The Operating Model Is the Real AI Harness

We finally stopped comparing models. I think we've stopped one layer too early.

When generative AI first entered the enterprise, almost every conversation revolved around models. Which model was smartest? Which benchmark mattered? Which provider was winning? Capability was treated as though it lived inside the model itself, so choosing the right model became the central strategic decision.

Over time that explanation became harder to sustain. Organisations discovered that the same model could produce radically different outcomes depending on the software wrapped around it. Context management, routing, memory, tools, evaluation and deterministic orchestration often mattered more than marginal improvements in the underlying model. The industry responded by shifting its attention towards harnesses. That was an important step forward because it acknowledged something that should have been obvious from the beginning: models never operate in isolation.

I think we’re about to make the same discovery again.

The assumption today is that capability lives in the harness rather than the model. Build a better harness and you’ll build a better intelligent system. That certainly explains more than the old model-centric view, but I don’t think it explains enough. The same model deployed through a similar harness can still produce radically different outcomes. The more I look across different organisations, the less convincing the usual explanations become. We often attribute those differences to execution quality or organisational culture, but I increasingly suspect both are symptoms of something more fundamental.

Every harness inherits the operating model it is built to serve.

That sounds abstract until you look at what a harness is actually trying to do. It has to coordinate work, retrieve knowledge, invoke tools, route decisions, escalate exceptions and operate within whatever governance exists around it. None of those responsibilities originate inside the harness. They originate in the operating model. The harness is where the operating model becomes software.

When I talk about the operating model, I don’t mean the organisation chart. I mean the system through which work actually happens. How decisions move through the organisation. Where accountability sits. How information flows. How teams coordinate. How humans and intelligent systems divide responsibility. Those aren’t background conditions sitting behind an AI implementation. They become part of the implementation because every intelligent system inherits them.

Seen from that perspective, the progression over the last few years reveals a pattern. We began by assuming capability was primarily a property of the model. We then discovered that capability emerged from the relationship between the model and its harness. I increasingly think we’re about to discover that the capability of the harness itself emerges from the operating model of the organisation in which it is deployed. Each layer constrains the possibilities of the layer beneath it because each layer inherits the conditions created by the one above.

That's why I find myself becoming less interested in model comparisons and more interested in organisational design. The strategic question is no longer which model to buy or which harness to build. It's whether the organisation itself has been designed to allow intelligent systems to succeed. Perhaps the better question is whether intelligent systems change the assumptions the organisation was designed around in the first place. Two organisations can make identical technology decisions and still experience radically different outcomes because the technology is inheriting two completely different operating models.

I’m beginning to think we’ve been treating organisational design as an implementation problem when it may actually be the primary design problem. The model still matters, and the harness still matters. But perhaps we’ve misunderstood what a harness really is. We talk about it as though it’s another software layer. Increasingly, I think it’s the software expression of the operating model itself.

The software doesn’t escape the organisation. It implements it. If that’s true, then designing intelligent systems and designing organisations are becoming the same problem.

◆

Stuart Winter-Tear

Independent AI Advisor | Author of UNHYPED

unhyped.pro