What four days at the Databricks Data + AI Summit taught me about making agentic analytics actually work
I spent this week at DAIS 2026 in San Francisco — four days of sprinting between buildings, conference rooms, and keynote halls to catch as many sessions as possible. I tried my best to get a full picture, not just the headline announcements.
Sessions included Intelligent Document Parsing with Lakeflow, AI Mesh: The Age of AI Products, Your AI Strategy is Only as Good as Your People Strategy, How J&J Medtech and Takeda Scale Data & AI, An Intro to Building & Scaling Agentic Apps, Agentic Analytics on the Databricks Lakehouse, and Latest Innovations in AI/BI for Business Users.
The energy was electric. The keynote session was attended by over 30,000 professionals – I had never been in a room like that before. The demos were impressive. And one slide — almost easy to miss in the middle of a packed breakout session — said something that has stayed with me:
“Most organizations don’t have an AI innovation problem. They have a reuse, governance, and trust problem.”
This diagnosis lands with particular weight in regulated, data-intensive industries like life sciences.
The Pattern We Keep Seeing
Across industries, a common AI adoption story plays out like this: a team identifies a high-value analytics problem, builds something smart to solve it, and sees real results. Then another team does the same — independently, for their own version of essentially the same question. Then another.
Each effort is genuinely useful. Each one also operates in isolation: its own data assumptions, its own logic, its own definition of the metrics that matter. There’s no shared semantic layer — no governed, agreed-upon source of truth that all these efforts draw from. No common framework for what “on track” means, or how performance is measured, or which numbers are approved for which audiences.
The result isn’t a lack of AI capability. It’s a proliferation of disconnected outputs that are difficult to reconcile, expensive to maintain, and nearly impossible to scale across a large user base.
The next use case starts from zero. Again.
This is the AI pilot trap. And based on the sessions I attended – looks like it’s everywhere.
What “Agentic Analytics” Actually Requires
The most useful framing I took away from DAIS came from the session on Agentic Analytics on the Databricks Lakehouse. The presenters used the example of a data analyst “Jess” and laid out three failure modes with uncomfortable clarity:
Accuracy — Agents lack the semantic grounding to write accurate queries. Ask an agent “what’s our dropout rate in the APAC region this quarter?” and it will give you an answer. Whether that answer matches your function-specific definition of dropout, your certified enrollment table, or your SME-approved denominator — that’s a different question entirely.
Governance — What data and tools are the agents allowed to access? Who approved this agent’s access to a certain table? Can it write to operational dashboards? Does it behave differently depending on who’s asking? Agents access data and tools in ways that weren’t designed for them.
Scale — Agents that work in a pilot cause performance and cost problems at enterprise scale. Lightning-speed query response times and predictable infrastructure costs aren’t luxuries when you’re running analytics across hundreds of trials for thousands of users.
These are the exact failure modes that could turn a promising clinical ops AI initiative into a governance escalation.
The Architecture That Changes the Equation
What Databricks laid out — and what I think is genuinely the right pattern for enterprise analytics teams — is a four-layer answer to these three problems:
Compute (Lakehouse RT) — gives you the scalable, cost-efficient foundation. Sub-second responses across large clinical datasets without unpredictable cloud bills.
Governance (Unity AI Gateway) — acts as a single control plane: agent registry, access control, contextual policies, cost budgets, and tracing. One place to govern both data and AI. One place to ask “who approved this?” and get an actual answer.
Semantics (Genie Ontology) — THIS, for me, what the most valuable aspect of the entire conference. It is the piece that made every room lean forward. Rather than each team maintaining their own definitions, you build a shared business ontology — enrollment milestone timelines, regional performance benchmarks, patient dropout criteria — once, in a governed layer, and every agent draws from it. This can be defined based on functional definitions. An agent querying trial performance doesn’t decide what “on track” means. The ontology does.
Agent Choice — deploy through Genie One, custom agents, or MCP endpoints, wherever your users actually work: in MS Teams, in your clinical ops portal, in a mobile dashboard during a site visit.
The phrase from the keynote that stuck: one platform, one identity model, one governance plane.
The Harder Question: Who Owns This?
The most practically useful session I attended was on “AI Products” — not AI features, not AI pilots, but AI products. The framing was simple: the domain owner defines the reusable business capability before the team implements the agent.
In clinical trial operations, that means someone has to own the answer to: what is our certified definition of enrollment rate? What’s the governed source of site dropout data? Which metrics are approved for executive reporting versus investigational use?
That’s not an engineering question. It’s a data leadership question. And the organizations that answer it first — that build once, govern once, and compose many times across their agent ecosystem — are the ones that will turn AI pilots into AI infrastructure.
The ones that don’t will keep rebuilding the same agent, for the same question, on the same data, six months from now.
Dilip Merala is a Data & AI Analytics Manager specializing in clinical trial operational analytics. He attended the Databricks Data + AI Summit 2026 in San Francisco.
