The path to AI productivity gains
Companies are investing heavily in AI, but many have yet to translate that investment into significant improvements in business performance. It’s still early, and much of AI’s promised economic value remains unrealized.
But Albert Meige, Director of Blueshift at Arthur D. Little (ADL), predicts companies will eventually climb the J-shaped productivity curve. Gains will come, he says, when companies redesign workflows around cognitive work that can now be delegated to AI. Central to that redesign, he argues, will be preserving the human capabilities that AI cannot readily reproduce.
AI stands to turn certain kinds of cognition into abundant commodities, Meige notes, but widely available capabilities offer little differentiation when competitors can access them, too.
Instead, Meige argues that the difference will increasingly be made by what companies don’t farm out to AI. The new scarcity lies in human initiative and the judgment required to wisely deploy AI. Cultivating human agency, cognitive diversity, and proprietary intelligence may therefore become a source of competitive advantage.
The new scarcity lies in human initiative and the judgment required to wisely deploy AI
What we lose when we add AI
As AI looms larger in the workplace, the risks of delegating too much cognition are also becoming clearer. What Meige terms the AI “convenience trap” for organizations has three parts. Leaders lose decision-making capacity and accountability. Uniformity and lack of marginal perspectives start to affect companies that use the same AI models trained on the same original data. And employee de-skilling becomes commonplace.
Growing evidence suggests heavy AI reliance can erode human critical faculties. A 2025 study in The Lancet Gastroenterology & Hepatology examined what happened to human expertise after four medical centers introduced AI tools to detect polyps during colonoscopies. After endoscopy providers began using AI, their ability to detect polyps without its help sagged significantly. A 2025 preprint by MIT researchers reported on brain connectivity among essay writers relying on either a large language model (LLM), a search engine, or their own brains to complete a task. The researchers found the weakest brain connectivity among users relying on the LLM, reporting, “Cognitive activity scaled down in relation to external tool use.” (The research has not yet been peer reviewed.)
Trade-offs with cognitive offloading aren’t new, as anyone can attest who memorized dozens of phone numbers in the 1980s and now finds themselves dependent on a digital contacts list. Researchers have documented declines in the ability to navigate without help among people who rely on GPS. Engineers who used slide rules, which require users to understand orders of magnitude, noticed people lost a deep feel for the numbers once electronic calculators took over.
But as AI tools grow more capable and we delegate more deliberative thinking to them, there may be more going on than the familiar trade of skill for convenience.
A 2026 Wharton School study described a phenomenon of cognitive surrender, in which people passively accept AI outputs with little critical judgment, overriding their own intuition and deliberation. The researchers framed this as a design and education challenge: how to capture the benefits of AI-assisted cognition without sacrificing critical thinking and accountability.
ADL has found related results. In a recent experiment, the company found that time-pressured employees who used AI to create a presentation had a poorer grasp of their topics than when they didn’t use AI, according to ADL’s Chairman and CEO Ignacio García Alves.
As he puts it, “The tool can make you lazy.”
Heavy AI reliance can erode human critical faculties
Avoiding the AI convenience trap
Avoiding that trap, Meige argues, requires three mutually reinforcing pillars: human agency, proprietary intelligence, and cognitive diversity.
Human agency inheres in the ability to make well-considered decisions, and leaders must be “architects of intent,” in Meige’s words. This means deliberately thinking problems through before turning to AI, clarifying intentions, and reducing the risk of anchoring bias. It also means remaining fully accountable for decisions, including those informed by AI.
This has implications for the talent pipeline. Because it’s often good at entry-level thinking tasks, AI may seem to make junior employees less important in the short term. Yet jettisoning those workers would create a lost generation, García Alves points out. Without them, who will make expert judgment calls about AI output in the coming years?
“The fact that you combine that AI system or the junior person with a senior person that has experience is extremely important,” García Alves says. “If there is this hole in the pyramid, what do you do?”
Proprietary intelligence starts with a limitation: much of what AI knows comes from information that has already been recorded. Humans can find fresh information beyond AI’s reach by engaging with experts. They can also learn through fast, iterative experimentation, Meige notes, creating original data and insights that do not yet exist in AI training sets.
Cognitive diversity, the crucial third pillar, emerges from disagreement, perspectives that resist easy reconciliation, and from the disciplined process of arguing ideas into stronger conclusions, according to Meige. It counteracts the gray sameness that can settle over organizations whose AI tools draw on the same pools of information.
What does cognitive diversity look like? According to García Alves, it means combining as many types of intelligence as possible, including AI, experts, and colleagues from a wide range of backgrounds.
