AI at work: Reasoning models and the future of business
We are now living in a new reality—one in which AI can think and reason like humans, solving complex problems that have stumped even the most capable experts. This reality emerged just a few months ago, when OpenAI released the first of its AI “reasoning” models, which can understand and solve problems by making logical inferences and adapting to new information. More recently, DeepSeek made waves with a reasoning model that was developed more quickly and cheaply than we thought possible, and Anthropic released a hybrid reasoning model that can handle both immediate responses and those that require deeper consideration.
Let’s decode what happens when AI “reasons,” and what this remarkable new capability will mean for your business.
Most current AI models rely heavily on pattern recognition to answer questions almost instantly, but reasoning AI takes a more deliberate approach. It engages in logical, multi-step analysis—a process called chain-of-thought reasoning—to break down complex problems into smaller, more manageable chunks. That lets the AI explore different paths and backtrack or pivot when it’s wrong, similar to how humans solve problems.
Until recently, the go-to method for improving AI model performance was feeding it increasingly massive data sets during the training stage. Reasoning models leverage a different strategy called test-time compute, which involves using more processing power and time during the actual problem-solving stage. This means the AI takes more time and uses more resources to think deeply and provide more complete, accurate answers.
Reasoning AI isn’t perfect: humans still have a premium on common sense, and AI struggles with tasks that require understanding context beyond logical reasoning, such as interpreting nuanced language. Still, reasoning capabilities make AI extraordinarily powerful, able to solve problems that stymie other systems.
Here’s one example of that power in action: Ethan Mollick, professor at the Wharton School of the University of Pennsylvania, wondered if OpenAI’s o1 reasoning model could spot a recently unearthed math error in a research paper that briefly sparked a panic about the safety of black plastic cooking utensils. He asked it to “carefully check the math in this paper,” and it quickly pinpointed the mistake.
As Mollick wrote, “When models are capable enough to not just process an entire academic paper, but to understand the context in which ‘checking math’ makes sense, and then actually check the results successfully, that radically changes what AIs can do.”
Reasoning models are racking up astonishing results on intelligence benchmarks, as Mollick points out. The GPQA Diamond benchmark tests high-level science knowledge that isn’t available online, and OpenAI o3 beat human experts with a score of 87.7%. In FrontierMath, a set of incredibly tough math problems, o3 scored 25.2%, a major improvement over previous models. And on ARC-AGI, a test designed to be doable for humans but hard for AIs, o3 scored 87.5%, besting both previous AIs and the baseline human level.
All this isn’t to say that AI is going to take the place of human expertise and judgment. But reasoning as a scalable, always-on resource represents a powerful new paradigm. This is a watershed moment—one that every leader and organization will need to come to terms with.
Reasoning AI offers huge promise for business, across industries. Think of its potential for research and development. AI can now propose hypotheses and simulate outcomes on its own—thinking that’s well beyond the capabilities of standard prompt-and-response models. That advancement could cut years off traditional R&D cycles and bring breakthroughs in fields from renewable energy to pharmaceuticals.
More broadly, reasoning AI will upend many of our assumptions about work. Leaders should keep two things in mind: First, these models can perform cognitive labor that is equivalent to or better than humans. In other words, they can perceive, understand, reason, and execute—sometimes even create—at levels that approach or surpass human abilities. For every task your team needs to tackle, ask yourself, “Can AI do this job?” If the situation doesn’t call for uniquely human skills like judgment, nuance, originality, or emotional intelligence, the answer is now yes. We need to imagine a new division of labor for humans and AI—and new approaches to managing that labor.
Second, reasoning models change the economics of work. Historically, “acquiring” reasoning meant hiring humans, but that’s no longer exclusively the case. You can now rent or purchase cognitive labor on a consumption basis, similar to acquiring any other input for your business, from electricity to equipment. And that’s a very big deal. With efficient and affordable reasoning capabilities, your organization and industry will radically change. I expect that disruption to come from AI-native firms rather than incumbent companies. AI natives will have a competitive edge simply because they’ve been weaving AI into every process from the start.
It’s still early days for AI reasoning—and these are my initial thoughts. I’m certain that reasoning will crack open possibilities—and opportunities for business—that I haven’t even begun to imagine.
For more insights on AI and the future of work, subscribe to this newsletter.