Elastic CIO Adnan Adil is telling enterprise technology leaders to stop treating AI architecture as a race to whatever model demo shipped last week. In custom content produced by MIT Technology Review Insights in partnership with Elastic, Adil argued that the parts likely to endure are less glamorous: usable data, controlled context, governance, observability and people who understand the business.
The argument lands as companies expand from AI assistants toward agentic systems that can fetch information, make decisions and run workflows across software systems. Those systems need more than a model endpoint. They need access to the right internal data, rules about what they may use, monitoring that shows when they fail and staff who can redesign processes around them.
Data quality remains the first choke point
Adil said data is a “durable” part of AI architecture because models cannot provide useful service without the right context. The problem is familiar to any CIO with a basement full of legacy systems: fragmented data ownership, inconsistent schemas and incomplete records do not magically become clean because a large language model is pointed at them.
Deloitte surveys cited in the MIT Technology Review Insights piece identify data quality as a major barrier to AI success. Gartner, also cited, predicts that through 2026 companies will abandon 60% of AI projects that lack AI-ready data. The prescribed fix is not exotic: clear standards, named owners, cleaned and labeled data, governance and retrieval pipelines that can serve current information when the model needs it.
Context engineering replaces prompt theater
The Elastic-backed piece also draws a line between prompt engineering and context engineering. Prompt engineering concerns the wording of a request. Context engineering covers the information environment around the model: what gets retrieved, how it is structured and what is left out.
That distinction matters because dumping more material into a model can make answers worse and bills higher. Extra context can bury the relevant facts, slow responses and increase token usage. Adil said effective context engineering requires minimal context, current and correct data, and machine-readable information.
The technical pieces named include retrieval-augmented generation and vector databases, but the mechanism is less mystical than vendors sometimes imply. The system has to locate relevant enterprise information, format it so the model can use it and avoid feeding the model documents it does not need.
Governance and observability cannot be stapled on later
Adil also argued that governance has to be designed into AI systems from the start. The controls he named include security, project oversight, data protection, cost management and architecture. Without those constraints, AI applications may retrieve or process more information than required, raising compute costs and creating security risk.
The piece points to prompt-based data leakage, model vulnerabilities and adversarial inputs as new exposures created by AI systems. It says companies need access controls, monitoring and oversight to protect sensitive information.
Observability is the companion piece. Elastic’s 2026 report, cited by the company, says 85% of IT decision makers expect to enable observability for internal generative AI applications. Adil said observability data can support cost control, engineering efficiency and decision-making. In practice, that means measuring how AI applications behave after launch: whether they are accurate, whether staff use them, where they fail and whether they are producing the expected business value.
Humans remain part of the stack
The final plank is staffing. Deloitte’s 2025 Tech Executive Survey, cited in the piece, found nearly 70% of respondents plan to expand teams in response to generative AI. Adil said people will be a major factor in whether AI becomes useful.
That claim cuts against the lazier version of the AI automation story. More autonomous tools still need employees who can evaluate outputs, govern workflows, adjust prompts and orchestration, manage change and preserve institutional knowledge. Adil said parts of the AI stack are changing quickly, while institutional knowledge and adaptability remain durable.
Elastic’s position is straightforward and self-interested, as infrastructure vendors tend to be: the model is only one layer. The less flashy layers underneath may decide whether enterprise AI becomes production software or another expensive pilot with a chatbot pasted on top.
This story draws on original reporting from MIT Technology Review.
