Ali Ghodsi, co-founder and CEO of Databricks, opened Databricks Data + AI Summit 2026 with a poll. He asked the 31,309 attendees from 174 countries whether AGI had arrived. About 90% said no. His response: you’re wrong — it’s already here. That moment set the tone for the entire week.
Here are the five things that defined the summit.
Key takeaways from Databricks Data + AI Summit 2026
Permalink to “Key takeaways from Databricks Data + AI Summit 2026”- Ready or not, AGI is here
- The context gap is now a business problem
- Closing the gap requires context-by-design
- Databricks is making its own move on context
- Systems of context are the new systems of record
Ready or not, AGI is here
Permalink to “Ready or not, AGI is here”Frontier models now solve the “Humanity’s Last Exam” benchmark: 2,500 hard questions spanning every discipline humans have organized knowledge into. By every definition used in 2009, AGI is here. The audience didn’t recognize that evidence because they’re still dealing with a fundamental question: if intelligence is solved, why is enterprise AI still failing?
It doesn’t come down to the models or the benchmarks, or even which frontier lab is winning. It comes down to context. And that’s the factor that most enterprises haven’t solved for yet.
The context gap is now a business problem
Permalink to “The context gap is now a business problem”As AI has become mainstream over the past two years, enterprise AI projects are becoming scrutinized more aggressively. And rightfully so: a PwC study found that 56% of CEOs report zero financial benefit from AI. The share of companies scrapping AI projects before reaching production went from 17% to 42% in a single year. Models got dramatically better: API costs dropped 99% between 2023 and 2025, and benchmark scores improved across every category. Still, enterprise results got worse.
Ali Ghodsi put it bluntly: “AI doesn’t have an intelligence problem. It has a context problem.”
The standard diagnosis has been organizational, blaming a lack of change management, wrong use cases, or too much technical complexity. The speakers at the summit offered a different diagnosis: the problem is architectural, not organizational.
Enterprise agents fail because they can’t read context. They know what words mean, but not what your business means. They know what “revenue” is but not how your company defines it, which table is authoritative, when the fiscal quarter closes, or why one metric disagrees with another one two rows down.
The intelligence is there, and it’s commoditizing. The context isn’t. And high intelligence with low context is a dangerous combination in production.
Ali called this explicitly in his opening keynote. Four problems are blocking enterprise AI: Context, control, cost, and choice. Context came first for a reason.
Closing the gap requires context-by-design
Permalink to “Closing the gap requires context-by-design”The most concrete proof of what it will take to fix the context gap came from Mastercard’s breakout session on engineering AI-ready data products with a context-by-design approach.
Mastercard processes 200+ billion transactions across 220+ markets. But Vivek Radhakrishnan, SVP of Data & Analytics, and Brian Piel, Vice President, didn’t talk about models. They talked about infrastructure. Specifically, how to build a federated data products layer where every product ships with its business context built in. That means aligned definitions, clear ownership, and embedded compliance constraints, built in from the start.
Brian put it plainly: “When we talk about enterprise-ready data for AI, we’re talking about adding the right context and making sure that everyone agrees on that.”
Mastercard used Atlan’s context agents to enrich more than 30,000 data assets, saving over 6,000 hours of human effort. Before, Vivek recalled, regulator questions would trigger “a bunch of emails, meetings, lots of follow-ups.” Now, the process is largely automated.
That’s as much a context infrastructure story as a governance one. Governance improved when the context got built.
Databricks is making its own move on context
Permalink to “Databricks is making its own move on context”Today’s agents do what’s called a “random walk”: at query time, they search across documents, tools, and systems to find relevant information. This approach is slow, expensive, and degrades answer quality because the agent is reasoning about what to look for while it’s simultaneously trying to answer your question. Several platforms at the summit offered approaches to solve this problem with context, including Databricks with its new Genie Ontology context store.
Genie Ontology builds a background graph of enterprise context across the lakehouse and external systems, including Google Drive, SharePoint, email, and calendar. That gives agents structured context that’s ready to go before users ask questions. An OntoRank algorithm (think PageRank for enterprise assets) identifies the most relevant knowledge by analyzing who created it, how often it’s accessed, and how it connects to other business concepts.
Databricks’ own instance has 4.5 million ontology snippets. Users can inspect them in the product, trace them back to underlying assets, and verify where Genie learned a fact. Ali Ghodsi called out Atlan by name in his opening keynote, saying that Genie customers “can bring your own semantics from Unity Catalog, whether you’re using one of our partners like Atlan.”
Genie Ontology isn’t a replacement for the context work enterprises need to do. It’s infrastructure that makes that work more valuable, but only if the context you’ve built is clean, governed, and connected. Organizations that have invested in their context layer will get dramatically more from Genie than those that haven’t.
Systems of context are the new systems of record
Permalink to “Systems of context are the new systems of record”Satya Nadella’s appearance at the conference, on the heels of his viral X article, asserted the clearest strategic shift of the week: “In the past we said, here is the system of record, system of engagement. Now we are creating systems of context.”
The shift redefines what enterprise infrastructure is for: not storing and retrieving data, but making intelligence applicable to a specific organization.
Databricks made the same bet explicit across every major announcement. The vision for Unity Catalog and the agent platform is to become the “agent system of record,” replacing fragmented SaaS stacks as the context and governance layer for all enterprise agents.
“Intelligence, for all intents and purposes, is commoditized,” said my colleague Austin Kronz during his session. “Your competitors have access to the same models you have. The thing that actually makes your agent more differentiated is your context and your data. Context is almost like the IP of your business.”
The practical consequence is that every architectural decision a data team makes today — how data products are structured, how metadata is managed, how definitions are governed — is now also a decision about AI readiness.
How to start harnessing and building context
Permalink to “How to start harnessing and building context”If you’re still mapping out what this shift looks like in your organization, here’s where to start:
1. Audit your definition layer. What do terms like “active customer,” “revenue,” or “escalated incident” mean in your company? If the answer varies by team, that’s a context problem. Before you introduce any tools, take inventory of and align your definitions.
2. Build context into your data products by design. Take a page out of Mastercard’s book by embedding definitions, compliance, and ownership into data products from the start. That saves time and avoids errors in a way that compounds.
3. Connect your context to your agents via MCP. Building context that’s accessible through multiple interfaces ensures it stays fresh and consistent, no matter where you’re working. Companies like Mastercard run Atlan’s MCP integration to deliver governed, enriched metadata to their internal agentic workflows, and Databricks Unity Catalog now covers MCP tools as a native asset type.
4. Treat context like code. Assign stewards, build review cadences, and version definitions so they always stay current. When an agent gives a wrong answer, you’ll be able to trace it to exactly what context it was running on and fix it at the source.
The context race is happening now
Permalink to “The context race is happening now”Ali’s poll at the opening of Databricks Summit was more revealing than the audience realized. The 90% who said AGI hadn’t arrived weren’t wrong about model intelligence. They didn’t realize that the model race is over, and the context race has begun.
Vivek Radhakrishnan of Mastercard put it more simply than any analyst has: “We’re not adding hundreds and thousands of people to help us with governance. We’re leveraging technology where it’s possible. And the good thing is: we’re already realizing this.”
Mastercard is realizing and acting on it. Atlan and Databricks are building infrastructure for it. The question isn’t whether your AI will eventually need a context layer, because it will. Now is the time to start building it.
