Holly Chen has spent her career building growth systems at Google, Slack, Loom, and now Samsara, where she leads growth marketing inside a much larger marketing organization. Samsara has about 4,000 employees, a 230-person marketing team, and roughly 35 people in growth marketing.
This conversation shows what AI adoption looks like after the first wave of curiosity. Samsara is not debating whether marketers should use AI. The harder work is turning hundreds of individual experiments into an operating system: who builds, who reviews, how tools get shared, how agents reach production, and how they’ve shipped 90 AI agents.
AI fluency now means systems fluency
In 2021, Holly says she was hiring specialists: paid, SEO, lifecycle, and other people who owned specific growth lanes. Today she is still hiring for expertise, collaboration, curiosity, and openness to learning, but the practical bar has shifted. In interviews, she asks candidates how they use AI and listens for the level of abstraction. Are they using chat as a thought partner? Are they building workflow agents? Or are they creating systems that other people can reuse?
That distinction matters because AI can either improve an old workflow or change the workflow itself. A personal to-do agent that reads Slack, Zoom, email, calendar, and Airtable is an efficiency layer on top of an existing routine. A sales super agent that holds company knowledge and generates personalized decks changes how product marketing enablement works. The marketer's job becomes knowing which kind of problem they are solving.
Build bottom-up with loose governance
The core operating model at Samsara is not a central AI team building everything for everyone. Holly describes two patterns: top-down mandates where leadership asks ops or engineers to build agents for the rest of the company, and bottom-up adoption where individuals build while ops or engineering consult, review, and help maintain.
That is the key lesson for enterprises. A top-down push can make AI important. It cannot understand every team's daily workflow well enough to keep improving the tool after version one. The people closest to the work have to become builders, while marketing ops creates the guardrails that keep the system usable.
At Samsara, those guardrails include a shared repo of approved skills, a knowledge base, brand guidelines, and a marketing ops layer that operates more like a consultant than an outsourced builder. The goal is not to slow down experimentation. It is to keep experiments connected to the right data, tools, and review paths once they start affecting other people.
Separate personal agents, team agents, and super agents
Holly breaks Samsara's agents into three categories. Personal productivity agents are built by individuals for their own work and do not need to be registered. A marketer can build a daily to-do agent, adjust it, borrow from a better shared skill, and keep using it without creating organizational overhead.
Team agents are different. They solve shared workflow problems and may pull from systems like Gong calls, Salesforce, calendars, or other business data. Those agents register with an AI governing body Holly calls iPod, run by marketing ops. When a team agent affects other people, an engineer reviews it to make sure it runs correctly, pulls the right data, and can go to production.
Functional super agents are the third tier. These are used across the organization and are usually built by marketing ops or revenue ops. Holly gives Samsara GPT as an example of a broader agent used by the entire org. The taxonomy gives the company a way to avoid treating every experiment like a production system, while still giving production systems a path to review.
Adoption needs rituals, not just tools
Samsara's transformation went through three stages. First came leadership mandate and cascade. The CEO and executive team made AI a business priority, then VPs, directors, and managers kept reinforcing the message. Second came learning and enablement: a marketing ops AI boot camp, org-wide tool access, and hackathons. Third came process and structure, including AI Power Hour and recurring org meeting slots where teams present what they built.
The hackathon format is especially practical. Instead of one intense build day, Samsara gave people four weeks before a team onsite. People self-organized into groups of roughly three, learned together, built an agent, presented to leadership, and competed for the top three agents to go into production.
The structure solves multiple adoption problems at once. It creates time to learn, a safe small group for questions, peer accountability, leadership visibility, and a production incentive. That makes AI adoption feel less like a training deck and more like practice.
Scale judgment by showing the build
Judgment is built through years of failures, learnings, repetition, and exposure. A static skill can capture some tactical know-how, but it cannot fully capture the intuition that comes from seeing hundreds of landing page tests.
That is why the sharing ritual focuses on process, not just output. In Samsara's AI Power Hour and org meetings, teams do not simply demo a finished agent. They explain how they improved it, how they iterated, what surprised them, what did not work, and what they learned.
This is one of our key takeaways from this conversation. If teams only share artifacts, people copy prompts without understanding judgment. If teams share the build process, the organization sees how better decisions are made.
The agent portfolio becomes its own systems problem
Samsara's use cases have grown from simple personal automations to more sophisticated marketing systems. Early examples included an agent that watched for new Zoom recordings in a marketing learning-share Slack channel, summarized them, and sent the summary to one person. Later examples included a Figma plugin that replaces ad copy in banner ads, a reverse-engineered Google Ads quality score workflow that rewrites ad copy and landing pages, and 1:1 ABM landing pages that pull from Salesforce, account plans, engagement data, and external intent signals.
At that point, the challenge shifts. The issue is no longer whether people can build. It is whether the organization can discover, orchestrate, and maintain what has been built.
Holly says maintenance is top of mind because some people build agents and then forget to maintain them. She compares the work to code: there are versions, dependencies, and long-term ownership questions. Not all 90 agents are production systems or repeatedly used, so governance has to distinguish between creative messiness and durable infrastructure.
Every marketer has to become a builder
Near the end of the conversation, Holly names the mistake she sees from overwhelmed marketing leaders. Every marketer needs to think like a systems builder. They need enough fluency to identify the workflow, build or adapt the right agent, understand where it needs governance, and keep improving it after the first version ships.
Samsara also tracks ROI in both productivity and business impact. A weekly two-hour process becoming 10 minutes is one kind of gain. An account-intelligence agent that helps sales identify expansion or renewal opportunities is another. The common thread is that builders present how the system changes their work, not just that they used AI.
Listen to the full conversation
Listen to the full episode for Holly's breakdown of Samsara's personal, team, and functional super agents; the iPod governance model; the four-week hackathon format; AI Power Hour; and the orchestration problem that appears once a 230-person marketing team has built 90 agents.
