For much of the last two years, enterprise AI progress was often measured in scale. Bigger context windows, longer prompts and larger document dumps became shorthand for smarter systems, as organisations tried to improve model performance by feeding assistants entire repositories of files, chats and knowledge bases. But as those experiments moved from demos into production, a more difficult lesson emerged: more information does not necessarily mean better decisions.
That is the argument Elastic is making as it sees customers shift from pilot projects to live deployments. According to Ravindra Ramnani, Head of Field Engineering, India, at Elastic, the real question is no longer how much data an AI model can accept, but whether it is being given the right data at the right moment. "The conversation therefore shifted from 'how much can we give the model' to 'what does the model actually need to know.' That question lies at the heart of context engineering," Ramnani says.
The concept is gaining traction as enterprises discover the limits of what some in the AI world have jokingly called tokenmaxxing, a trend that began in Silicon Valley in early 2026 and treated high token consumption as a badge of productivity. In some companies, internal leaderboards ranked employees by tokens burned, turning usage into a status symbol regardless of whether the output was useful. Ramnani dismisses that mindset as a poor proxy for value. "It is the AI-era equivalent of judging developers by lines of code," he says.
The same logic, he argues, has seeped into system design. Teams often assume that if a model can process more context, they should give it everything: instructions, historical conversations, archived documents and related records. But models do not automatically distinguish signal from noise the way a skilled analyst would. The result can be slower responses, higher costs and answers that still require heavy human cleanup because the model has been distracted by irrelevant material.
That problem becomes far more serious once AI is no longer just drafting text but taking action. In production environments, an agent working from incomplete or poorly curated context can make consequential mistakes. In sectors such as banking compliance or security operations, a wrong autonomous action can propagate before a human has time to intervene. The issue is not simply wasted compute. It is operational risk.
Context engineering, as Elastic describes it, is the discipline of deciding what a model should see at each step of its reasoning. That includes the instructions that frame the task, the tools available to the model, the information retrieved from enterprise systems, the memory carried over from earlier turns and, just as importantly, the material deliberately left out. In enterprise settings, retrieval becomes the hardest part because the model must be grounded in trusted internal knowledge without being buried under it.
Elastic frames the distinction with a simple analogy: tokenmaxxing is like handing someone an entire filing cabinet and asking them to find the relevant page. Context engineering is giving them the three pages they actually need, already tabbed. Both contain information. Only one is designed to support better decisions.
The company says relevance is not just a search problem but a safety mechanism. For autonomous AI to act reliably, the information it receives must be smaller in volume, higher in signal and governed by permissions so the agent can only access what it is authorised to use. That combination of relevance and control, Elastic argues, is what separates a dependable production system from an expensive proof of concept.
The shift is especially relevant in India, where enterprises across banking, technology, telecom and large-scale services are under pressure to deploy AI quickly while maintaining compliance and operational discipline. As organisations move beyond experimentation, the appeal of bigger prompts is being replaced by a more practical question: how do you make AI useful, accurate and safe in real workflows?
Elastic's answer is that the next phase of enterprise AI will not be won by the largest context window alone. It will be won by systems that know what to include, what to exclude and how to ground every response in the right enterprise knowledge. In that sense, the industry's focus is moving away from tokenmaxxing and toward a more mature discipline: context engineering.