In the augmentation view, AI expands human cognitive capability by making analysis cheaper. As the cost of analysis drops, inquiries that used to be uneconomical become worth pursuing. People can do more analysis, and the resulting value compounds.

Here, I’m going to argue in favor of this augmentation loop, both as a prediction of how the future will unfold and how organizations should look to adopt AI.

First, I’ll begin with the argument against replacement, because replacement is the easier one to imagine. If AI replaces human labor, then its success is going to reduce its own economic justification. Replacing people with AI reduces demand for workers while increasing the supply of people available for the remaining jobs. That lowers the cost of human labor, changing the economics of having a person rather than an AI do the work.

Replacement saves money only when the automation pays back the effort of building it. As an example, an Excel macro might pay for itself after five repetitions, while an AI agent might need 5,000. Once the easy tasks are automated, the remaining ones are harder. If hiring a person also becomes cheaper, the agent must run longer to justify the investment. This won't necessarily happen uniformly across occupations or industries, but it creates a natural break on AI adoption.

For AI providers, this raises a question about how long replacement-led growth can continue. For AI consumers, it raises the risk of investing heavily in workflows that become uneconomical.

Replacement also carries costs beyond the economics of automation: it replaces the act of doing with the act of monitoring and reduces opportunities to learn by doing. Lisanne Bainbridge, in her work The Ironies of Automation, looks at why increasing automation doesn’t pay off as much as it should. The gist is that humans remain essential to the process, but the skills they need to actively participate atrophy.

If you move a person from doing a task to monitoring a task, their work becomes more cognitively demanding. Consider, as an example, air traffic controllers. Air traffic control illustrates the demands of sustained monitoring: it is widely regarded as stressful, cognitively demanding work. In an agentic world, everyone becomes an air traffic controller.

The other irony is about preparedness. Automation removes the routine work that prepares people for difficult situations. When an exceptional situation arises and a person does need to intervene, the case is more likely to be particularly complex. Without that buildup of experience, the worker is less prepared and less likely to succeed. As a result, the exceptions become most costly, which negates the original cost saving of the automation.

This learning by doing is actually a big part of the benefit of having a human worker. If you think about having a junior worker—no offense to people just starting out in their careers—the value of a junior employee is often much lower than a senior employee because the senior employee can complete the same tasks faster.

But you also have to consider the discounted value of future work by your junior employee. Basically, you don’t get senior employees without training junior employees, and that discounted value of their future work offsets lower productivity today.

Now let’s turn our attention to the augmentation loop. Consider the cost and value of answer a question. For an organization, if the answer is worth less than the cost to obtain it, an organization has no reason to pursue it. This is an expression of bounded rationality, the idea that people and organizations seek out not the best solution, but a solution that is “good enough”.

Agentic AI lowers the cost of answering questions by handling specific parts of the process: collecting information and assembling context, connecting analytical tools, collecting and analyzing results, and presenting them in a human-readable form. It leaves room for the things that humans are better at: value judgments, considering wider implications, and bringing in distributed and latent knowledge.

As the cost of answering questions falls, more answers become worth pursuing: their answers become worth more than they cost to find. New roles will form around answering them. Those questions vary by business and firm, so identifying yours is central to taking advantage of AI.

The answers to these questions often open new lines of inquiry: they reveal gaps in our knowledge, challenge our assumptions, or open new possibilities. When those new questions are worth investigating, they sustain the augmentation loop. AI makes further inquiry cheaper, and the resulting insights can improve decisions and create a reason to use it again. That feedback depends on the value of the insights, not merely the volume of answers.

There’s a useful analogy here to Jevons’ Paradox. During the Industrial Revolution, more efficient steam engines could produce more energy from the same amount of coal. You might think that coal demand would decrease; we could get more energy out of a piece of coal, so we would use less of it. Instead, as engines became more efficient, energy became cheaper, so we started using more of it. The trend was rapidly expanding coal consumption rather than the expected decrease.

In the augmentation loop, that expanded activity means humans can pursue more questions and do more with the answers. The idea is to benefit from agentic AI being used to increase human agency: allowing humans to do more because they’re not bogged down by tasks that we can offload.

Many of the tasks we can offload involve coordination. Users need to move data, reconcile data from different systems, prepare inputs, and communicate across organizational or technical boundaries. That is the type of thing that AI is very good at today. Delegating these tasks gives agency not only to the AI, but ultimately to human workers as well as they are given more capacity to pursue meaningful, impactful work.

But does this augmentation loop still leave a distinctive role for humans? There is a clear objection here: models can reason too, and models can often reason better than humans. It’s important that we differentiate reasoning and judging.

A model can reason well. It can consider implications, challenge assumptions, and compare options. But there is still a role for humans in providing extra context that lives outside the system, choosing which goals and priorities are important, and reconciling things when competing values or ethics are at play. Reasoning through a trade-off does not, by itself, settle which trade-off people should accept.

One way to distinguish human and AI roles is to ask how an answer can be assessed. Reasoning can be tested against objective criteria when an answer is verifiable; broader judgment involves choosing among competing priorities and values.

A good example of model reasoning is OpenAI’s highly publicized finding that the Navier–Stokes equations have a singularity in certain situations. The reason AI could do this—a result that has evaded human mathematicians for decades—is that it is a question with a definitive answer. Mathematical problems have objective criteria of what constitutes a valid answer. There is a right or wrong.

For judgment, think about a Supreme Court case. The task is not simply to verify an answer, but to weigh conflicting interpretations and determine which should be accepted as correct. A model could make that judgment too; the distinction is not that humans alone are capable of it. Rather, decisions that settle competing values are better left to humans, who should remain accountable for the choices made. In the augmentation loop, AI can inform that judgment without taking over responsibility for it.

That distinction between reasoning and judgment has implications for organizational design. Agentic AI creates new organizational-design problems as work becomes more abstract. AI handles more tasks with definitive answers: data gathering, reconciliation, and routine calculations. Human work shifts toward framing problems, weighing priorities, and evaluating outcomes.

This is where organizational design becomes a source of competitive advantage. Competitive advantage will come from asking good questions. As the ability to answer questions becomes increasingly commoditized, firms will need to distinguish themselves by asking good questions and using the answers to make good decisions. A process that connects inquiry to judgment will separate firms that benefit from AI from those that do not.

To identify the new roles and tasks that will emerge as AI matures, ask people what they would do with more time. Most already know: “It would be great to do this, but I never get around to it.” Their answers point to opportunities in your organization.

Another tactic is to consider tasks that gain or compound value when repeated at scale. One scenario analysis offers a glimpse of a possible future; 1,000 scenarios provide much more information. A task of limited value on its own may become more valuable when repeated. A second opportunity is to increase the frequency of analysis. Work currently done once a year or once a quarter could become continuous, allowing people to act without waiting for the next scheduled update.

Whether you’re running more scenarios or updating an analysis constantly, the measure of success needs to change. Measure the value of output through being able to make better-informed decisions, not simply being able to make more decisions or faster decisions.

Augmentation is an opportunity, not an inevitability. AI creates the capacity to investigate more questions, explore more possibilities, and make better-informed decisions. But that capacity needs direction.

The organizations that benefit most will not necessarily be those that automate the most work or produce the most analysis. They will be those that recognize which questions are worth asking, exercise sound judgment of the answers, and create the conditions for people to act on what they learn.