Inside Google Cloud's marketing org, the top 20% of AI adopters are also the people who've done the most training. Not the most tool-buying. Not the most prompt-engineering Twitter threads consumed. The most deliberate skill-building. Sarah Kennedy Ellis, Google Cloud's VP of Global Demand & Growth, laid this out at SaaStr AI 2026, and the implications for marketing ops teams are worth pulling apart.
Kennedy Ellis has run marketing at Marketo (acquired by Adobe for $4.75B) and led Adobe's enterprise software marketing before Google. Her core argument: Google wants to be "Customer Zero" for AI in marketing. Not the flashy demo version. The actual day-to-day org running on its own agents, feeding failures back to the product teams. Most of what she described applies whether you have 5,000 marketers or 5.
The Blocker Is Process, Not the Model
Kennedy Ellis framed it bluntly: the greatest friction in a workflow is the biggest inhibitor to adoption, well above agent quality on any given day. Teams investing in change management and training are the ones getting real productivity. Teams waiting for a better model are stuck.
This tracks with broader data. About 78% of organizations now use AI in at least one area, up from 55% in 2023. Generative AI usage sits at 71%, up from 65% last year. Access isn't the constraint anymore. The constraint is that existing processes weren't designed for agents, and nobody's done the work to redesign them.
For ops teams, that's actually good news. It means the bottleneck is squarely in your domain: workflow architecture, data quality, process documentation. The "boring middle" of data cleaning, taxonomy alignment, and traceability that experts keep flagging as the prerequisite for reliable AI outputs. Skip it, and you get faster garbage.
Train in 5 Minutes or Don't Bother
Time is the real budget constraint. Kennedy Ellis said her team genuinely had about 5 minutes a week for learning. So Google built around that reality with "AI Boost Bites": 5-to-7-minute videos, some as short as 2 minutes, each covering one specific task. Early ones were basic ("how do you create slides with Gemini"). They've since progressed to multi-agent orchestration across a campaign.
They gamified it. Internal competitions, task completion requirements, badges. Kennedy Ellis called badges "gamification from 20 years ago," and it still worked. The program started internal, got adopted fast, and Google eventually published it free on YouTube, where it's passed a million views.
The lesson for B2B ops: if your enablement program requires a 90-minute workshop that nobody attends, you don't have an adoption problem. You have a design problem. The skill half-life in this space is under 2.5 years, which means continuous micro-training isn't optional. Build the cadence around the gap your team actually has.
Scaled Content Where Quality Went Up, Not Down
The headline case was the Gemini in Chrome launch. Google needed thousands of creative assets. They cut production time by about 70%, going from weeks to days. But the 70% isn't the interesting part. Everyone quotes time savings.
The interesting part: conversion rate lifted. Volume went way up, and quality went up with it. Personalization down to the individual level at a scale that wasn't possible before, and it converted better. Kennedy Ellis's rule for where to go hard: high volume plus limited human judgment required to get a high-quality outcome. That's the zone where agents earn their keep.
If you're scaling content and quality is dropping, you're doing the volume without the judgment. That distinction matters for measurement. The leading indicator isn't output count; it's whether conversion holds or improves as volume increases.
Agents Don't Ask for a Job Description
Kennedy Ellis made what she called a "maybe unpopular" point: the sales-marketing structural line matters less than accountability. RevOps started the merge. Agents finish it, because agents start from outcomes, not org charts. Humans have spent years trying to align on shared outcomes across GTM functions. Agents begin there by default.
Her hiring filter has shifted accordingly. The opening question: "Tell me what you built." Not managed. Built. Your resume is becoming a collection of the agents you've constructed. You're not just bringing yourself to a role; you're bringing a team. How agent portability works technically is still unsettled, but the candidates thinking that way are the ones who can build an agent-led function.
The Governance Problem Is Coming
Google let agents proliferate for about 18 months. They wanted the garden full. Then the sheer number forced a curation process: which teams are actually scaling something reusable versus running a one-off experiment. The fix wasn't a crackdown. It was shared infrastructure, so a good agent built by the Chrome team becomes available to the Cloud team on the same platform.
For ops teams earlier in the journey, that sequence matters. Encourage building first. But plan the governance layer now, because the alternative is a mess of disconnected agents with no shared data model, no QA loop, and no way to measure what's actually working across teams. AI-native firms are reportedly running about 25% smaller with flatter hierarchies and roughly 30% higher valuations per employee. That math only works if the agents are coordinated, not siloed.
What to Actually Do This Week
Pick one end-to-end workflow, not one task. Go where your data is cleanest and your pain is highest. Build a 5-minute training bite for that single workflow. Measure whether quality holds as volume scales. Treat every agent like a new hire that needs real onboarding: context, training data, guardrails, and a clear success metric.
The hypothesis: if we redesign one high-friction workflow around an agent (instead of bolting the agent onto the existing process), then cycle time and output quality will both improve, because the constraint was never model capability. It was process debt.
Kennedy Ellis's meta-point landed harder than any individual tactic. Marketers used to get their gratification from the output. Now the input is the most valuable thing you create, because the agent handles the output. For ops pros who've always known the system matters more than any single campaign, that's not a new idea. It's just the first time a Google VP said it into a microphone at a conference.
