“The model eats the world” is redefining competitive edge. Now what?

Welcome back to AI@Work, a newsletter and video series that decodes the future of business. 

As we enter the new year, we are also entering a world of work where the real competitive edge is how fast you can teach an agent to beat the best human at a job. That may sound radical, but the patterns are already visible at Frontier Firms across industries. Beyond AI becoming a daily part of our tools and workflows, it is fundamentally changing how new value-generating capabilities are created. 

For decades, organizations have run on a mix of deterministic software, fixed processes, and human expertise. But something new is emerging. In domain after domain, we’re seeing that there is no amount of code, and in many cases no amount of human effort, that can consistently outperform a well-trained model paired with the right evaluation loop. I describe this shift as “the model eats the world.” 

This represents a new value paradigm—one where advantage comes not from how much expert knowledge you have, but from how quickly you can build, scale, and improve that expertise with agents. 

The model eats the world—and unlocks the future 

A common refrain of AI skepticism is: Isn’t AI hitting diminishing returns? Haven’t we reached the edge of what these models can do? 

Most people still think of AI as a general-purpose reasoning machine—broad, flexible, and capable of answering most queries. But the real breakthrough comes when you feed a model high-quality data in a particular domain—finance, physics, marketing, accounting, customer support, engineering. It stops behaving like a general-purpose chatbot and starts performing like a specialist in the field.  

Take for example, this collaboration between Yale and Google. Using 27 billion parameters, Yale researchers trained a Google model to understand how human cells behave. The model generated a novel hypothesis to make certain tumors more visible to the immune system, which is critical for effective cancer treatment. Here's the breakthrough: when scientists tested the AI’s prediction in a real lab setting, it worked—the insight translated into measurable biological results.  

These highly specialized models require the creation of measurable evaluation steps to teach the model the nuance of the domain and arm it with new capabilities. As the model conquers those eval sets, you’re effectively creating a custom fork of that model—an agent that sharpens with every iteration. The outcome is a system that keeps improving as long as you keep raising the bar. 

This unlocks new paradigms:  

• Expertise stops being a scarce, human-only resource. 
• Specialized capabilities become inexpensive to generate. 
• Every company can create its own domain experts—agents trained on its IP, processes, and way of working. 

When expertise becomes abundant, companies no longer gain advantage simply by hiring more specialists than the competition. They gain it by training and improving agents faster than competitors, and then effectively applying that expertise. 

Hill climbing powers the capability flywheel  

What’s driving this new engine of value creation is something we at Microsoft call “hill climbing”—replacing roadmaps with feedback loops and feature lists with measurable capability gains. The pattern looks like this: 

1. Pick a domain: Financial analysis, HR workflows, customer onboarding, supply-chain exceptions, product design—any space with a documentable workflow. 
2. Establish an eval set: A clear definition of what “good” work looks like in that domain—accuracy, structure, reasoning, completeness, compliance, speed, or even style. 
3. Train until the model hits the bar: Not once, but repeatedly and reliably, under many conditions. 
4. Retire the original eval set and create a harder one: Raise expectations. Increase complexity. Push the model up the hill. 
5. Repeat: The model gets better, then better again, often surpassing the average human and approaching the best human. 

This is how teams at Frontier Firms are now producing capabilities. It’s how we’re building increasingly specialized agents, like Agent Mode for Excel. And it’s how every organization will soon create specialist agents trained on their own institutional knowledge and workflows. 

This loop works just as well for service delivery or engineering as it does for product development. That’s the larger point: hill climbing isn’t a software-building concept. It’s a capability-building concept. Any domain with measurable work is a candidate. 

What it all means for your competitive edge 

Most of the attention today is on support functions—finance, HR, legal, operations—and AI is rapidly transforming them. But real competitive advantage will come from the core value-creating engines of firms: how products get built, how services are delivered, how engineering generates new IP, and design explores and iterates. 

In product development, agents will assess trade-offs, validate logic, and test edge cases. In services and consulting, expertise will be captured, shared, and strengthened across teams--enabling firms to deliver consistent, high-quality work at an unprecedented pace. In engineering and design, agents will help teams widen the aperture of possibility, catch errors earlier, and lean into new ideas faster than human-only processes ever could. 

The companies that pull ahead will be the ones that build high-velocity, eval-driven loops into these core areas. Rapid iteration means that even small gains compound quickly. The organizations that treat AI as a feature—or as a tool bolted onto old processes—will fall behind just as quickly. 

Get started now 

The shift to “the model eats the world” paradigm can feel overwhelming, but the first step is surprisingly small: pick a domain and start hill climbing. 

Choose a workflow that creates value for your business. Build an eval set that defines what “good” results look like. Tune the model to create an agent that can execute the workflow and produce those results. Give a small cross-functional team the time and space to iterate. Watch how quickly the agent improves—and how quickly your employees start working differently around it. 

That capability-improvement loop—train, test, raise the bar—that’s where new value is created. Once you’ve seen it work in one place, you’ll recognize its pattern everywhere. And once you know how to run it, you’ll know how to compete in the era ahead. 

3 more things 

Listen to this podcast: Influential economist Tyler Cowen joined the WorkLab podcast to share insights on the counterintuitive relationship between AI negativity in society and progress toward harnessing the technology to transform business.  

Check out these trends: Susanna Ray highlights seven ways AI is reshaping work in 2026—from digital coworkers to breakthroughs in security, science, and software—showing why leaders who pair humans and AI will win.  

Consider this question: How is AI redefining the way your organization develops new capabilities? Let me know in the comments.