Key Takeaways

  • Satya Nadella rejects the idea that vertical monopolies from silicon to models will own enterprise software, betting instead on horizontal specialization.
  • Enterprise deployments depend on model routers and decision engines that dispatch discrete jobs to the best model for each task.
  • A new services category is emerging that blends business evaluation with agent architecture, distinct from classic systems integration or management consulting.
  • Consumer giant Unilever automated order-to-cash and trade claims by building an agentic workflow around business evals.

The Death of the Monolithic AI Stack

Tech history tends to swing between integrated stacks and horizontal layers. Right now, founders worry that hyper-scalers or frontier labs will capture the entire value chain from silicon chips to end-user applications. Nadella does not buy that thesis.

Nadella expects horizontal layers to win out over single-vendor walled gardens. As he told Harry Stebbings, “my hope is that this becomes a little bit of a horizontal specialization with multiple winners in each layer and then the ability for people to compose.”

If Nadella is right, the winning architecture will not bind an enterprise to one provider. Instead, systems will sit above the models. “If you're building an application like an agentic system you could even have something like what I said which is a router with even a decision engine built into it and then you just route to the best model for a given task,” Nadella noted. In that world, raw model weights matter less than the orchestration logic that directs traffic based on latency, cost, and reliability.

The New Implementation Economy

Enterprise software historically required massive armies of systems integrators like Accenture or management consultancies like McKinsey. Nadella sees an entirely new category forming around autonomous agents, but it does not fit older operational templates.

“This one I think has different contours right because it's the combination of tech and business if I could say it's kind of... not classic management consulting, it's not classic systems integration, and it's not classic sort of technology support,” Nadella explained.

Consider Unilever. The consumer goods company tackled a back-office headache: trade claims and order-to-cash processing. Retailers deduct trade claims for promotions, damaged goods, or late shipments, creating massive manual reconciliation backlogs. Instead of hiring traditional consultants to patch their legacy ERP, Unilever built an agentic workflow. Nadella observed that “it started with what's the business, what's the eval, what's the way to conceptualize the agentic architecture. So it's sort of business consulting combined with technical consulting.”

In this setup, the real moat is not training a bespoke model from scratch. The moat is creating private evaluation benchmarks that capture actual business rules, then structuring agent architectures to hit those targets.

What to Do With This

Audit your product roadmap and strip out any hardcoded dependencies on a single frontier model. Build a simple routing layer this week that switches between two competing model providers based on task cost and response latency. If your application cannot swap underlying model providers with a single configuration flag, you are building an integration, not a defensible product.