Gartner estimates that 80% of organizations increased AI investments in 2026, yet only one in five shows measurable ROI. Despite that, the belief in AI to move the needle for enterprise organizations is still very much alive, as evidenced by the more than 20,000 data and AI leaders who gathered in San Francisco this week for Snowflake Summit 2026. There, the companies that have cracked the code on production AI shared what they actually did differently.
Here are the five things that defined the week.
Key Takeaways from Snowflake Summit 2026
Permalink to “Key Takeaways from Snowflake Summit 2026”-
The model is commoditized, so your data is the price of entry.
-
But your data is only as good as the context you have on it.
-
MCP is becoming enterprise infrastructure.
-
Governance is an accelerant, not a brake.
-
Context needs a lifecycle, and it has to stay yours.
The era of AI experimentation is over
Permalink to “The era of AI experimentation is over”Sridhar Ramaswamy, CEO of Snowflake, opened the summit with a line in his keynote that set the tone for the week: “AI isn’t a promise anymore. It’s creating real outcomes, real tangible opportunities.”
Then came the evidence: Sridhar described how Canva compressed weeks of analyst work on user behavior and feature impact into near real-time product decisions. And a European utility company cut query time from weeks to two seconds, reduced compute costs by 85%, and completed a migration in 12 weeks that previously took months.
Emmanuel Frenehard, Chief Digital Officer of Sanofi, explained how his team deployed an AI concierge that gives pharmaceutical sales reps live pre-call plans, patient context, and icebreakers, all pulled from production systems in real time. And Samsung’s Jung Suh, Head of Digital Commerce Team and Corporate Executive Vice President at Samsung Electronics, told the crowd about how hours of launch-day analytics now take seconds, with 1,000 executives and marketers globally acting on the same agent.
None of these wins came from a better model. The results were possible because the companies’ AI was connected to their own unique data. Manish Sharma, Chief Strategy & Services Officer at Accenture, put it simply: “85% of our clients have a data problem. AI is not the challenge.”
For production AI, data beats models
Permalink to “For production AI, data beats models”The most repeated argument across Summit sessions was the one Sridhar made first: “The model is not your unique advantage, because your competitor has that model too.”
Every company has access to the same models, so AI on its own isn’t a source of differentiation. What your competitors don’t have is your company’s data, like customer history, operational context, and institutional knowledge. Getting that information closer to your AI reduces friction, accelerates workflows, and differentiates AI beyond the models. As Christian Kleinerman, EVP of Product at Snowflake, said, “we are in the friction elimination business.”
Snowflake’s product roadmap reiterates that. CoCo, its AI coding agent (formerly Cortex Code), cuts data migrations from six months to six days. CoWork, the personal AI work agent for knowledge workers, promises to learn how you work and run in the background across tools you already use. And a new interactive query compiler delivers 40x faster compile times on early workloads.
But context is the competitive moat
Permalink to “But context is the competitive moat”Data may be a differentiator in the AI race, but the argument goes a step further. Because AI might have access to your data, but that doesn’t guarantee agents understand what it means. That’s where context becomes the real moat.
In her session on the Enterprise Context Layer, Prukalpa Sankar, Co-founder and Co-CEO at Atlan, shared that despite years of infrastructure investment, 56% of CEOs report zero financial benefit from AI. Without an understanding of what data means or how the business operates, there’s no volume of data or level of intelligence that will make AI work in production.
“Over the last decade, intelligence has compounded thousands-fold,” she said, pointing to how much better LLMs have gotten at acing exams. “Context, on the other hand, is very old. It’s still locked in dashboards, Slack chats, and the heads of analysts.”
The fix wasn’t a better model or a new platform. It was a portable, vendor-agnostic context layer that could route semantic definitions to any agent across any system: “Without the context layer, none of this works at all.”
The industry is leaning the same way. The Open Semantic Interchange (OSI) effort showcased at Summit pushes toward vendor-neutral semantic models that travel across tools, rather than getting locked into one semantic layer. Interchange like that is a critical feature of a context layer, however, not a substitute for one.
MCP is becoming enterprise infrastructure
Permalink to “MCP is becoming enterprise infrastructure”There was no shortage of MCP headlines at Summit, and that tracks with a trend we’re seeing at Atlan. Since September 2025, our own data has shown a 58x increase in agents using MCP as their primary connectivity protocol. As MCP makes its way into the core enterprise infrastructure, every serious AI platform either already supports it or will soon.
Where connectivity used to be the differentiator for MCP, now it’s table stakes. As Joe DosSantos, VP of Enterprise Data & Analytics at Workday, put it: “MCP makes it easy to interoperate, but doesn’t make any of the agents you’re calling any smarter.”
Snowflake’s intent to acquire Natoma aims to bring MCP connectivity into the Snowflake AI Data Cloud, so when agents connect to Gmail, Slack, Jira, GitHub, or Microsoft 365, every interaction is governed by Snowflake’s access controls, with human-in-the-loop for sensitive operations. “This expands Snowflake’s governance perimeter from data to AI action and workflows across the enterprise,” Christian Kleinerman explained.
Atlan’s MCP server takes the same bet and applies it to context. With a single MCP call, Atlan gives AI agents governed access to enterprise context like definitions, lineage, permissions, semantic models, regardless of which agent framework or data platform they’re running on. The value is no longer just the connectivity, but also the enforcement that makes agents smarter.
