Tech moves fast, but you're still playing catch-up?

That's exactly why 100K+ engineers working at Google, Meta, and Apple read The Code twice a week.

Here's what you get:

All delivered twice a week in just 2 short emails.

Join 100K+ engineers

Hello, fellow humans! Today, we’re looking at three studies that give us valuable insight into what makes AI implementations work, and what that means. The bottom line is that we need to be thinking more about how the organization works and less about the technical challenges of MCP and RAG. Don’t get me wrong, those are both still hurdles, but we’re discovering that they’re actually the easy part.

The blistering pace of AI development has created a catch-22 for organizations trying to leverage AI; on the one hand, AI platforms are incredibly capable compared to previous computer systems and, in the hands of skilled humans, can generate moderately successful work incredibly quickly. It promises so much value (depending on how you define that word), but AI is still unreliable. AI still requires skill, knowledgeable direction, and supervision to create valuable work. In other words, getting value from AI is much less a technological hurdle and more of a human and organizational one.

I’ve written about the risks for implementing AI at enterprise scale and how AI is best thought of as a force multiplier rather than a cost reducer. Three major pieces of research dropped in the past week that, taken together, reveal something crucial about how work is actually transforming.

All three of these studies agree that success comes down to organizational design, not technical capability. Let’s take a look at what that means.

McKinsey Global Institute released its comprehensive Agents, Robots, and Us analysis with a startling headline: 57% of US work hours could theoretically be automated. Before you panic or celebrate, here's the critical word: theoretically. Their actual finding is more nuanced and more interesting—human skills won't disappear, McKinsey predicts that they'll evolve into what they call "skill partnerships" with AI.

They've developed something called the Skill Change Index to measure this. Digital and information-processing skills face the most exposure, but here's where it gets interesting: they project $2.9 trillion in economic value by 2030, and that value comes from collaboration and workflow redesign, not from elimination.

That means that the money isn't in replacing people; it's in fundamentally rethinking how work gets done when humans and AI systems work together.

McKinsey identifies seven work archetypes ranging from people-centric (least AI-dependent) to agent-centric (most AI-dependent). Most roles sit somewhere in the middle of that spectrum. This isn't a binary shift—it's a gradient, and where your work falls on that gradient determines what you need to focus on.

The first throughline emerges here: We're measuring the wrong thing if we're just counting what AI can do. We need to measure how human skills change when AI enters the picture.

Product Talk’s Teresa Torres published Lessons from Nine Real AI Product Teams, and in the world of AI, even if you’re not on a product team, you’re on a product team.

This is where Product Talk's research becomes invaluable. They analyzed nine real AI product teams and found patterns that contradict a lot of conventional wisdom.

First, small teams win. We're talking 2-3 people, cross-functional, domain-expert-led. Not the large, specialized AI teams you might expect. Why? Because domain expertise drives product decisions more than AI technical knowledge. The people who deeply understand the problem space make better choices about AI implementation than AI specialists who don't.

Second, they all started narrow. Every successful team began with a specific use case, built evaluation capabilities around it, and then expanded. The teams that tried to solve everything at once struggled or failed.

Third, their evaluation systems evolved dramatically—from simple spreadsheet assessments to sophisticated frameworks. But they built those frameworks through use, not upfront.

If you're starting an AI initiative tomorrow, here's your playbook:

The second throughline: Small, focused, domain-expert-led beats large, comprehensive, and technically sophisticated every single time.

The third study, from Organizational Physics, makes the case that Google solved only half the problem with their multi-agent coordination breakthrough.

Google's breakthrough in multi-agent coordination is technically impressive—they've solved how to make AI agents work together efficiently. But the analysis found that that's only half the problem, and it might even be the easier half.

The harder problem is the organizational alignment.

It’s a provocative argument: you need to treat AI agents like human team members with defined roles, clear responsibilities, reporting structures, and accountability frameworks. It means decision rights and escalation paths. It means performance evaluation.

All of that is important because organizational intelligence—not just machine intelligence—determines whether AI implementation succeeds or fails. You can have the most sophisticated multi-agent system in the world, but if it doesn't align with how your organization actually makes decisions, communicates, and executes, it won't deliver value.

The kicker: the principles for aligning humans with humans are identical to the principles for aligning humans with AI. If your organization struggles with clarity around roles, responsibilities, and decision-making among people, adding AI agents will amplify those problems, not solve them.

When you synthesize these three research pieces, a framework emerges. Think of it as three layers you need to get right simultaneously:

The Strategic Layer comes from McKinsey's work. You need to map your organization's work exposure using something like their Skill Change Index. Identify which archetypes apply to different roles in your organization. Calculate potential value through workflow redesign, not elimination. Set realistic timeframes—McKinsey is projecting to 2030 for a reason.

The Tactical Layer comes from Product Talk's findings. Form small, domain-expert-led teams. Start narrow with high-value use cases. Build evaluation capabilities early and iteratively. Expand only after you've validated your approach.

The Organizational Layer comes from Organizational Physics' insights. Design org structures that can accommodate AI agents as team members. Establish clear roles, responsibilities, and reporting lines that include both humans and AI. Apply your existing organizational alignment principles—if they work for humans, they'll work for AI integration. Invest in organizational intelligence, because that's what determines success.

If you're leading an organization through AI transformation:

The uncomfortable truth emerging from all three of these analyses: most organizations are over-indexed on technical capability and under-indexed on organizational design.

We're buying AI tools, hiring AI specialists, and building AI strategies. But we're not redesigning our organizations to work effectively with AI. We're not starting small and learning fast. We're not treating AI agents as team members that need proper integration into our existing structures.

The $2.9 trillion opportunity McKinsey identifies won't come from the AI itself. It'll come from organizations that figure out the other two layers—the tactical implementation and the organizational integration.

The final throughline: The technical problem is increasingly solved. The organizational problem remains wide open. And that's actually good news, because organizational design is something you can influence directly, starting tomorrow.

The question isn't whether AI will transform work. It's whether your organization will transform how it works to capture that value. The teams getting it right aren't the ones with the most sophisticated AI. They're the ones with the clearest organizational design, the smallest and most focused teams, and the deepest domain expertise driving decisions.

That's the partnership framework that actually works.