From Building to Bounding
Anthropic has said its AI now writes more than eighty per cent of the code going into its own products. At the same time, the same company called for the world to have a way to halt frontier AI development if it ever moves too far, too fast.
Those two facts belong together. The first is a capability statement: the tool has started building the tool, and doing most of the building. The second is the company that builds it saying, in public, that someone needs to be able to pull the brake.
It is easy to file the first fact under “impressive” and ignore the second. Most of us instinctively do, and that gets it backwards. The capability is becoming the cheap, abundant part; the brake is the hard, scarce part. And most leaders are still spending their attention on the wrong one.
For almost the whole history of business, capability has been the thing in short supply. We were limited by how much skilled production we could afford: how many engineers, analysts, designers, writers we could hire and keep busy. Strategy was, in large part, the art of spending a scarce resource well.
That resource is becoming abundant, not infinite and not free, but abundant enough that, for most knowledge work, “can we build it” is no longer the binding constraint. When a capable system can draft the first version of almost any document or block of code in minutes, the building stops being the bottleneck. And when a constraint dissolves, the value tied up in it does not vanish. It moves to whatever is scarce next.
What is scarce now is not building. It is bounding.
The people who build serious autonomous systems already understand this, even if the language has not reached most boardrooms yet. Read any careful account of how a production AI agent is designed and the same preoccupation shows up. The clever part is largely taken for granted. The hard part is working out how much damage the thing can do when it is wrong, because it will be wrong sometimes, and it now acts faster than a person can step in. So you design, from the start, a hard limit on how far a single mistake can travel: what the system can touch, what it can change, what it must hand back to a human before it acts. Engineers call it blast-radius control. It is the same instinct Anthropic was reaching for with its halt proposal, brought down to the level of a single workflow.
Learning to contain capability, alongside the rush to acquire it, is about to become one of the central disciplines of running an organisation. And the natural reaction points the wrong way.
When production gets cheap, the urge is to produce more (or to bank the saving and cut costs). The instinct is to wire in another agent and automate another step, as fast as possible. It feels like progress, and it makes a good slide in the quarterly update. But unbounded capability is a liability that has not gone wrong yet. An organisation that has wired autonomous systems into its operations without deciding clearly what each one is allowed to do has not pulled ahead. It has taken on a debt that comes due the first time one of them acts confidently on something false.
The organisations that come out of this phase in front will be the ones that decided most clearly what should be built at all, and drew the cleanest lines around what their systems are permitted to touch. When one of those systems goes wrong, and some will, a bounded one keeps the damage closer to small. That is harder and less glamorous than launching another pilot. It is also the work that lasts, because no model upgrade does it for you.
This is what we find ourselves arguing for, again and again. The scarce skills in the AI era sit elsewhere. They are judgement, deciding what should exist; constraint, deciding what the machine may decide on its own and what comes back to a human; and accountability, naming who owns the outcome when a system acts in your name. Those are design decisions about the shape of the work itself. They are what we mean by restoring the true form of the work: rebuilding the operation so the machine does its building inside boundaries you set on purpose. That is the opposite of bolting AI onto whatever is already there.
There is a plain question that tends to find the gap. Somewhere in your operation, something is already being built or decided by a system faster than anyone can check it. Where is it? And who drew the boundary around it? If the first half is easy to answer and the second is not, you have found the first piece of work worth redesigning.
The building is getting cheaper by the month, so the bounding has become the work. Find the one place a system is already running ahead of you, and draw the line there first.
Mark Bunce is the founder of Trueform Consultancy, focused on redesigning how work gets done using modern technology.