Jensen Huang describes AI as a stack of five layers. Energy at the bottom, then chips, then the infrastructure that runs them, then the models, and at the top, the applications people actually use.

The first four layers get almost all the attention. Model launches make headlines. Chip supply moves markets. Data centre deals are announced in gigawatts. If you followed AI only through the news, you would think the race is won or lost somewhere in those four layers.

For most organisations, that race is already decided, and it was never theirs to run. Very few will build a model, design a chip or operate a hyperscale data centre. They will buy those layers the way they buy electricity, and the price and quality will keep improving whether or not they pay attention.

That leaves layer five. The application. It is the only layer most organisations actually control, and I think it is where nearly all the value will land.

Layer five is easy to misread as “using ChatGPT”. A model on its own answers a question. An application decides which question gets asked, what information the model is allowed to see, what it is allowed to do with the answer, how anyone knows the answer was right, and who is accountable when it was not. The model is one component. The rest is the application, and the rest is most of the work.

The people building AI agents have started giving that rest a name: the harness. An agent is a model plus its harness, and the harness is everything except the model. The instructions, the tools, the checks, the limits, the loop that decides what happens next. You rent the model. You build the harness, and it accounts for much of the difference between two systems running the same model.

That is also where most AI projects stall. The demonstration works in a meeting. The service has to work on a Tuesday afternoon, for the wrong user, against a stale record, with an auditor reading the logs a year later. Closing that gap is real engineering, but most of it is engineering organisations already understand: identity, permissions, evaluation, records management, procurement, change control. It needs applying with care to a component that behaves differently from ordinary software.

There is a second reason layer five matters, and it is close to home. The first four layers are largely global. Broadly the same models run in Melbourne as in San Francisco. Layer five is local. It is shaped by privacy law and data sovereignty, by procurement rules, by how a organisations actually processes an application or how an agency handles an information request. That context does not come in the box, and it does not travel well. The organisations that do best at layer five will be the ones that understand their own context deeply and design for it.

That is the work I care about, and it is what this publication is for. Each post will take one decision at layer five and work through it: whether something needs an agent at all, what to measure before you trust it, what it should never be allowed to do, and how to tell a promising pilot from a service that will hold up. Practical, specific, and grounded in how things actually work here.

I will write roughly once a fortnight. If you build, buy or govern AI systems, particularly in Australia, I hope it is useful.

Working at layer five.

Ujjwal