The Agentic Enterprise Needs a Supply Side

A POV on Sridhar Ramaswamy's Snowflake Summit 2026 keynote, and the question I think it leaves open.

After listening to Sridhar Ramaswamy's keynote at Snowflake Summit 2026, I came away convinced the destination he laid out is the right one. I also came away thinking the announcement leaves a real question unanswered, and it's the one I'd most want to talk about with any CDAO who watched it.

Sridhar, Snowflake's CEO, opened with a confident case for what Snowflake is calling "the Agentic Enterprise." Four components: enterprise data, AI models, the applications businesses already run on, and an agentic control plane that coordinates across them. The flagship announcements were Snowflake Intelligence (a personal "talk to your data" agent for every knowledge worker) and Cortex Code, which Snowflake affectionately calls "CoCo" (the natural-language coding agent for the builders behind the experience). Julie Sweet, Chair and CEO of Accenture, joined to talk about the data foundations every client needs to get there. Manish Sharma, Chief Strategy and Services Officer at Accenture, walked through how the partnership ties AI investment to specific outcomes in procurement, finance, and operations. Emmanuel Frenehard, Executive Vice President and Chief Digital Officer at Sanofi, demoed Concierge, a pharma rep's AI assistant pulling a pre-call plan against five years of unified data on Snowflake. Daniela Amodei, President and co-founder of Anthropic, closed with the case that trust is what lets enterprises move faster, not what slows them down.

The destination is hard to argue with. Every business user will work alongside an AI agent. Every engineer will move faster. The gap between a business question and a working answer will keep closing.

What's open, in my view, is the work in between.

What Snowflake Announced, in One Paragraph

Snowflake Intelligence puts a personalized work agent in front of every employee. Sridhar described it as the agent that pulls forecasts for a sales leader, summarizes meetings, drafts campaign personalization for a marketer, and manages a candidate pipeline for an HR lead. It sits on top of Snowflake's MCP-powered governance perimeter, which now extends natively to Drive, Gmail, Zoom, Jira, Slack, GitHub, and Microsoft 365. Cortex Code, "CoCo," is the natural-language coding agent for the builders behind that experience. The pitch is specific: a migration that took six months now takes six days, pipelines are written conversationally, and CoCo understands the team's tables, roles, and security boundaries well enough to test, deploy, and optimize. The agentic control plane coordinates across data, models, and applications, with governance built in from day one. Many of the reasons a CDAO would have hesitated three years ago are addressed.

The Asymmetry I Keep Coming Back To

There is a quiet asymmetry in this picture: Snowflake Intelligence significantly scales the demand side of data, leaving the supply side untouched. Every business user, equipped with an agentic AI assistant that asks better and more frequent questions, generates more requests against the data layer than that same employee did last year. The agent for a sales leader pulls forecasts; the agent for a marketer drafts a campaign; the agent for an HR lead manages a pipeline. Each of those interactions either lands against a data product that already exists or generates a request to build one. The agents ask well, and they ask constantly.

CoCo scales the productivity of the engineers who build those data products. It does this well, and I want to be specific about that. Sridhar's six-months-to-six-days claim is credible for the kind of work it describes, and an AI that understands tables, roles, and security boundaries is a meaningful step forward. CoCo makes the engineer who is sitting at the keyboard faster, and that is a real gain.

Here's where I'd push back, or at least pause. The asymmetry sits between those two things. CoCo changes how quickly the engineer in the chair can ship the next pipeline. It does not change how many pipelines can be in flight at once, who is supervising them, what happens when an engineer leaves, or how the team meets a wave of new requests from agents that did not exist last quarter. Sridhar's line is worth sitting with: "The thing that used to require a specialist in your team now just requires an idea." My read is that an idea still has to be typed into CoCo by someone who knows what good looks like, what tests apply, what governance the work has to satisfy. That someone is a data engineer. The data engineering team is not growing at the rate the demand side is.

I want to be clear that I don't see this as a criticism of the announcement. It's a description of what I think the announcement leaves open. Snowflake has built an excellent demand-side story and an honest acceleration story for the engineers behind the curtain. The supply-side scaling answer is something I think the customer still has to assemble.

