Pair a person with an AI and, on the right kind of task, the pair beats the person working alone. That is the centaur model, borrowed from chess, where teams of humans and machines once outplayed either on its own. It is a genuinely useful frame for AI at work, as long as it is stated honestly, with the conditions attached.
The honest version is narrower than the marketing version. The verified claim is not that human plus AI beats everything, everywhere. It is that on tasks within the tool's competence, an AI-assisted human outperforms an unassisted one, and that the human's role in steering is what keeps the pairing from failing.
What the evidence supports
The clearest field evidence is the randomized trial of 758 BCG consultants. On tasks inside the model's frontier, consultants using GPT-4 completed 12.2% more tasks, 25.1% faster, and at roughly 40% higher quality than those with no AI. That is the centaur advantage in numbers: the assisted human beat the unassisted one, decisively, on suitable work.
Two features of the finding matter as much as the headline. First, the gain was largest for lower performers, which the study observed as a compression of the performance distribution; the pairing lifted the people who had the most room to gain. The same shape appears in the support-center study, where AI raised novice output by up to 34% and barely moved the experts. Second, and crucially, the advantage was conditional. In the same BCG experiment, on a task built to fall outside the model's competence, AI-assisted consultants did worse than those without the tool, because they trusted a confident wrong answer. The centaur only wins when the human keeps hold of the reins.
The advantage is a design problem, not a slogan
Because the benefit is conditional, capturing it is a design problem: which tasks get the pairing, and how the human stays in the loop. A large audit of worker preferences found that people themselves want this division drawn carefully, automating low-value tasks while preserving shared agency, AI executing and the human keeping judgment, on high-value work. That preference is not sentiment; it maps onto where the centaur model actually performs. Put the pairing on suitable, verifiable tasks with the human steering, and it wins. Put it on tasks past the frontier, or let the human disengage, and it loses.
What the evidence does not license is the stronger boast that human plus AI always beats AI alone, or human alone, on every task. On some narrow, well-bounded tasks the tool alone may be enough; on tasks past its frontier the human alone is safer. The defensible, useful claim is the qualitative one: for a wide band of real knowledge work, a skilled human steering a capable tool is the strongest of the available options, and the skill of steering is the scarce ingredient.
This is why the centaur is a management claim as much as a technical one. The pairing does not assemble itself; someone has to decide which tasks fall inside the frontier, route them to the human-and-AI team, and keep the riskier work under closer human control. Left to chance, people apply the tool everywhere, capture the wins on suitable tasks, and quietly rack up losses on the unsuitable ones, ending up roughly where they started. The advantage is real, but it has to be engineered.
Trust the person, not the credential
As Ricardo Semler has told it, Semco once put a maintenance technician without a degree in charge of two hundred people, because everyone trusted him, on the view that competence does not arrive with a diploma. The centaur model rewards the same instinct. The advantage does not go to the most credentialed operator; it goes to the one with the judgment to know which tasks the tool can handle and when its confident answer should be doubted. That judgment is the human half of the centaur, and it is not optional equipment.
Half the animal is still human
The centaur is a useful image precisely because it refuses to erase the human. The machine supplies speed and breadth; the person supplies aim, judgment, and accountability. Remove the human half and you do not get a faster centaur, you get a confident animal running in the wrong direction. The organizations that win with AI are not the ones that hand tasks to the tool and walk away. They are the ones that build the pairing deliberately, on the right work, with the person firmly in the saddle.
Sources
- Dell'Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Lakhani et al. (2025). Navigating the Jagged Technological Frontier (Organization Science). RCT with 758 BCG consultants: on tasks inside AI's frontier, GPT-4 users did 12.2% more tasks, 25.1% faster, ~40% higher quality; on tasks outside it they erred more. https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838
- Brynjolfsson, Li & Raymond (2025). Generative AI at Work (Quarterly Journal of Economics). Field study of 5,172 support agents: AI raised resolutions per hour ~15% on average, up to ~34% for novices and near zero for the most experienced. https://academic.oup.com/qje/article/140/2/889/7990658
- Shao, Zope, Jiang, Pei, Brynjolfsson, Yang et al. (2025). Future of Work with AI Agents (arXiv 2506.06576). Across 1,500 workers, 844 tasks and 104 occupations, people want AI to automate low-value tasks but prefer shared agency (AI executes, human keeps judgment) on high-value work. https://arxiv.org/abs/2506.06576