The same AI can point in opposite directions. On one path it lifts the least experienced worker toward the level of the best, spreading capability across a whole workforce. On the other it concentrates power in the hands of a few people who own and direct the machines, while everyone else becomes a spectator to their own work. The technology is identical. The outcome is a choice.
The democratizing evidence
The case that AI levels skill is strong and empirical. In a field study of over 5,000 support agents, generative AI raised output most for novices, up to about 34%, and barely at all for experts, because it captured the tacit expertise of the best workers and made it available to the rest. In the BCG consultants' trial, AI compressed the performance distribution, lifting lower performers most. In developer experiments, the largest gains from GitHub Copilot went to the less experienced. Across three different settings the pattern repeats: AI narrows the gap between the expert and the beginner. Handled one way, it is one of the most democratizing tools knowledge work has seen.
The concentrating counter-current
But the same period is producing the opposite architecture. As organizations move toward AI agents, work on visibility into AI agents warns that delegating tasks to autonomous systems raises exactly the question of who controls them, and that accountability requires identifiers, monitoring, and logs, mechanisms that carry their own trade-offs for the concentration of power. McKinsey's account of the agentic organization sketches a structure in which a human team of just two to five people supervises an agent factory of fifty to a hundred agents running an entire process. Read optimistically, that is leverage. Read another way, it is a design in which a handful of operators command enormous capacity and everyone formerly in that process is gone. The tool that could lift the novice can also make the novice unnecessary.
Which path an organization walks is not set by the model. It is set by who gets access to the tools, who is trusted to direct the agents, whether the workings are transparent or locked behind a few gatekeepers, and whether the gains are shared or captured. Those are governance choices, made by people, not properties of the technology.
Trust does not come with a diploma
As Ricardo Semler has told it, Semco ran factories led by engineers without degrees and once put a maintenance technician without any diploma in charge of two hundred people, because everyone trusted him, on the principle that competence and trust do not arrive with a credential. That instinct is the democratizing choice in miniature. An organization that decides only a credentialed few may direct its AI, that treats the tools as instruments of a new priesthood, will concentrate power by default. One that decides capability is distributed, and puts the tools in the hands of whoever can use them well regardless of title, will spread it.
There is a companion instinct in Semler's story about profit sharing and open books: a fixed slice of each unit's profit paid to the people who produced it, with the books open so anyone could check the math. His point was that people behave like owners when they are treated as owners and can see the numbers. The AI question rhymes with it. Democratized AI shares both the tools and the gains, and keeps the workings visible. Concentrated AI hoards both and keeps the workings opaque.
The distinction matters because the two paths feel identical at the start. Both begin with a pilot, a productivity number, a rollout plan. The fork is quiet: who is allowed to direct the agents once they scale, whether an ordinary worker can see and question what a system did on their behalf, and whether the hours the tool saves come back to people as relief or leave as roles. By the time those choices harden into structure they are expensive to reverse, which is why they are worth making deliberately while the technology is still new enough that either future is genuinely open.
Choose the path on purpose
The uncomfortable truth is that the concentrating path is the easy one. It requires no deliberate effort; it is where cost-cutting and control instincts lead if left alone. Democratizing AI takes intention: widening access, making the tools legible, trusting people the org chart would not have trusted, and sharing the value the technology creates. Both futures are technically available right now, in the same building, with the same software. The one that arrives will be the one the organization actually chooses, whether or not it admits it is choosing.
Sources
- 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
- Peng, Kalliamvakou, Cihon & Demirer (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot (arXiv 2302.06590). Developers with Copilot finished a coding task ~55.8% faster than the control group; later RCTs replicate larger gains for less-experienced developers. https://arxiv.org/abs/2302.06590
- 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
- Chan et al. (2024). Visibility into AI Agents (ACM FAccT '24). As work is delegated to autonomous agents, accountability requires agent IDs, real-time monitoring and activity logs, each with trade-offs for privacy and concentration of power. https://doi.org/10.1145/3630106.3658948
- McKinsey (2025). The Agentic Organization. In client work, a human team of 2 to 5 people can supervise an 'agent factory' of 50 to 100 specialized AI agents running an end-to-end process (consulting estimate, not audited evidence). https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/the-agentic-organization-contours-of-the-next-paradigm-for-the-ai-era