AI systems are moving quickly from narrow assistants toward agentic systems that can retrieve information, make decisions, and execute workflows across systems. That shift gives IT leaders a practical problem: they need to invest now without betting everything on technology that may change again soon.
The most reliable answer is to focus on the parts of AI architecture that remain useful even as models improve. The source article identifies four foundations that support production-ready AI: data quality, context engineering, governance, and human expertise.
Data quality is the first scaling problem
AI models depend on the information they can reach. If that information is incomplete, badly organized, fragmented, or poorly governed, the output becomes less reliable. The source links poor data quality to AI hallucinations, bias, and unreliable results.
Many enterprises still work across legacy systems, inconsistent data structures, unclear ownership, and incomplete datasets. Those problems do not disappear because a model is powerful. AI can use data, but it cannot automatically repair the organizational and technical issues underneath that data.
Adnan Adil, CIO of Elastic, puts the point directly:
“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.”
For IT leaders, the practical implication is that an AI strategy cannot begin and end with model choice. It has to include connected data, clear standards, clean and labeled information, real-time retrieval, and ownership that makes the system maintainable.
The source also cites Gartner’s prediction that companies will abandon 60% of all AI projects through 2026 if they are not supported by AI-ready data. That makes data architecture more than a back-office concern. It is one of the main conditions for whether AI work reaches reliable deployment.
Context engineering turns data into useful input
Having data available is not the same as giving the model the right information at the right time. Context engineering is the discipline of selecting, organizing, and presenting the information a model needs for a specific query or task.
The source distinguishes this from prompt engineering. Prompt engineering focuses on how a request is worded. Context engineering designs the broader information environment around the model, including retrieval, structure, and the machine-readable presentation of data.
This matters because more information is not always better. If a model receives too much context, relevant details can be diluted. The source also notes that excess context can raise costs and slow response times.
Effective context engineering depends on a modern, unified data foundation, along with retrieval and memory systems such as retrieval augmented generation (RAG) and vector databases. It also requires decisions about what information matters, what should be excluded, and when different kinds of information should be used.
Adil summarizes the requirement this way:
“Minimum context, correct and current data, and machine-readable information are critical to effective context engineering,”
That framing is useful because it keeps context engineering grounded. The goal is not to surround the model with every possible piece of information. The goal is to supply the smallest useful set of accurate, current, structured information for the job.
Governance and observability make AI controllable
As AI systems become more embedded in workflows, organizations need to know how those systems use data, how they behave, and where they fail. The source presents governance and LLM observability as key tools for maintaining that control.
Without clear controls around retrieval, workflows, and model usage, AI systems may process far more information than needed. That can create inefficiency and raise operating costs through additional computing resources, token consumption, and API charges.
Governance also connects directly to security. The source notes that AI expands the attack surface through risks such as prompt-based data leakage, model vulnerabilities, and adversarial inputs. Sensitive information requires access controls, monitoring, and oversight.
Adil says essential controls related to security, granular cost management, project controls, data security, and architecture are often insufficient. The lesson is that governance cannot be treated as a later layer. It needs to be built into architecture, workflows, and decision-making from the beginning.
When governance is present early, observability becomes more effective. LLM observability and benchmarking help teams evaluate accuracy and utility over time, monitor adoption patterns, and adjust systems as conditions change.
The source says observability is also important for the ROI of AI initiatives because benefits are often indirect and business value depends heavily on adoption and use. A 2026 report from Elastic found that 85% of IT decision makers expect to enable LLM observability for their internal generative AI apps.
Adil explains the business value plainly:
“Observability is actually huge. We can use observability data for cost control, decision-making, and engineering efficiency,”
Human expertise remains part of the architecture
The source’s fourth foundation is people. AI architecture is not only infrastructure, data pipelines, and controls. It also depends on teams that can design, integrate, evaluate, govern, and adapt these systems as the technology changes.
Nearly 70% of respondents in Deloitte’s 2025 Tech Executive Survey report plan to grow teams in direct response to generative AI. The source frames this as a contrast to widely reported AI-related cuts.
Adil’s view is that people will shape whether AI becomes useful:
“We think the people aspect is largely what's going to make AI impactful going forward.”
As AI becomes more deeply embedded in operations, organizations need people who can govern workflows, evaluate outputs, redesign processes, and manage change. The source also points to the need for skills in prompt engineering, orchestration, and change management as tools become more autonomous.
Institutional knowledge matters here. Turnover can bring new ideas, but the source notes that it also carries costs in continuity, organizational understanding, and innovation. Human-centered strategy therefore needs to be part of AI execution, not a separate concern after deployment.
The durable bet is the system around the model
The source article’s core argument is that model progress should not distract leaders from the underlying system required to make AI reliable at scale. Data quality, context engineering, governance, observability, and human expertise are less likely to become obsolete than any single model decision.
That is especially important as AI systems evolve from single-task assistants toward more autonomous agents. The more responsibility these systems take on, the more they depend on accurate data, relevant context, transparent controls, visible performance, and skilled teams.
Adil closes the source article with a broader view of what this could mean for work:
“We fundamentally believe that with these tools, velocity of work will get much faster,”
For IT leaders, the practical path is clear. Build AI architecture around the durable foundations first, then let model choices evolve on top of a system designed for reliability, governance, and scale.