Inside Snowflake’s Agentic Enterprise and AI Bets Beyond the Summit Keynote

Every enterprise software company now sells the same dream. Describe what you want, and an AI agent does the work. No dashboards, SQL, or waiting on the data team. At its Summit 26 keynote in San Francisco, Snowflake unveiled a sharper version, renaming its two fastest-growing products and promising every employee a personal agent to act on the company’s most sensitive data.

The dream pitches beautifully from a stage. It runs much harder in production, where a confident wrong answer costs money, and a leaked number ends careers.

Snowflake executives Vivek Raghunathan, SVP of engineering, runs a 2,500-person organization and is tearing up how the company writes software; Anahita Tafvizi, chief data and AI officer, makes sure that when an agent answers a question about revenue, the number is right and the person asking is allowed to see it. One bet is speed. The other is trust. Pull either, and the whole machine seizes.

Snowflake renamed Cortex Code, its AI coding agent, to CoCo, and Snowflake Intelligence, its knowledge-work agent, to CoWork. CoCo is now the fastest-growing product in the company’s history, with more than 7,100 users and roughly 13,900 customers behind it. Numbers buy confidence, but say nothing about whether the foundation underneath can take the weight.

A Contrarian Bet: Hire Green, Ship Fast

Vivek reached for Andy Grove to describe the moment. “Let chaos reign, then rein in chaos,” he says. “We’re very much in the ‘let chaos reign’ phase of this transformation.” Roughly 70% to 80% of Snowflake’s recent engineering hires have less than two years of experience, a deliberate break from an industry still fighting over senior talent. “Our thesis is that this new world empowers builders. We believe we have enough senior scaffolding in place to coach and develop these engineers successfully,” he said.

AI-native development, he asserts, is a reskilling problem, not an upskilling one. “We’re fundamentally changing how software gets built. In that environment, not having 15 years of experience doing things one particular way can actually be an advantage.”

It is a clean argument, and a convenient one when junior engineers cost a fraction of senior ones. The whole bet rests on three words: “enough senior scaffolding.” Thin it out, and a fast org of green hires is one outage, one ungoverned query away from learning why the scaffolding was there. His proof point is Cortex Desktop, built by four engineers: one mid-level, three new graduates. “Our newest engineers understand AI-native workflows better than our most experienced engineers. Meanwhile, senior engineers bring deep domain expertise, architectural judgment, and context.” He calls it a bidirectional apprenticeship.

So what happens to the engineers who do not adapt fast enough? Vivek had a tidy frame ready. “Every transformation has pioneers, settlers, and resisters. Right now we’re identifying the pioneers.” He expects the gap to narrow over time. 

Context Beats Confidence

Anahita started her career as a physicist, chasing things that are demonstrably true. Her job now is making sure an AI agent does not confidently tell an executive something that is not.

“If I ask an analyst, ‘What was Q1 revenue?’ there is only one answer, down to the penny. It is not probabilistic,” she says. “But if you ask an AI agent and it gives you an answer very confidently, how do you know it’s right? AI agents often answer confidently, even when they’re wrong.” Two questions kept her up at night: “How do you ensure the answer is accurate? How do you know it’s well governed?”

A model wired into a database has no idea how a business defines its own fiscal year. “Fiscal year, customer acquisition cost, lifetime value, those definitions vary from organization to organization. You can’t just connect a general-purpose agent to a database and expect it to know which table, column, date range, and business definition to use. It will give you an answer confidently. But you need to know whether it picked the right source and applied the right logic.” Her team spent the past year building the semantic layer that supplies that context.

For years, Anahita noted, technical skill was an accidental security guard. “Previously, you had to know SQL or Python to access the data. Now anyone in any function can ask questions in natural language.” When everyone can ask anything, governance is the only thing between a company and a leak.

Vivek’s fearless juniors are safe to unleash only because Anahita’s foundation defines what an agent may touch. Her foundation is worth the effort only because his org moves fast enough to use it. Speed and trust are not two strategies running side by side; they are one strategy that fails the moment either slips. 

BI Dashboard Demoted not Dead 

Anahita claims the dashboard is not dead so much as demoted: it still matters for shared metrics, now paired with a conversational agent for the follow-ups. “Last year, before many of these capabilities existed, we did a BI migration into Streamlit. It took six months, and we were considered fast,” she says. “Today, that same project might take days.”

That speed has a cost few people are pricing in. A dashboard was contestable; anyone could trace the query behind a number and call it wrong. An agent hands you a finished answer and swallows the reasoning behind it, which is exactly what makes a wrong one dangerous.

In February, roughly $285 billion in software value evaporated in about 48 hours, as investors bet AI agents would gut the per-seat model that bankrolled enterprise software for two decades. The same wave moved in step with close to a quarter-million tech layoffs last year. So the industry’s favorite reassurance, ‘AI augments people’, doesn’t seem to be working.

Anahita refused both the easy reassurance and the doom. “It’s extremely difficult to predict the future of jobs,” she says. “If you add value, there will be a job for you at Snowflake and across the industry. As one layer of work gets automated, another layer becomes more important.” Her message to her team is sharper: “If you don’t keep up with the rate of change, it will be difficult to remain employable.” Honest, and the exact point where “augmentation” stops describing what is happening to people.

Then there is a hard reality: women hold roughly a quarter of AI roles, collapsing toward single digits at the top. Anahita is one of the exceptions and treats it as personal. She was often the only woman in the room in physics, and is now raising two daughters drawn to STEM. She points to progress at Snowflake, where about half of this year’s Summit speakers were women. “Visibility matters,” she says. “Showing examples of what is possible matters. Creating role models matters. But it is also about welcoming those perspectives in the room and creating an environment where each person can be at their best.”

But if agents swallow the entry-level work that used to earn a first promotion, the bottom rung thins out, first for the groups already struggling to reach it. Nobody in enterprise AI owns that problem yet, and it stays invisible until a pipeline has already run dry.

The Control Layer Verdict

Speed without trust ships a liability; trust without speed ships nothing. 

Snowflake floored the engine this week, in front of a market that just wiped out a quarter-trillion dollars betting on exactly this future. The engineering is ahead of the field. The governance is serious. But when the dashboard dies, so does the argument over the number. An agent hands you a finished answer and swallows the reasoning behind it. The real prize is sitting in the governed layer, where a handful of people now decide what the whole company is allowed to ask.