were, reading the headlines about the ladder.
I saw exactly this at an independent school not long ago, sitting with the whole leadership team, the head included. They had done the work. They had asked parents, pupils and former pupils what they thought, they knew they needed a real plan for how AI would run the school and support learning, and they were asking the right questions. Then I asked how many of them had used the agent tools, the ones that write and run code and operate software on your behalf. Every one had assumed those were for developers and had nothing to do with a school. They were making do with Microsoft Copilot and paying consultants to force it to behave. An hour with a few other systems changed the conversation. They could see the capability, and the strategy they had walked in with started to look like the wrong shape.
I find this more hopeful than not, and not as a consolation. What holds most people back is not a model they are barred from having. It is the work of building the system around the intelligence they can already get, almost all of which is cheap and freely available. That work is hard, but it is learnable, and unlike the model race it compounds. The model layer is becoming a commodity. The layer that holds your context, your judgement and your way of working, the part nobody can download, is not. Time spent there is not time spent chasing a launch. It builds something that stays yours.
So the question worth carrying out of all this is not whether you are behind on the model. On the model you are almost certainly fine. The better question, the one none of them was asking, is plainer and harder. What does your work actually consist of, task by task, and what would it take to build a system that lets ordinary intelligence do the repetitive part of it well. That is a question about your own operation, not about the labs, and it does not have an answer you can buy.
There is a version of the next few years where the convenient thing wins by default. The model already sitting inside your messaging app, the one that has learned how your company works without anyone deciding it should, becomes impossible to remove, and you end up renting your own context back from whoever arrived first. There is another version where you treat the system around the model as the thing worth owning and the model as a part you can replace. Both are open right now. The difference between them is not access to the best models. It is whether you used this stretch of time to build.
There is one more fear worth meeting head on, because it sits under a lot of the worry I keep hearing: that all of this mainly means fewer people doing less. It does not. Some tasks will be absorbed, and some roles built around them will not survive in their present shape, and that is worth being honest about. But it sits next to something the displacement story keeps missing. Rebuilding a company around this technology is an enormous job, and almost nobody has started. It needs people who can break a business into tasks, design the systems that carry intelligence into the work, and keep them honest once they run. That is years of work, and while it is being done it will ask for more hands than it frees.
It also opens problems that were never worth solving. For a long time there were narrow needs nobody fixed, because the fix cost more to build than it was worth. As capable intelligence falls towards costing nothing, that changes. Whole categories of problem that were never economic to touch become worth touching, and most have not been noticed yet, let alone built.
The clock in the headlines is not yours. The one that matters is running inside your own walls, and it has been running slow for three years. That is the one to watch. The reason it has run slow is the same reason there is so much still to build, more of it than we have begun to count.
If this was useful, consider subscribing. The next piece looks at what the system around the model is actually made of, and how a small team can build enough of it to stop renting its own context back from the companies that hold the frontier.
Craig Hepburn is an AI strategist and Perplexity Fellow. Twenty years building at the frontier of digital, from Microsoft and Nokia to Art Basel and UEFA. Now building at the frontier of agentic intelligence.
Great article and metaphor