Monks Spent Two Years Building What Nvidia Now Uses to Sell Everyone Else

The best product demo Nvidia can show advertising agencies isn't a graphics card benchmark. It's Monks.

S4 Capital's production arm spent two years - starting roughly when ChatGPT went mainstream - building internal AI infrastructure from scratch. The result: 8 of their top 10 clients regularly use AI tools, and according to Adweek, about 40% of their 7,000-person workforce actively uses Monks Flow, an internal platform threading together multiple AI models for research, content generation, and media planning. Now Nvidia is taking that story on the road, using real-time AI video generation as the hook and Monks as the proof that it works.

This is how first-mover advantage works in the current AI cycle. The companies that will set the terms for the next five years aren't the ones with the best access to models - everyone has roughly equal access to the same APIs. They're the ones who spent the last 24 months making mistakes, building workflows, and teaching their organizations how to move.

The Two-Year Gap Is Structural, Not Technological

When McKinsey's operations practice published its latest analysis on scaling agentic AI for operational breakthroughs this week, one claim stood out: waiting to implement carries real risk. That warning is easy to dismiss as urgency-selling. But the pattern across sectors makes it harder to wave away.

In life sciences, Stanford professor James Zou described how AI agents are being integrated into biology R&D and drug discovery - not just analyzing literature, but running multi-step reasoning across molecular structures, drug interactions, and immunology data. The pharmaceutical companies that started building these workflows two years ago aren't just running faster experiments. They're developing institutional intuition about when to trust the agent, when to intervene, and how to structure problems so AI can actually solve them.

In creative production, Nvidia's pitch to agencies is built around Monks having figured this out already. Henry Cowling, Monks' chief innovation officer, described the goal as going "from a blank page to a finished commercial with an absolute minimum of human involvement." The technology didn't change overnight to enable that. The organizational capacity did.

In operations broadly, McKinsey partner Michael Chang framed agentic AI as fundamentally about doing things, not analyzing them - a distinction that matters enormously for where the value lands. The companies capturing that value are the ones whose people already understand how agentic systems fail, and have built workflows around those failure modes.

This is what a two-year head start actually buys: not better tools, but better judgment about tools.

What "Building Internal AI Infrastructure" Actually Means

There's a tendency to equate AI adoption with access - getting an API key, enabling a feature flag, subscribing to a platform. Monks Flow is a useful corrective to this assumption.

The platform doesn't just provide access to AI models. It connects them: research models feeding into content generation pipelines feeding into media planning workflows. Forty percent of 7,000 employees using it regularly means the tool had to be usable by people whose primary job isn't AI - producers, strategists, account directors. That usability doesn't come from the underlying model. It comes from two years of iteration on what those specific people actually need the system to do.

This pattern looks familiar. What separates Klarna's AI agent deployment from most enterprise experiments isn't model quality - it's governance architecture: the decisions about where agents have authority, where humans retain control, and how failures get caught before they compound. The full breakdown of why Klarna's approach works applies directly to what Monks built at the production layer.

In biotech, the equivalent infrastructure is the curation layer that makes AI-generated hypotheses scientifically credible rather than just statistically plausible. Zou's point about AI scientists isn't that they replace human researchers - it's that they can run parallel hypothesis generation at a scale no human team can match. Calibrating when those hypotheses deserve wet-lab resources requires judgment that only comes from operating with AI agents in real research contexts over time.

The throughline: the moat isn't the model. It's the 18-24 months of organizational learning that accumulates when you actually run things through AI and see what breaks.

This is the kind of structural pattern STI's research tracks systematically - particularly how early operational decisions compound into durable advantages or durable gaps depending on when organizations commit.

Why McKinsey's Warning Is About the Gap, Not the Timeline

"Waiting to implement could be risky" sounds like consulting urgency. In context, it's a more specific claim.

The risk McKinsey is describing isn't primarily that organizations will fall behind on feature access. Every enterprise can buy the same Nvidia hardware, access the same foundation models, hire from the same talent pool. The risk is falling behind on operational fluency - the organizational capacity to deploy AI in ways that actually change outcomes rather than producing sophisticated reports about what AI could theoretically do.

The earlier NRF 2026 data on retailer hesitation illustrated the failure mode clearly: companies exploring agentic AI "with visible caution," refusing to hand decision-making authority to systems they haven't had time to calibrate. That caution is rational for an organization without operational experience. It's also self-reinforcing - you can't build calibration without deployment, and you won't deploy without calibration.

Monks doesn't have this problem because they solved it two years ago. Their 40% adoption rate isn't a success metric to celebrate. It's proof that the tools passed a much harder test: they became genuinely useful to non-AI-native workers doing real work under real pressure.

The Talent Reallocation Signal

One detail from Nvidia's pitch deserves more attention than it usually gets. Cowling described the goal as going from blank page to finished commercial with "minimum human involvement." Nvidia's Richard Kerris framed the broader shift: "Your computer has now become your real assistant."

These aren't just efficiency claims. They describe a different organizational design - one where the humans who were doing execution work are now doing judgment work. That transition is disruptive regardless of company size. The organizations that started it two years ago have a workforce that has already made this adjustment. The organizations starting now face the disruption at the same moment they're trying to scale adoption. That's a harder problem, and it doesn't get easier with better tools.

The Cross-Industry Signal

When the same dynamic plays out simultaneously in operations consulting, pharmaceutical R&D, and advertising production, it's not three separate AI trends. It's one structural shift expressing itself across industries at different speeds.

The pattern: agentic systems that can do things - not just recommend, not just analyze, but execute multi-step tasks - compress the value chain for whoever deploys them first. That compression creates advantages that are hard to close because they're embedded in organizational practice, not technology access.

The McKinsey AI budget analysis from last week flagged a related implication: enterprises systematically underfund the governance layer that makes agentic AI safe to deploy at scale. The organizations that started deploying two years ago have been building that governance layer implicitly, through operational experience. Everyone else is now trying to construct it deliberately, at speed, while also trying to deploy.

This is the actual gap McKinsey is pointing at. Not "you'll miss out on AI features." Something more like: there's a growing class of organizations for whom AI is already native, and catching up requires solving two problems simultaneously instead of sequentially.

The Decision That Compounds

Nvidia is using Monks as a sales pitch because Monks is the answer to the question every skeptical CMO asks: "Has this actually worked, at scale, with real workers doing real jobs?"

The answer is yes - but it took two years to build.

That's the uncomfortable implication of where this is heading. The companies, research labs, and agencies that will define best practice for the next wave of agentic AI - more autonomous, more integrated, more consequential - are largely the ones who already went through the learning curve on the current wave. Their advantage isn't a product feature. It's accumulated judgment about failure modes, governance requirements, and where human oversight actually matters.

For organizations still calibrating whether to move: McKinsey's operational argument, Zou's biotech case, and Monks' production numbers are all pointing at the same thing. The question isn't whether agentic AI will change how your organization works. The question is whether you build operational fluency before or after the organizations in your competitive set do.

If you're mapping your organization against these patterns, our analysis tools can help surface where the gaps are and which decisions are most likely to compound over time.