The Unsolved Part is Ambition

Around 2010, an air-quality monitor on the roof of the U.S. Embassy in Beijing began publishing real-time PM2.5 readings. The technology was unremarkable. China already monitored air quality, and pollution was not a new problem. What changed was that an invisible constraint became impossible to ignore. The conversation moved from whether pollution was a problem to how serious it was and what to do about it. Xinjin Zhao tells this story to make a point worth holding onto: many of the most consequential changes begin not with a technology but with a constraint that leadership translates into a clear ambition.

The AI race has the same shape, and most organisations are reading it backwards.

Today's competition is framed around model capability — compute, chips, context windows, open versus closed weights. That is the part vendors compete on, and they are competing hard. It is largely settled as a buying problem: the capability you need is available, from several providers, at a price that keeps falling. The unsolved part is ambition. What decision is an agent actually allowed to make? Inside which workflow? Toward which outcome? Against which constraints that are non-negotiable? Those questions stay invisible until something breaks — and then they are answered in an incident review instead of a design session.

Zhao reaches for Apollo, and it holds. Kennedy did not ask NASA to be more innovative, and he did not specify the rocket. He named the objective, the success criteria, and the timeline, then left enormous room for innovation underneath. Thousands of people could work independently because the ambition above them was clear. Most enterprises deploying AI do the reverse. They buy the technology first and define the ambition never. The result is dozens of disconnected pilots, each locally reasonable, none aligned to a decision the organisation has actually made about where it is going.

Defining that ambition is not a vision statement. It is architecture work, and it has structure. Where is the organisation going, and why. How decisions get made, and by whom. How work flows end to end. What rhythms keep it aligned. Only then: where AI and automation sit inside all of that. The order is not decoration. Each layer is a prerequisite for the next, and the agents everyone wants to deploy sit at the end of the sequence, not the start. An agent is a decision-maker you have hired without an interview. It inherits whatever decision architecture you already have. If that architecture is implicit, the agent automates the implicit version — at speed, at scale, and without anyone having chosen it.

This is why so many AI deployments underperform without failing visibly. The technology works. The model returns a plausible answer. But the answer is in service of an ambition nobody defined, so it produces motion without direction. The constraint that would have made the work cohere — a clear statement of what the system is for — was never made visible, the way the embassy monitor made air quality visible. Nobody aligned around it because nobody could see it.

The harder task for leadership, as Zhao puts it, is not encouraging innovation but defining ambition. We would put it more bluntly. Architecture built before deployment is a competitive advantage. Architecture built after an incident is damage control. The winners in this wave will not be the organisations with the most advanced model. They will be the ones who decided, in advance and on purpose, what they wanted the technology to do — and built the operating architecture that makes that decision real.

→ Want to know which decision your organisation is actually automating? The Architecture Index is a 15-minute, self-serve read on exactly that — no call required: index.mandantis.com