The Databricks Data + AI Summit has experienced near-exponential growth over the past several years. As regular attendees, we have watched it transform from a technical, data engineering conference into one of the largest corporate enterprise data and AI events in the world.
With more than 31,000 in-person attendees filling the Moscone Center in San Francisco this year, the sheer scale of the audience reflected how far interest autonomous, agentic AI and the trusted data that powers intelligent decision-making has extended beyond the technical teams who build these systems. Business leaders, product managers, solution architects, and C-level executives are now actively part of the conversation, which speaks volumes about the cross-functional reality of AI adoption today.
A year ago, most of the conversations centered on raw model performance, parameter counts, and what generative technology might be capable of someday. This year, the conference felt grounded in the operational realities of deploying AI inside a real enterprise. Everyone knows that frontier LLMs work. Where people now struggle is how to feed them the right data, govern them responsibly, control runtime costs, and make them useful at scale.
That change was reflected in the event’s biggest announcements, which centered around four themes that appeared repeatedly across keynotes, product launches, and customer stories: context, control, cost, and choice. Together, these principles outlined Databricks' vision for helping organizations move AI from experimentation into governed, production-ready systems.
At the summit, little time was spent debating model intelligence. Databricks made it clear that today's models are already capable; the bottleneck is helping them understand the specific businesses they are supposed to support.
Ali Ghodsi, CEO of Databricks, captured this idea early in his opening keynote:
"AI does not have an intelligence problem. It has a context problem."
He added:
"Getting this context into the AI is harder than one could imagine. It's actually really, really difficult and getting the perfect enterprise context is elusive."
Anyone who has tried moving AI beyond a prototype will recognize the problem. Enterprise knowledge is fragmented across documents, meetings, legacy systems, messaging tools, and operational databases, all using slightly different language and definitions. Without a shared grounding layer, agents hallucinate, misinterpret internal logic, or fail outright.
Databricks’ answer is an enterprise context layer built into the platform. The headline announcement was Genie Ontology, a living system that maps and connects organizational knowledge across tools like SharePoint, Google Drive, email, calendars, and the Lakehouse.
Rather than relying on static semantic models that require constant manual upkeep, it uses OntoRank to dynamically surface the most relevant and authoritative business context as data changes.
This directly powers Genie One, which has evolved from a text-to-SQL interface into a more capable AI coworker. It can draft reports, surface anomalies, and initiate downstream workflows based on a continuously updated understanding of how the business operates.
If the last few years were spent making AI smarter, the next few years will be spent making AI more informed. The Lakehouse, Unity Catalog, and ingestion pipelines are now fundamentally context infrastructure.
If context was the headline theme of the week, governance and runtime control were a neck-and-neck second. For many organizations, the fear ungoverned agent access remains a barrier to moving autonomous agents into production.
As Ghodsi put it:
"We have to be able to control the AI. Make sure that we have security rules, policies, and auditability over the AI, the data, and the infrastructure."
This is why governance continues to center on Unity Catalog. Every announcement – Genie Ontology, the AI Gateway, Sandbox, App Spaces – is grounded in the same permissions and auditability model that has always governed data in the lakehouse. What changed at this summit is that Unity Catalog’s governance now extends natively to agents, not just tables and dashboards.
A key mechanism enabling this is the Unity AI Gateway, which sits between agents and enterprise systems. Instead of granting broad or persistent access, it issues scoped, auditable tokens governed by Unity Catalog policies. It also routes requests in real time, shifting workloads between models based on cost, capability, and policy constraints.
As agent usage scales, cost becomes a limiting factor as much as capability.
The Unity AI Gateway enforces budget controls at the level of users, teams, and applications, preventing runaway agent loops from generating uncontrolled spend. It also applies contextual policies to ensure agents do not violate access rules or expose sensitive data during execution.
A key evolution this year was the move from simple cost limits to cost visibility. Rather than discovering a runaway agent after the invoice arrives, teams can now attribute cost back to the specific user, application, or workflow that generated it. This turns AI spend from a black box into something a finance or platform team can forecast and manage the way they would any other line of infrastructure cost.
