Last week, the Snowplow team joined thousands of data leaders, engineers, and AI practitioners at Snowflake Summit 2026 in San Francisco. At last year’s Summit, the open question was whether AI belonged in the enterprise data stack. This year nobody was asking that. The conversation had moved on to whether the stack is actually ready for what AI is starting to demand of it.

Snowflake CEO Sridhar Ramaswamy framed it directly in his opening keynote: unifying enterprise data and making it governed, secure, and AI-ready is what separates real business transformation from AI experiments that never scale. The Snowflake product announcements followed that line, positioning Snowflake as the Enterprise Data Layer, while the session our co-founder and CTO Yali Sassoon gave anchors Snowplow as the Customer Context Layer. Five things stood out.

‍

1. Enterprise AI has moved from insight to action

For years the promise of the data stack was insight: better dashboards, better models, a better understanding of what already happened. The theme running through Snowflake Summit 2026 was action. The announcements and the sessions assumed AI isn't there to tell you what customers did, but to do things on their behalf and on yours.

That shift sounds subtle. However, it changes what the data underneath has to be capable of, and it's the thread that connects everything below. We at Snowplow call this the Customer Context Layer, and that has been our bread and butter since day one (over 14 years ago).

‍

 2. Snowflake now has two agents, one for doing the work and one for building it

Snowflake opened the week by renaming Snowflake Intelligence to Snowflake CoWork, a personal work agent that helps knowledge workers reason across enterprise data, automate workflows, and take governed actions across the tools their teams already use. Alongside it, Cortex Code became Snowflake CoCo, a coding agent aimed at the data and engineering teams building the infrastructure underneath. CoCo generates pipelines, ML models, and agents from natural language, grounded in your actual Snowflake catalog, lineage, and permissions from the first prompt.

The split is the point. CoWork does the work; CoCo builds the systems that do it. Together they are Snowflake's answer to what an agentic environment looks like for the people doing the work and the people building the tools.

‍

3. Three Cortex announcements worth watching

Beyond the two agents, three Cortex announcements stood out:

Cortex Sense (private preview soon) learns how an organisation defines its own business, including its workflows and the relationships between its data assets, so business reasoning is grounded from day one.

Cortex Training (in private preview) lets teams fine-tune open-weight foundation models to their own domain and cost requirements without managing GPU capacity.

Streaming feature support (generally available soon) serves online features in 10ms from Snowflake Feature Store, with under two seconds of data freshness from ingestion to serving.

‍

Streaming feature support is the one to watch for anyone trying to serve agents in real time, for reasons that become clear in takeaway 5 below. Spoiler: the answer is Snowplow Signals.

‍

4. Agentic AI is quietly breaking customer data infrastructure

This was the core of Yali's session, "When the Interface is an Agent: Re-imagining Your Customer Context Layer." His argument was that agentic AI is creating data infrastructure problems most teams haven't reckoned with, and the window to get ahead of them is closing.

Two shifts drive it. The first is that agents now make up more than half of all website traffic, and filtering them out is no longer enough. Whether an agent is acting for your customer, your competitor, or no one useful at all changes how you should respond to it.

The second, and more interesting, is what happens when you put an agent inside your own experience. Yali used a simple example. Three customers walk into a shoe shop: one wants to buy, one wants to return, one is browsing. Today they all get the same interface and do the work of navigating to what they need. With an agent, the experience composes itself around each person's intent. The customer stops reading the map.

‍

That generates an order of magnitude more data than a traditional digital experience. It also makes two questions answerable for the first time. Yali pointed to Gary Angel, one of the early figures in digital analytics, who observed that the two things teams most wanted to know about a customer had always been opaque: what is this person trying to do, and did we give them what they needed? Tracked properly, agentic experiences finally close that gap.

The hard part is activation. A customer-facing agent has a finite context window, and dumping a full event stream into it helps no one. The real engineering problem is compressing a high-velocity behavioural stream into something dense enough to be useful without eating the budget the agent needs to reason. As Yali put it, that means turning detailed tabular data into a tight paragraph of text. It's a new problem.

‍

5. Most teams can't yet answer the three questions that matter

Three things earn a place in an agent's context window: who the customer is, what they are trying to do right now, and whether they need help. Real-time identity stitching, intent inference, and frustration detection aren't additions to that model. They are the model.

This is where the Snowflake announcements and Yali's session turn out to be one conversation rather than two. CoWork and the agentic infrastructure around it assume the data feeding those agents is accurate, real-time, and current enough to reason on. Historical warehouse data tells an agent what a customer did. It doesn't tell the agent what they're trying to do right now, whether they've hit a wall, or whether they're even human.

Snowplow Signals is built to fill that gap. It takes the behavioral event stream, processes it in real time, and delivers a tight context object an agent can act on. Paired with Snowflake's 10ms streaming features, the latency objection falls away. The infrastructure to give agents real-time customer context now exists.

Yali closed with three questions he put to the room. Can you see, in your Snowflake data, how your customer-facing agents are reasoning versus what your customers are actually doing? Can you take your behavioral data and feed it to those agents in a form they can use? Can you tell the agents visiting your applications apart from the humans? Most teams we spoke with at Summit couldn't answer yes to all three.

‍

Where the work is

The teams investing now in how they collect, structure, and deliver behavioral data are the ones whose agents will work intelligently. The rest will spend the next couple of years debugging outputs and wondering why the context window isn't doing its job. That distance is the infrastructure gap, and it's widening.

Snowplow built a blueprint showing how Snowplow and Snowflake CoWork fit together in practice. If you're working through any of this, it's a good place to start. Or get in touch directly to go deeper. 

‍