There is a moment at most technology conferences when the distance between the keynote stage and the real world becomes impossible to ignore. At the Databricks Data + AI Summit 2026 in San Francisco, that moment came early.

One remark from the keynote stayed with me: AI does not have an intelligence problem. It has a context problem. It neatly captured an issue many organisations have been grappling with over the past two years.

The question was never whether models could reason, generate or execute. The question was always what happens when those capabilities are applied to real enterprise data, fragmented across systems, inconsistently governed and rarely structured for the decisions the business actually needs to make.

The gap between the AI pilot and production

This is not a new observation. But the scale at which it is now manifesting is. In conversations with mid-market enterprises across the United States, and in discussions throughout the week in San Francisco, the same gap kept surfacing: most organisations are not behind on AI capability. They are behind on data discipline. That gap determines whether a pilot becomes a production system or remains a pilot indefinitely.

Representatives from companies across industries were not discussing AI models as much as they were discussing governance, data freshness, cost management and ownership. The questions that surfaced most often were not about which model to choose. They were about what happens when an AI agent queries a live database autonomously without governance boundaries in place, or when business decisions are made using data that is already hours out of date.

One practitioner summed it up well: building the agent is roughly 1% of the work. The remaining 99% is the infrastructure no one planned for. In most cases, that infrastructure gap traces back to the same root cause: the data layer was never made ready for what came next.

Data readiness is not a preparatory step. It is the work

The pattern I see repeatedly, both at industry events and in conversations with organisations adopting AI, is that data readiness is treated as a prerequisite, something to complete before the real AI work begins. It is not. It is the real AI work.

What 'data ready' actually requires is more demanding than most enterprise AI roadmaps acknowledge. It means a unified data layer where analytical and operational data are accessible from the same governed source. It means a semantic layer where AI systems understand what data means in business terms, not just what it contains. It means governance that is auditable at the agent level. And it means data quality that is continuous rather than a one-time clean-up exercise.

Modern data infrastructure platforms are increasingly converging around these same requirements. The technologies may differ, but the underlying challenge they are trying to solve is remarkably similar. Organisations that recognise this are asking the right question: not which platform to choose, but whether they are building the governance and data quality foundation that will make any platform worth the investment.

The mid-market is solving a different AI problem

One assumption that surfaced repeatedly in conversations around the summit is that enterprise AI maturity is largely a function of scale. In practice, the distinction is less about size and more about operating reality.

Large enterprises often have the resources to absorb complexity. They can build dedicated AI teams, establish governance functions and invest in multi-year transformation programmes. Their challenge is orchestrating AI across thousands of employees, business units and legacy systems.

Mid-market organisations operate under very different constraints. They rarely have the luxury of specialised AI teams or extensive implementation capacity. Instead, lean IT functions are expected to modernise decades of accumulated applications while keeping day-to-day operations running. Every AI initiative must therefore deliver measurable business value without introducing another layer of operational complexity.

What makes these organisations particularly interesting is that they often move faster once the data foundation is in place. Decisions are made closer to the business, governance is less fragmented, and business leaders are more directly involved in defining the outcomes AI is expected to deliver. The limiting factor is rarely the willingness to adopt AI. It is the readiness of the underlying data estate.

What matters now

The conversations that will determine enterprise AI outcomes over the next three years are not the ones happening in model selection committees or vendor briefings. They are the ones happening, or in many organisations, not yet happening, between data leaders and business owners about what the underlying data estate is actually capable of supporting.

Most leadership teams have spent the past two years asking the right questions about AI capability. The question that has not received proportionate attention is simpler and more fundamental: is the data estate that AI will run on fit for purpose? Not in principle, not in a future-state roadmap, but now, with the governance it currently has, the quality controls it currently runs and the unified access it currently provides, or does not.

The organisations that are pulling ahead in enterprise AI did not get there by moving faster on models. They got there by making an unglamorous decision earlier than their peers: treating data readiness as core infrastructure rather than a preparatory step. That decision compounds over time. Every new AI capability that emerges will deploy faster, govern more effectively and deliver more reliably for organisations that invested in their data foundations. For those who did not, the gap does not remain constant. It widens with every new wave of innovation.

The author is a technology entrepreneur and co-founder, Cymetrix Software, a Wondrlab Company.