Four AI Architecture Fundamentals That Outlast the Model Hype Cycle

IT leaders weighing which AI investments will still matter in six months are being pointed back to four architectural fundamentals: data quality, context engineering, governance and observability, and human expertise. A new report published by MIT Technology Review Insights in partnership with Elastic argues these elements stay relevant regardless of how fast the underlying models change.

The framing matters because the churn is real. Organizations are expanding AI use cases and moving toward agentic systems that retrieve information, make decisions, and run workflows across multiple systems. That pace introduces risk, and the report positions structural fundamentals as a hedge against betting on capabilities that may be obsolete by the next release.

Data quality is the durable base layer

The report puts data first, and the stakes are quantified: Gartner predicts companies will abandon 60% of all AI projects through 2026 if they lack AI-ready data. Most enterprises are working against themselves here, relying on legacy systems, inconsistent data structures, fragmented ownership, and incomplete datasets.

The uncomfortable point is that AI cannot fix the problem that limits AI. Poor data quality produces hallucinations, bias, and unreliable outputs, and no model upgrade compensates for a broken foundation.

"The data is a durable part of AI architecture because without it, these models won't run, won't provide the right context, or won't give the right level of services that we're looking to implement," said Adnan Adil, CIO of Elastic.

The recommended baseline includes clear data standards and ownership, clean and labeled data, and pipelines built for real-time retrieval, ideally designed into the architecture from the start rather than retrofitted.

Context engineering, not just prompt engineering

The report draws a distinction that is becoming central to production AI work. Prompt engineering concerns how a request is worded. Context engineering designs the entire information environment around the model, retrieving the right data and presenting it in a structured, machine-readable form.

More context is not better. Feeding a model excessive information can dilute the relevant details, raise costs, and slow response times. The discipline is largely about exclusion: deciding what matters, what to leave out, and when different types of information apply. It leans on retrieval and memory systems such as retrieval augmented generation (RAG) and vector databases sitting on a unified data foundation.

"Minimum context, correct and current data, and machine-readable information are critical to effective context engineering," Adil said.

Governance and observability, built in early

The report's third element ties technical control to cost discipline. Without clear controls around retrieval, workflows, and model usage, AI systems tend to process far more information than needed, and that inefficiency shows up directly in token consumption and API charges.

Governance also widens into security. AI expands the attack surface with risks including prompt-based data leakage, model vulnerabilities, and adversarial inputs, which the report says require strong access controls, monitoring, and oversight. The argument is that these controls cannot be added as an afterthought and need to be embedded in architecture and workflows from the outset.

Observability follows from governance. According to a 2026 Elastic report, 85% of IT decision makers expect to enable LLM observability for their internal generative AI apps. The rationale extends beyond monitoring uptime: because AI's business value is often indirect and depends on how systems are actually adopted, observability data is what lets teams measure performance against expectations and justify ROI.

"Observability is actually huge. We can use observability data for cost control, decision-making, and engineering efficiency," Adil said.

People remain the constraint

The fourth element cuts against the narrative that AI reduces headcount. Nearly 70% of respondents in Deloitte's 2025 Tech Executive Survey said they plan to grow teams in direct response to generative AI, a contrast with widely reported AI-related cuts.

The skills in demand shift toward orchestration, prompt engineering, and change management, alongside people who can govern workflows, evaluate outputs, and redesign processes as conditions change. The report also flags a cost that rarely appears in AI budgets: turnover erodes system continuity and institutional knowledge, both of which are hard to rebuild.

"Many aspects of the stack are moving very, very fast, but institutional knowledge and the ability to adapt remain durable," Adil said.

The through-line across all four elements is a deliberate bias toward the parts of an AI system that do not need replacing every time a new model ships. For teams across Asia-Pacific building AI on top of fragmented legacy data estates, the data-readiness figure is the one worth taking seriously before committing to agentic deployments.