What the Supply Side Actually Has to Do

The work between a business question and a production-grade data product is not typing. It is design, build, test, deployment, monitoring, governance, and the institutional memory that makes the next pipeline cheaper than the last. A coding assistant compresses the typing step inside that loop. The rest of the loop still requires engineers.

This is why I think most prior attempts at data democratization reached a ceiling. Self-service BI gave business users a consumption layer; it never let them change the data layer (Emmanuel made the same point from the stage when he described five years of investment producing "thousands of Power BI dashboards" before the AI work began at Sanofi). Low-code platforms made the work easier; the work still needed a person. dbt lowered the bar on building pipelines; the bar still said "engineer." CoCo is the latest, and best, version of that idea. It is also, I'd argue, the same shape of solution: one engineer, accelerated.

What a CDAO actually needs, when every business user has a Snowflake Intelligence agent in front of them, is a way to scale the engineering side of the house without scaling the engineering side of the headcount. They need the engineering team to deliver a multiple of what they delivered last year, without losing the controls that let them sleep at night.

What Scales

The shape of a real supply-side answer is a team of specialized data engineering agents that does the engineering work end-to-end, under engineering oversight. This is the domain of autonomous data engineering: CoCo, by design, is a developer tool. It lives in the terminal, and it's built for an engineer who already knows what good code looks like and what the team's standards say it has to satisfy. That isn't the audience Snowflake Intelligence is meant to serve. The business user with a personal "talk to your data" agent in front of them has no place in CoCo's surface, and I'd argue that's the right call. CoCo is a tool for the people behind the data layer, not the people in front of it.

A team of data engineering agents is a different category of tool, and it sits in a different place. What struck me listening to Sridhar describe Snowflake Intelligence is that the natural completion of that vision is for every business user to have not just an agent that asks, but a team behind the asking, one that can take a business requirement and translate it the whole way down to the technical implementation. Not just the SQL. The full loop: design, build, test, deploy, monitor, and govern a data product against the team's standards, end to end. This is where multi-agent systems earn their keep. The engineering team is still in charge: they set the standards, and they supervise through Mission Control. The engineer's role moves from "produce" to "approve, supervise, redirect," and the team stops being the bottleneck every new request has to queue behind.

This is what we're building Maia to do. The team-of-agents architecture handles the design, build, test, deployment, and monitoring work. The Context Engine encodes the team's standards once, so every agent reads them on every pipeline — an applied form of context engineering. Mission Control gives the engineering lead the human-in-the-loop supervisory surface (a board of work in progress, with the ability to inspect, redirect, or roll back). CoCo accelerates one engineer typing. Maia becomes a member of the engineering team, running the work while the team manages execution. From the business user's perspective, it's the difference between an agent that pulls forecasts from data that already exists and a data engineering team standing behind that agent, able to build the next pipeline when the data doesn't.

These are different shapes of solution for different problems. CoCo answers a real question: how do I make the engineer at the keyboard faster? Maia answers a different one: how do I give every business user with an agent in front of them a data engineering team behind them, capable of taking a business requirement to a production data product, end to end?

The Roadmap Implication

A CDAO doing AI strategy in the wake of Summit has a real planning question to answer. I think the agentic enterprise vision is the right destination, and Snowflake Intelligence and CoCo are real and useful pieces of the picture. The piece that is not announced is the one that lets the data engineering team meet the demand wave the rest of the vision creates. For an executive audience working through this, our CDAO's guide to data automation covers the same ground in depth.

Without this critical piece, the rollout I'd expect to see looks like this: business users get agents, agents generate requests, the requests pile up on a data engineering team that is now typing faster but still the same size, the backlog grows, and the executive sponsor who championed the agentic enterprise vision answers the question "why isn't this delivering yet?" with "we are building the pipelines as quickly as we can."

My honest framing for the next twelve months is that the agentic enterprise creates a new bottleneck. Not "business users can't ask good questions of their data." That one is gone. The new bottleneck is whether the team behind the data can keep up with the asking. Solving it requires a different category of tool than a coding assistant. It requires a team behind every business user with an agent.