By PVN Pavankumar
Two companies, same industry, same size, run the same AI agent on the same model against the same problem. One pilot becomes part of how the business runs. The other dies in a slide deck. The model was identical. I have spent the last year designing the north star architecture for how a very large enterprise becomes AI-native, and this is the lesson almost no one selling AI will say out loud. Everyone can buy the same frontier models. A capability everyone can buy is not an advantage. The advantage is the context you feed the model, and context is the one thing no one can buy.
A supplier is quietly becoming a problem. Procurement has logged late deliveries, finance is sitting on an unresolved dispute with the same vendor, legal is renegotiating the contract. Drop the most capable model in the world into the procurement system and it will tell you, fluently, that two deliveries were late. It will say nothing about the dispute or the contract, because those live in systems it cannot see. This is AI-first: capable, but boxed inside one application. AI-native is the opposite, intelligence that spans the whole landscape.
It records, but it does not remember
For fifty years, enterprise systems have stored what happened with perfect fidelity: a purchase approved, a discount granted, a ticket closed. What they almost never store is why. Two customers identical on paper receive very different discounts, and the system cannot tell you why, because the reasoning was never written down. A human in the room could explain it in thirty seconds. The system cannot, and neither can the agent built on it, so it proposes the average answer, not the right one. The shift now underway is exactly this: from systems of record that capture transactions to systems of context that capture the reasoning behind them.
The moat you cannot buy
Context is what closes that gap: the decisions, exceptions, and precedents that explain how your business actually behaves. No rival can license it, because no one else has lived your decisions. And it compounds. Better context produces better decisions, better decisions leave behind a richer trail, and that trail makes the next decision sharper still. This is the flywheel, and the loop feeds itself. Data was the moat of the last decade. Context is the moat of the next.
The model reasons. The harness governs.
The most underestimated part of an AI-native architecture is the harness: the enterprise controls wrapped around the model. Identity, so an agent acts as a governed participant. Authority boundaries, so it touches only what it is permitted to. Observability, so you can see what it did and why. The same model in two different harnesses performs at two completely different levels. The harness, not the model, sets the ceiling.
Three disciplines are forming around it. Context engineering assembles the right slice of information for each task. Harness engineering wraps models in the controls above, so governance is built in rather than bolted on. Specification engineering defines agent behaviour precisely enough to test whether it works.
The warning for India
This warning lands hardest here. Much of our technology industry was built on running and extending enterprise systems with skilled people, and the instinct now is to pitch AI as a faster way to do that same work. That is the trap. Speed without context produces pilots that impress in a demo and quietly stall in production. The firms that win will not be the ones that deploy an agent first. They will be the ones that turn decades of hard-won judgment, the kind that lives in engineers’ heads and undocumented code, into structured context no rival can copy.
Context is not just memory. It is the mechanism through which an enterprise learns. Successful decisions are retained and propagated into future ones, much as evolution preserves the genes that work and carries them forward across generations. Organisations that do this well become self-improving systems. Those that do not are condemned to relearn the same lessons over and over again.
The smartest model will not decide who wins the next decade. Everyone will have it. The best-remembered context will.
The writer is VP, Office of the CTO at SAP
Disclaimer: The views expressed are the author’s own and do not reflect the official policy or position of Financial Express.
