AI Generates More Human Work, Not Less, and That Changes Everything for Those Who Lead
There is a narrative that circulates comfortably in boardrooms: artificial intelligence is going to eliminate positions, reduce payroll, and free up capital. It is a comfortable narrative because it takes the form of a clean financial decision. The problem is that the data does not support it.
Jeff Bezos said it plainly in a recent interview on CNBC: AI is not going to empty the labor market — it is going to generate a talent shortage. His analogy was precise. An engineer who spent years digging a trench with a shovel does not disappear when handed an excavator. He digs more, faster, on projects that were not previously viable. The work is elevated, not extinguished.
What is happening at the real frontier of AI adoption confirms this thesis in a way that should be deeply uncomfortable for those who made staffing decisions based on the opposite narrative.
When Automation Multiplies Expert Work
Dan Shipper, CEO of Every, published an analysis that is worth reading carefully. His company automated everything it could possibly automate using AI agents. The result was that the team grew from four to more than thirty people. Not in spite of the automation, but precisely because of it.
The mechanics behind this phenomenon are less paradoxical than they appear. When AI takes on the standardized parts of a process, it does not eliminate the need for expert judgment — it multiplies it. Someone has to define what counts as a good result. Someone has to review the agent's output before it reaches the client. Someone has to decide what to do with that output within the broader organizational context. AI collapses the middle task. Humans hold up both ends.
Shipper describes this using a process geometry that carries concrete organizational implications: at the beginning, humans establish the framework. In the middle, AI executes. At the end, humans judge, extend, and decide. That is not a cycle that reduces the human burden. It is a cycle that shifts that burden toward decisions of greater cognitive density.
Data from Anthropic on the use of its models with real users points in the same direction. In typical knowledge work tasks, execution time falls by approximately 80%. That saving does not translate into less work; it translates into a higher volume of initiatives, faster decision cycles, and a larger surface area of human coordination. McKinsey estimates that with the adoption of AI agents at scale, around 57% of working hours in the United States are technically automatable with technology available today. If that figure were realized, the potential additional economic value would reach $2.9 trillion annually by 2030 in that market alone. The problem does not lie in the capability of the technology. It lies in who supervises, coordinates, and integrates that new volume of output.
The MIT Sloan research that tracked the impact of AI between 2010 and 2023 found something that rarely appears in headlines: when AI automates only a portion of the tasks within a role, employment in that role can actually grow. And in high-salary roles with high exposure to AI, employment growth was approximately 3% over five years. This is not destruction. It is reconfiguration.
The Organizational Cost of Believing the Wrong Narrative
What interests me about Shipper's analysis is not only the mechanics of the process. It is what it reveals about the conversations that many organizations are actively avoiding.
When an executive team adopts AI under the premise that it will reduce their dependence on expert human talent, they are building a strategy on a false premise. And strategies built on false premises do not collapse all at once. They rot slowly. The most common symptom is a growing backlog of decisions that AI cannot make, piling up on top of a team that was downsized or that was never trained to operate within the new framework.
What Shipper identifies as the new organizational bottleneck is a governance problem, not a technology problem. AI produces at a speed that the human supervisory structure cannot always absorb. And when that gap is never named, the organization begins to operate on outputs that nobody truly reviewed carefully — they were only reviewed quickly. The difference between those two things has consequences that take months to become visible, and when they do, they present themselves as inexplicable errors.
There is another effect that few organizations are measuring honestly: the homogenization of output. When everyone in an industry uses the same models to produce documents, analyses, presentations, and communications, the result is a convergence toward legible mediocrity. Shipper states this bluntly: abundance generates uniformity, and uniformity destroys differential value. A financial analysis that looks like every competitor's does not provide any advantage. A communication strategy that sounds like the industry average does not build positioning. In that context, the real scarcity becomes the human judgment that produces something that does not resemble what AI would choose by default.
Goldman Sachs Research arrived at a similar conclusion from a different angle. Their analysis finds that so far there is no statistically significant correlation between local exposure to AI and growth in unemployment, layoff rates, wages, or hours worked. Zero measurable macro impact, despite the sheer volume of narrative about job destruction. What they are observing instead is a redistribution of tasks within existing roles, accompanied by growing demand for the skills that AI cannot replicate: complex coordination, contextual judgment, interpersonal trust.
The Work the Organization Does Not Yet See
There is a type of work that AI adoption creates and that few organizations account for correctly: the work of keeping agents functioning well.
Shipper has a team at his company dedicated exclusively to ensuring that AI agents operate within acceptable parameters. This is not a temporary implementation cost. It is a structural operational cost. Models degrade in certain contexts, produce outputs that require continuous calibration, and the threshold of what counts as "good enough" shifts over time and with client expectations. That requires engineers, judgment, and decisions that cannot be delegated back to the AI.
Boston Consulting Group estimates that over the next two to three years, between 50% and 55% of positions in the United States will be significantly reconfigured by AI. Reconfigured, not eliminated. That distinction is not semantic. It means that the organization that enters that process without having prepared its people to operate within frameworks of supervision, judgment, and output integration will find that it has powerful tools and human capacity that is misaligned with what those tools require.
The most costly mistake an executive team can make right now is not moving too slowly with the technology. It is moving at technological speed while operating the human structure at the speed of the past. AI accelerates the production cycle. If the organization does not simultaneously build the capacity for supervision, governance, and judgment at that same scale, what accelerates is not value. It is the volume of outputs that nobody is truly validating.
The question worth holding onto over the coming months is not how many positions AI can automate. It is how many expert judgment positions the organization needs to create so that the automation produces something worth having.
Shipper sums this up with a phrase that deserves far more attention than it typically receives in C-level conversations: once a situation has been reduced to text, it has been converted into a corpus. And a corpus is a corpse. What the human needs to do is precisely what has not yet happened — what cannot already be documented — what needs to be named right now, in this context, with this client, under these conditions. That is where AI falls short. And that, paradoxically, is where there is the most work left to be done.
