Denise Persson runs marketing for Snowflake. That’s a 700-person org, new-business pipeline she’s personally accountable for, and a level of compliance and data risk most of us never have to think about. She came back to SaaStr AI 2026 to talk about what actually changes when you deploy agents across a marketing team at that scale.
The headline she gave us: she doesn’t log into a dashboard in the morning anymore. She interrogates her data in plain English. Nobody on her team gets Slack messages from her asking “why did pipeline move in US West?” because she just asks the data directly.
The Top 5 Takeaways
1. The dashboard is dead, or at least dying. Dashboards only ever answered “what happened.” They never answered “why.” So you’d ping someone, schedule a meeting, sit with the sales team and argue about what the numbers meant. Persson now asks her data the why directly and gets recommendations back in real time. Her quote: nobody gets Slack messages from her anymore, because she can finally get the answers she could never get before.
2. Talking to your data killed the sales-marketing data war. Every B2B leader has lived this. Marketing says the campaign worked. Sales says it didn’t source revenue or “doesn’t count.” You burn hours aligning on whose dashboard is right before you ever discuss the actual business. One source of truth ends that. The data now tells you where a deal was sourced, who touched it, what happened on the site. The fight over interpretation goes away, and so does the time you spent on it.
3. Better data work isn’t optional, it’s the whole game. Bad data plus AI doesn’t give you bad decisions. It gives you bad decisions faster and at scale, because the agent amplifies whatever you feed it. Persson’s advice to anyone starting out: invest in your data estate first. Skip it and it bites you a year from now. It’s the Salesforce hygiene lesson from 15 years ago, except the cost of getting it wrong compounds far faster.
4. The budget reality: deliver 40-50% growth with flat or fewer resources. That’s the actual mandate. Nobody is walking into next year’s planning asking for more headcount. Persson was blunt: if you ask for more bodies in 2026, leadership will look at you like you don’t understand where the company is. The expectation now is that AI absorbs the growth, not new hires.
5. The hiring profile flipped from tools to temperament. The old job spec was a list of certifications: Marketo, Salesforce, the platforms. Now the soft skills matter more than the stack. Adaptability, curiosity, self-leadership, change management, the willingness to learn at the speed things are moving. The GTM engineer is the role Snowflake hires for. Business analysts, much less so.
A 30% Reduction in Cost Per Opportunity
Persson didn’t just talk philosophy. The proof point she led with: a 30% reduction in cost per opportunity over six months, driven by pulling fragmented media channels into one place and letting the system recommend daily optimizations instead of waiting until a campaign ended to learn it failed.
The morning brief is the other unlock. She gets a daily skill report that goes well past pipeline. Org health. Who joined Snowflake marketing this week, who left, whether there’s an attrition issue forming. Even travel and expenses she’d rather not look at manually now surface on their own. Intelligence that used to live only with finance is now a question she asks before her first meeting.
How They Built AI Fluency Across 700 People
This is the part most teams underestimate. Persson called it the single biggest investment of the last year, and she runs it as inspiration, not mandate. Her words: she doesn’t believe in the stick.
The system:
- Weekly AI skills training for the team
- A weekly AI challenge where someone records a short video on an agent or skill they built, and challenges someone else to share next
- Function-level AI hackathons, because what the comms team needs differs from what digital marketing needs
- An AI council and a quarterly company-wide AI day
- A usage leaderboard, with a heavy caveat she repeats every month (more on that below)
- “What matters,” their quarterly OKRs, where every single person has to set an AI goal. It can be small. It can be learning one thing. The point is everyone moves.
The result that surprised her most: the top of the leaderboard isn’t the people you’d predict. Her top three power users came off the brand team. They didn’t stay siloed either. They’re the ones now running into other functions to help with hackathons. The innovation showed up where she least expected it.
The Governance Layer is Managed by a Centralized AI Engineering Team
At Snowflake’s scale and risk tolerance, you can’t just let a thousand agents bloom unchecked. A wrong email to a customer is a brand impression that lasts. So they built a control plane.
A centralized AI engineering team sits on top of everything. Any skill that’s going to be used by more than a few people has to be certified before it ships. Their company-wide GTM agent, Raven, is used across both sales and marketing, and every skill inside it is centrally certified. The dual job of that team: make sure agents behave correctly, and stop the company from building the same agent five times.
On cost, Snowflake made a deliberate call: AI spend sits at the company level, and marketing gets effectively unlimited access right now. The CEO didn’t want anyone’s departmental budget to throttle experimentation. Persson was honest that this is a 2026 decision that probably changes, because usage is going through the roof and the bill is real.
Where the Human Still Wins
Persson’s read on the human-versus-agent line: authenticity is becoming high value precisely because so much is now synthetic. People are getting skeptical about what’s real. A dancing-dog video, fine, nobody cares it’s fake. But trust in a brand is different. That’s where humans spend their time now, on the uniqueness and authenticity of the brand, the stuff agents can’t manufacture.
Two more shifts worth stealing:
Events are surging. Ten years ago everyone declared events dead and pivoted all-digital. Now the demand for in-person experiences is, in her words, going off the roof. People are craving the room.
Enablement is getting rebuilt. Snowflake moved sales enablement, partner enablement, and customer training under marketing, because content was being duplicated across the company. The new model: build content once, generate every derivative asset for every segment, and ship self-service enablement agents so sellers get training at the moment they need it instead of sitting through a session that’s either too basic or too advanced. They’re even using roleplay agents so reps can practice a pitch against an agent loaded with company intelligence instead of cornering their manager.
The 3 Mistakes Denise Made (And the Ones She Sees Everywhere)
Even at Snowflake, the playbook isn’t clean. Here’s where she’s tripped, by her own admission and from reading between the lines.
1. The token leaderboard measured the wrong thing. A leaderboard ranked on usage rewards activity, not outcomes. An audience member called out the tension directly: more tokens means more cost, not necessarily more results. Persson now caveats the leaderboard every single month, telling the team it doesn’t matter if you only used 100 tokens, what matters is the business outcome. If you have to verbally correct your own metric every time you show it, the metric is sending the wrong signal. Build the leaderboard around outcomes from the start, not consumption.
2. “Let everyone build everything” is creating sprawl they’ll have to rein in. Persson admitted it plainly: they’re encouraging building at every level right now, and it’s going to come to a point where they have to pull it back. Duplicate agents are already being built across the company. She drew the exact parallel herself, to the SaaS app explosion of 15 years ago, when marketing bought a hundred tools and IT eventually had to come in and impose order. They know the control layer is coming. The cost of waiting is the cleanup.
3. Unlimited AI spend was the right call for experimentation and the wrong call for cost discipline. Centralizing AI budget and removing limits got people leaning in, which was the goal. But she conceded usage is going off the roof, the spend is significant, and they’re already spotting agents across the company doing the same job twice. She expects to walk this back in 2026. The lesson: unlimited access buys you adoption speed and a bill you eventually have to reckon with.
4. The activation layer is still half-built. This one she sees as the current gap, not a past error. They automated the analysis side: which use case to promote to which account, a workflow that used to eat enormous time. What they haven’t cracked is full activation. The campaign still can’t fully launch itself. That’s why the GTM engineer role exists and why that team’s time is the most demand-constrained resource in the building. The analysis got cheap. The doing didn’t, yet.
Persson’s closing point on the future of the function: nobody can paint a clear picture of what marketing looks like in three years. But you can be part of shaping it, or you can opt out. That’s the choice she’s putting in front of her team, and it’s the right frame for the rest of us too.
Have a question for Dear SaaStr? Submit it at saastr.ai/ai-mentor.
