Dan Shipper runs Every, a subscriber-supported media company that has been embedding AI agents throughout its operations for the past two years. He doubled his headcount during that time, going from 15 to 30 employees. When a Lenny’s Newsletter interviewer asked him to explain, his answer was direct: “Automation is a lie. Every time you automate something, you need a human on top of it, making sure that it continues working.”
That sentence should be on a poster for every product team projecting headcount reductions based on AI. It is not pessimism. It is a pattern with 160 years of economic history behind it, and it applies more directly to software and knowledge work than to any previous wave of automation.
Here are five places that pattern is showing up right now, and what it means for the product managers and designers worried about where they fit.
1. The Jevons paradox (the reason this keeps happening)
In 1865, economist William Stanley Jevons published “The Coal Question,” in which he observed something counterintuitive: more efficient steam engines increased total coal consumption rather than reducing it. Making coal cheaper to use expanded demand enough to offset the efficiency gains entirely. Total consumption rose.
Fortune reported in April 2026 that the same mechanism applies to AI in professional services. The cheaper AI makes a knowledge-work task, the more of that task organizations will buy. More efficient legal research does not mean fewer lawyers billed; it means more legal questions get asked. The same logic extends to software: when coding gets cheaper, the backlog of economically viable software projects gets longer, and more projects get started.
If this sounds abstract, consider what a five-person startup without AI can afford to build in a year compared to a five-person startup running coding agents. The second team ships more features, tests more ideas, and generates more product decisions per month. That is more work for product managers and designers. The constraint shifts from “how fast can we build this?” to “how clearly can we specify what to build?” That is a judgment problem, which is a PM problem.
2. Dan Shipper’s headcount experiment at Every
Dan Shipper’s account is worth reading carefully because it documents the paradox from the inside of a company that is actually doing this, not projecting what might happen. Every went from 15 to 30 employees while adding AI agents throughout its publishing and operations workflow.
The distinction Shipper draws is between what an agent does and what it takes to make an agent reliable. An agent can write a newsletter, generate podcast summaries, draft social posts. But someone has to set its context, review its output, catch the cases where it drifts from the original intent, and make the judgment calls the agent cannot make. The person doing that work is a human employee. At Every, more agents meant more people to supervise them.
Shipper also made a specific prediction about which roles benefit most: he believes PMs and designers will thrive in an agent-heavy organization because those roles involve the context-setting and judgment that agents cannot do themselves. That is a direct claim about the skills you already have, made by someone who has spent two years running the experiment.
3. AI coding tools and the work they create downstream
The productivity case for AI coding tools is real. Developers complete individual tasks faster when they have an assistant generating code. What is less often discussed is what happens to the rest of the software development cycle when the coding phase accelerates.
The GitLab 2025 DevSecOps survey, which surveyed 3,266 DevSecOps professionals worldwide, found what it called the “AI Paradox”: faster coding creates new bottlenecks in testing, security review, and cross-team coordination that cost teams close to a full workday per engineer each week. When a team ships pull requests faster, it generates more surface area to cover in quality assurance, more edge cases to catch in security review, and more integrations to coordinate across systems that were not designed to change at this pace. AI-assisted coding also tends to increase code churn, which DevOps researchers define as code that gets discarded within two weeks of being written. More code gets produced, and more of it turns out to be wrong.
Part of this is the nature of generated code. An AI assistant is optimizing for the task it was given. Long-term maintainability is a separate concern, and it does not show up in the model’s objective function. It will write code that passes the tests you have, leaving you to anticipate the tests you have not yet written. Someone has to think about that gap.
For a PM, this means more releases to coordinate, more tradeoffs to make about what ships now versus what waits, and more conversations with engineers about what “done” means for a given feature. The raw output rate of a development team goes up. The demand on product judgment goes up with it.
4. What the labor market data shows
The World Economic Forum’s Future of Jobs Report 2025 drew on a survey of over 1,000 leading employers and projected 170 million new roles created globally through 2030, against 92 million displaced. Net: 78 million more jobs.
The roles projected to grow are not exclusively technical. Frontline sales, customer service, and training roles appear alongside data analysts and software engineers. The reason is the same mechanism Jevons observed: more technology needs more people to sell it, support it, and explain it to customers who encounter it for the first time. The expansion of one productive input expands demand for the surrounding inputs.
The data is worth noting not because it is reassuring, but because it redirects the question. “Will AI eliminate my job?” already has an answer in the data. The better question is “what version of my job grows in an AI-heavy market?” For PMs and designers, the answer is the version that involves telling agents what to build and evaluating whether they got it right.
5. The oversight loop nobody budgeted for
Every AI agent in production creates a specific new category of work: checking whether the agent is still doing what you want it to do. This is not a temporary limitation that better models will solve. It is structural. An agent optimizes for the objective it was given at setup time, and the business context keeps changing. Someone has to catch the cases where the agent is no longer aligned with what the business actually wants.
Shipper’s framing is precise about this. The agent writes the thing. A person has to ensure the thing is working. Evaluating output is a different job than generating it. It requires knowing what good looks like in the first place, understanding the user well enough to catch the cases where the agent confused its training with the current context, and making the judgment call about when output is good enough to ship.
Those skills are not learned from a machine. They come from doing the PM or design work for long enough to have a model of what users actually want, as opposed to what the spec says or what the most recent user interview suggested. The teams that will get the most value from AI agents are not the teams that deploy the most of them. They are the teams with the judgment to evaluate what those agents produce.
The pattern and what it means this week
Across these five items, the same thing is happening. AI does not reduce the demand for human judgment. It shifts where that judgment is needed and raises the stakes of each individual decision. When a team ships more software faster, each product decision matters more because the cost of shipping the wrong thing at speed is higher than the cost of shipping it slowly.
This is good news if your skills involve figuring out what the user wants and whether the current output delivers it. That skill becomes more valuable as the output rate goes up.
The concrete action is to stop treating AI as a headcount substitution story and start treating it as a specification problem. The bottleneck is moving from “can we build this?” to “can we describe this clearly enough that an agent builds the right thing, and do we have someone good enough to catch when it does not?” That is your leverage point. The teams that recognize it first will have an advantage, and it is not a technical advantage.
References
| Source | Author / Org | Year | Supports |
|---|---|---|---|
| The AI paradox: More automation, more humans, more work | Shipper, Lenny’s Newsletter | 2026 | Dan Shipper’s Every headcount doubling and agent-oversight observation |
| A 160-year-old paradox explains why AI will create more lawyers and accountants | Fortune | 2026 | Jevons paradox applied to AI and the net-jobs-creation prediction |
| Future of Jobs Report 2025 | World Economic Forum | 2025 | WEF net +78M jobs projection through 2030 based on survey of 1,000+ employers |
| DevSecOps Survey | GitLab | 2025 | AI coding tools create downstream bottlenecks in testing, security review, and code churn |