“In our business, if you want to be creative, if you want to create collective intelligence and diversity, we need to have people who think differently,” García Alves says. “We have engineers, and we have people with a business education. We might have people with an art education. We might have people from different countries. We might have people that are autistic, et cetera.”
Combining people of differing ages and experience levels also matters. Demand for senior people is higher now, García Alves says, and value emerges when well-seasoned senior and tech-native junior employees work together with AI.
Generative AI (GenAI) can be good at brainstorming new ideas, notes Rob Austin, a Professor of Digital Innovation at Ivey Business School and former chair of Harvard University’s CIO executive program. Austin studies what he terms “divergent intelligence,” which he defines as an organization’s capacity to maintain sufficiently broad and independent viewpoints to recognize value in unfamiliar forms.
But AI’s tendency toward sycophancy might weaken its ability to pursue peripheral paths or spot the value in unconventional ideas, Austin notes. That’s a weakness it shares with many humans, he adds.
“There is this kind of fundamentally human problem associated with seeing value, even when we succeed in creating it,” Austin says. “And that’s, I think, where things like divergent thinking come in.”
Cognitive diversity counteracts the gray sameness that can settle over organizations whose AI tools draw on the same pools of information
Welcoming & valuing neurodiversity
One source of that divergent thinking is neurodiversity, such as autistic employees and those with attention deficit hyperactivity disorder (ADHD). Research from Harvard Business Review suggests neurodiverse teams that include at least one neurodivergent person can outperform neurotypical ones.
Austin thinks both GenAI and neurodiversity can help maintain divergent intelligence within a company and overcome pressure to conform. The employees described earlier in this article, for instance, were hired through an autism hiring program. They were willing to point out what others didn’t.
But hiring for cognitive difference is not enough. Inclusion must be explicit and constructive, going beyond simply hiring neurodivergent employees, if companies are to benefit from those perspectives.
Welcoming neurodivergent employees means examining and redesigning features of organizational life usually left implicit, Austin says. For example, organizations might send meeting agendas in advance, offer virtual participation, or allow employees to contribute through chat rather than requiring everyone to participate in the same way.
Another key tactic is asking all employees how they work best — how they prefer to receive feedback, whether they are derailed by interruptions, and other usually implicit features of the work environment — and then interacting according to their needs. Austin cites New York City engineering firm Ultranauts, three-quarters of whose professionals are autistic, as a company that has built this type of inclusion into its foundations.
Austin says the broader goal is to make explicit the workplace assumptions that usually go unquestioned and design around differences in how people work.
Both GenAI and neurodiversity can help maintain divergent intelligence within a company and overcome pressure to conform
How cognition will make the difference
The practical challenge, then, is not merely preserving these forms of human cognition but continuing to develop them.
Valuing and nurturing expertise may be especially important because it takes expertise to evaluate AI outputs — particularly on complex questions. Harvard Business School researchers have described AI as a “jagged technology frontier,” with its strengths and weaknesses varying in nonintuitive ways relative to those of humans.
“It’s very difficult to draw a line and say ‘Outside this, don’t use generative AI. Inside this, do use generative AI,’” Austin says. “Our intuition and the way we categorize problems doesn’t align very well with what GenAI is good at or bad at…. Moreover, it’s a moving target, because GenAI is getting better at things.”
For now, García Alves calls for explicit instruction in higher-order thinking skills — reflection, synthesis, judgment, understanding of context — so employees learn to challenge AI outputs as a matter of course.
“We need to train the junior person to be a little bit critical, so that they always say, ‘Does that make sense?’” García Alves adds.
Because AI’s capabilities are uneven and constantly changing, Austin argues that people closest to the work also need room to experiment with it.
“You have to actually try it to know whether it’s going to work well,” Austin says. “Put the tools in [workers’] hands and let them figure out … what it’s good at, what it’s not good at.”
For CEOs, the emerging challenge is not simply deciding what to automate, but determining which human capabilities need to be preserved and strengthened alongside AI. As AI itself becomes more widely available, competitive advantage may increasingly depend on how organizations combine it with judgment, expertise, initiative, and cognitive diversity — and on which forms of cognition they deliberately choose not to outsource.
“We need to train the junior person to be a little bit critical, so that they always say, ‘Does that make sense?'"
Key Takeaways
- Make human judgment your competitive edge. As AI takes on routine thinking, keep strengthening human judgment, initiative, and higher-order thinking.
- Don’t fall into the AI convenience trap. Overreliance can erode expertise, accountability, and independent decision-making.
- Build what AI can’t commodify. Invest in human agency, proprietary intelligence, and cognitive diversity as sources of differentiation.
- Seek out thinkers who see what others miss. Differences in neurotype, background, age, expertise, and thinking styles can challenge groupthink and uncover overlooked opportunities.