Governance is an accelerant, not a brake
Permalink to “Governance is an accelerant, not a brake”The pace at which AI is moving has raised an uncomfortable but important debate about how to move fast without sacrificing safety.
Daniela Amodei, President and Co-Founder of Anthropic, reframed the safety-versus-speed debate by looking at how you can have both: “Trust is an accelerant. Trust is something that helps you go faster.” At the end of the day, she said, no leader wants an untrustworthy model: “I’ve never had a customer meeting where the CEO said to me, ‘I would love if Claude could hallucinate more.’”
Caitlin Halferty, Chief Data & Analytics Officer at Thomson Reuters, spoke from the other side of the table. Her team serves lawyers, tax professionals, and auditors — people whose decisions carry high stakes for professional liability. Today, more than a million professionals use Thomson Reuters’ AI legal assistant daily, built on Snowflake’s governed data foundation.
“Governance has enabled us to derisk and accelerate our AI transformation — both internally and externally,” she stated. “We’ve moved beyond pilot to production. This is finance-validated metrics embedded in our key workflows.”
On the product side, Snowflake announced agent identity (so you can tell when activity is happening in an agentic context and apply different policies), data movement policies, multi-party approvals for sensitive operations, and AI guardrails for CoCo and CoWork. The consistent message: governance should be embedded at the agent layer from day one, not left to figure out later.
That covers the data and the agents. Atlan applies the same principle one layer up, governing the context lifecycle itself. This ensures the definitions feeding those agents stay versioned, trusted, and auditable as the business changes.
Context is the new enterprise advantage
Permalink to “Context is the new enterprise advantage”This year’s Snowflake Summit confirmed that the data foundation question is largely solved, and the model question is largely commoditized. Now, the defining enterprise variable is context.
The organizations winning right now are those that have started treating context the same way they treated data: as governed infrastructure, not tribal knowledge. “Context is your IP going forward,” Prukalpa asserted. “You need to be careful where you put it.”
Jonathan Chen, Director of Enterprise AI/ML Solutions at Workday, emphasized the importance of trust and self-awareness. He described how Workday’s production agent, Midas, was designed to refuse questions it couldn’t answer with confidence. When asked about “Christmas trees,” it said it had no relevant data rather than guessing. That trustworthy, transparent behavior is what earns adoption: “Knowing what you don’t know is actually a feature, not a bug.”
But perhaps the most universal mindset shift came from Cindi Howson, Chief Data & AI Strategy Officer at ThoughtSpot. “We are not talking about your BI semantic layers that were really just metrics,” she said. “An agentic semantic layer has to include the ontology, the context, the knowledge graph.”
The next unsolved problem: context needs a development lifecycle
Permalink to “The next unsolved problem: context needs a development lifecycle”If data was last year’s problem and connectivity is now solved by MCP, the question that kept surfacing at Summit was what happens to context after you define it. A semantic model isn’t a one-time artifact. Definitions drift, the business changes, and an agent that was accurate in spring quietly degrades by summer. The teams furthest along weren’t just building context — they were keeping it true over time.
That’s a different discipline from defining context once. It means treating context the way mature engineering teams treat software: versioned, governed, and continuously improved as the business changes, with feedback and memory so the system gets better with use instead of starting cold every time. Most organizations haven’t operationalized this yet. It’s the part of the stack still wide open.
This is where Atlan is focused: not just storing context, but owning its lifecycle — so definitions stay governed, every agent interaction makes the next answer a little better, and your context compounds as an asset instead of decaying. Connectivity gets an agent to your data. A managed context lifecycle is what keeps it trustworthy six months later — and it’s the clearest test to hold any vendor against, including us.
The companies that walked out of Summit with a clear next step were the ones that stopped asking whether context mattered and started asking how to build and maintain it at scale. That question — who owns context, how it gets built, and how feedback and memory keep it current as the business changes — is what the next 18 months will be decided on.
Atlan and Snowflake, better together
Permalink to “Atlan and Snowflake, better together”Snowflake’s roadmap points at the same diagnosis, and that’s worth pausing on. Horizon Context, its new managed layer for enriching agents with governed context, is Snowflake investing in a category Atlan has been compounding for nearly a decade. When a platform of Snowflake’s scale builds toward the same conclusion, the context layer stops being a question and starts becoming standard.
As it becomes standard, the question shifts from whether you need a context layer to how much one has to carry. Horizon Context is Snowflake’s governed semantic foundation, now reaching beyond Snowflake through its preview of new connectors. It turns that metadata into governed business definitions and makes them available to the agents and BI tools your teams already use.
At enterprise scale, though, the bar keeps rising. Getting one agent a good answer is one thing. Keeping every agent across the business working together from the same trusted context as the business evolves is another challenge entirely.
That broader job is what an enterprise context layer is for, and it’s the work Atlan has focused on for years. Atlan uses AI agents to build context that covers how the work actually gets done and what each agent is allowed to do, not only what the metrics mean, and governs it as the business changes. It delivers that context to every agent across the business and feeds the governed definitions into Snowflake as Snowflake Semantic Views for CoCo and CoWork, auto-generating work that would otherwise be manual.
That speed is the point of the partnership, and it keeps your most valuable business IP — your context — open and interoperable.