Another clear theme was avoiding lock-in as AI systems become more embedded in core enterprise workflows.
As Ghodsi noted:
"Over the last 20, 30 years, the main thing that's slowing down organizations is the fact that they're getting locked into various types of software that they keep adding to their complicated stacks."
This shows up in Lakebase, a serverless Postgres layer designed for operational workloads. Its branching capabilities allow teams to spin up isolated environments for testing, evaluation, and agent development without duplicating production data. Because it is built on Postgres, it remains portable rather than proprietary.
Lakebase also reduces the need for traditional change data capture pipelines. By making operational data immediately available for analytics and agent reasoning, it shortens the distance between live systems and decision-making.
With open-source tools like Delta Lake and Iceberg and open-source frameworks like Omnigent, Databricks is committed to the idea that enterprises will choose flexibility over convenience. The AI landscape is moving too fast for anyone to lock their core data context into a single proprietary model or vendor ecosystem.
Agentic AI dominated the expo floor, but the technical tracks focused on what it takes to make multi-agent systems reliable, secure, and cost-efficient in production. Jonathan Frankle, Databricks’ Chief AI Scientist, noted that the industry is moving away from raw token usage as a vanity metric and toward efficiency and outcome-based measurement.
To support this shift, Databricks introduced Omnigent, an open-source orchestration framework for multi-agent systems. Instead of relying on a single model to handle complex problem spaces, Omnigent coordinates specialized tools and routes tasks to the appropriate system. It then applies evaluation layers to validate outputs before they reach production, improving both reliability and efficiency.
Beneath this orchestration layer sits the Unity AI Gateway, now positioned as the runtime control plane for the entire AI estate. It manages routing, budgets, tracing, and security policies across all agents and models, regardless of origin.
Databricks also extended this system into application delivery through Databricks Apps and App Spaces, which introduce serverless micro-apps as a lightweight alternative to traditional dashboards. With the Genie App Builder, non-technical teams can assemble and deploy agent-backed applications without provisioning infrastructure. Instead of a small number of dashboards, organizations can now support a large number of narrowly scoped internal apps, each governed and isolated through App Spaces.
Finally, Sandbox provides a controlled execution environment for experimentation. Agents can run code, test workflows, and interact with data using shallow clones, reducing the risk of production impact while enabling faster iteration.
Customer sessions reflected how organizations are rethinking about data in practice. A quote from Mukesh Ambani, Chairman & Managing Director at Reliance Industries Limited, captured the direction:
"For many years, enterprises used data mainly to look back. The next generation will use data to make better decisions, act faster and continuously improve. This is the shift from dashboards to decisions."
This was not framed as aspiration, but as something already underway. For decades, enterprise data systems focused on static dashboards and retrospective reporting. Agentic systems change that by embedding data directly into operational decision-making.
We saw this play out in several high-profile enterprise use cases:
Databricks reflected the same shift internally. CustomerLake replaces batch-driven marketing workflows with continuous decision loops where identity resolution and personalization update dynamically in response to behavior. Lakewatch applies the same principle to security operations, automating threat detection and triage so analysts operate from guided actions rather than static dashboards.
The biggest challenge in enterprise AI right now is whether models can be trusted to run inside real systems without breaking cost structures, governance rules, or business logic along the way.
Across Databricks’ announcements, it was clear that context, governance, cost control, and openness are being unified in the lakehouse as a single operating model for enterprise AI. That convergence is what separates experimentation from production.
At Concord, we see this transition up close with teams trying to move AI systems from prototypes into something that can carry production workloads. We connect modern data platforms like Databricks with the operational layers needed to make AI usable in real environments, helping organizations move from isolated prototypes to governed systems that can scale safely.
If you’re working through how to bring that structure into your AI roadmap, or where your current systems are still missing the foundation needed to scale, we’re always open to a conversation.
Not sure on your next step? We'd love to hear about your business challenges. No pitch. No strings attached.
