We are all a little hooked on the model

Since Claude Fable 5 was approved for use again, I have been building roughly five times more than I did before. Anthropic keeps extending my access in small steps, a few days here, a week there, and today the window was extended again, to 19 July. It feels a little like a dealer sampling me into dependency, and I say that fondly, because I am a very willing customer.

What I notice is what this does to me. I plan around the model. I wait for it. I have started building my week on something I do not control, and I am not sure I would have admitted that a month ago.

A few of the things I read last week:

• AI super-users are about 5x more productive than non-adopters, yet only 29% of organisations see significant ROI from generative AI (Writer / Workplace Intelligence, AI Adoption in the Enterprise 2026).
• Ethan Mollick let a model work on its own for about fourteen hours and it came back with software he estimates at two to seventeen weeks of human engineering, for a token bill of roughly $251; what predicted getting real work out of the tools was domain expertise, not job title (Ethan Mollick, One Useful Thing).
• Otto Scharmer warns against an "intelligence monoculture" and asks a plain budget question: what share of a leader's attention and money goes to automation, and what share to sensing and judgement (MIT Sloan Management Review).
• At a firm that told engineers to double their merged pull requests, throughput reached 2.09 times the old baseline and the load moved onto the human reviewers, whose work roughly doubled (field study, arXiv).
• SaaStr runs its go-to-market with about 20 AI agents and roughly 1.2 humans; Jason Lemkin says training the agents took "30 days, not 30 minutes" and the oversight is daily, not optional (Lenny's Podcast).

Reading these next to my own week, I think they are all describing the same imbalance.

The fast part is the part we do not own

The Writer study puts numbers on something most of us can feel. The super-users really are about five times more productive, and yet only 29% of organisations see significant returns. My own five times is real, and it is exactly the number companies are chasing when they pour attention into every new model and every new tool. But the gains sit with individuals, and the organisation around those individuals stays as it was.

That is what model dependency looks like from the inside. We plan around the model, we wait for it, we build our weeks on it, and none of us owns it. The thing improving fastest is the thing we least control, while the things we fully control, our own workflows, our roles, our capacity to review work, are the things almost nobody redesigns. Otto Scharmer's budget question is useful because it is so plain: what share of our attention and money goes to automation, and what share to sensing and judgement? I suspect in most organisations the answer would be uncomfortable to say out loud.

The road is the part we own

I wrote this morning that giving someone a Lamborghini does not help much when the road still has the same speed bumps. The five-times person still slows down for the same approval chain and the same handoffs as before, so the road has to be redesigned first. The arXiv study shows what happens when it is not. A firm told its engineers to double their merged pull requests, throughput reached 2.09 times the old baseline, and the extra work moved onto the human reviewers, whose load roughly doubled. The car got faster. The road did not change.

There is a fair objection to all of this. Chasing the newest model is rational, because the capability jumps are real, my own five times proves it, and waiting until the organisation is ready can mean falling behind people who did not wait. I do not think anyone chasing models is being careless. But my answer would be: take the model, enjoy the five times, and spend at least as much on the road as on the car. The model gains arrive on someone else's schedule and they can be taken away. My access ends on 19 July unless someone decides otherwise, and everything I have planned around it goes too. The redesigned workflow is different. It compounds, and it stays ours. SaaStr makes the same point from the other side: twenty agents and 1.2 humans sounds like pure model magic, but Jason Lemkin says training them took thirty days, not thirty minutes, and the oversight is daily. Most of the work was road work.

If you work in L&D, HR, or transformation

Two small things we could do this week, and I mean this week. First, map what breaks if a model we rely on disappears next Friday. Which processes stop, who gets stuck, what have we quietly stopped knowing how to do ourselves? Mine comes with a real date attached, so I will be doing this exercise whether I want to or not. Second, pick one handoff or one approval chain and redesign it, instead of piloting another tool. The pilot is more exciting, I know. The redesigned handoff is the one that will still be there in a year.

The provocation

So the question I want to sit with, and I would genuinely like to hear how others answer it: this quarter, is the money and attention going to the car or to the road? If we looked at our calendars and budgets for the next three months, which one would they say we believe in?