AI Is Making Us More Capable. But Not More Coherent.

I have been using AI heavily enough now to notice that the constraint has moved.

For much of my professional life, there were useful things I simply could not do quickly enough. Research took time. A first draft had to be written. A prototype needed someone with the right skills. Comparing ten alternatives could consume an afternoon. Producing another version carried enough cost that, at some point, the work naturally stopped.

That cost has fallen dramatically. I can now research, draft, compare, restructure, prototype and iterate across domains that once required more time, more people or more specialized help. The work still requires judgment. Outputs still need checking. Taste still matters, as does knowing when the model is wrong. But capability is no longer the constraint it was.

What I did not expect was how quickly a different constraint would become visible. When another version is almost always available, deciding that the work is finished becomes part of the work. Another draft is possible. Another comparison. Another research branch. Another landing page. Another argument to test.

That was the question I kept returning to during DES Málaga: if AI gives us productive capability back, who gets it?

I did not arrive at the conference as someone looking at technology from the outside. I have spent more than two decades working in technology, advertising, entrepreneurship, sales, cybersecurity, digital systems and business transformation. I use AI every day. I remain more interested in what these tools make possible than in finding reasons to reject them.

DES itself was not framed as an anti-technology conversation either. Its 2026 theme, “Machines learn, people lead,” explicitly paired AI-driven efficiency with human judgment and leadership. Across sessions on work, talent, health, hospitality and enterprise transformation, the premise was that intelligent systems would increase what organizations could do while people remained responsible for direction.

That premise is reasonable. What felt less resolved was what happens after the capability increase arrives.

---

The Productivity Surplus

One of the more useful conversations I had at DES happened away from the stage with people from the local startup community. We were talking about AI and productivity, but not in the abstract. We were talking about the ordinary experience of being able to do things that previously required more time, more people or a different skill set.

A founder can research a market, draft an investor update, restructure a proposal, test positioning, build a rough prototype and compare several strategic options in the same day. A small team can cover more functional ground than it could a few years ago. An individual can move from question to usable first pass with remarkably little friction.

That is a real gain. The mistake is assuming the gain tells us what happens next.

Productivity is often discussed as though the benefit lives inside the number. A task takes two hours, then forty minutes, so productivity improved. That part is straightforward. What the metric does not tell us is who now controls the eighty minutes that became available.

The employee may finish earlier. The manager may add another responsibility. The firm may reduce staffing. The team may spend more time checking quality. A founder may start another project. A customer may receive the same service faster. The company may keep the same workload and capture the gain as margin. All of those are plausible uses of the same efficiency gain. AI changes the cost of producing the work. It does not decide how the difference will be allocated. Efficiency creates a surplus. Incentives allocate it.

Early evidence makes this question more interesting, not less. A large randomized field experiment across 66 firms and 7,137 knowledge workers found that employees who actually used a generative AI tool spent about two fewer hours per week on email and reduced work outside regular hours. The researchers did not detect a corresponding increase in the quantity or composition of tasks from individual AI access.

If I were trying only to argue that AI inevitably intensifies work, that finding would be inconvenient. It is more useful than that. It shows that saved time can remain with the worker, at least under some conditions.

At a wider level, the destination of the gain is still unsettled. A 2026 review from the International Labour Organization finds real but uneven productivity gains from generative AI, while worker-reported time savings have not yet translated cleanly into higher measured output, earnings or employment. The same review points to work organization, autonomy and job quality as unresolved parts of the transition. In other words, AI can create a surplus without determining who captures it.

The software removes some cost. The organization still decides what counts as enough.

---

When The Benchmark Moves

This is not a problem unique to AI. Technologies that make work faster often change the baseline of what organizations consider normal. Spreadsheets changed financial work. Email changed communication. Search engines changed research. Cloud software changed collaboration. Mobile connectivity changed expectations of reachability.

Once a task reliably takes less time, estimates can change. Turnaround expectations can change. Staffing assumptions can change. What counted as unusually fast can become the new baseline.

AI may accelerate that process because it affects many cognitive tasks at once. Writing, summarizing, coding, analysis, research, translation, ideation, documentation and customer support can all become cheaper in the same organization, although the gains vary greatly by task and user.

The important unit, then, is not simply minutes saved on a task. It is what happens to the benchmark after enough people save those minutes.

This is where individual experience and organizational economics can diverge. A worker may genuinely feel that AI makes a task easier while also discovering that the organization now expects more tasks. Both observations can be true. The tool improved the local process. The surrounding system changed what counted as an acceptable amount of work.

The same can happen to a founder without any manager imposing it. More capability creates more visible possibility. A project that would once have been dismissed because it required another hire or several days of work can now look feasible before lunch. The bottleneck shifts from execution toward selection.

My own version of this is almost embarrassingly practical. I rarely need help finding the next thing AI could do. I need to decide whether the next thing deserves to be done.

Earlier tools often stopped because I reached the edge of time, skill or available help. Generative AI moves several of those edges outward. The benefit is obvious. The cost is that a constraint that once arrived from the tool increasingly has to come from judgment.

That is different from saying AI causes overwork. It does not. It means one source of stopping friction has weakened, which makes the quality of the surrounding choices more consequential.

---

Capability Is External

This distinction has become important enough that I now use two words more carefully.

AI can expand capability very quickly. It can increase what I can produce, compare, simulate or learn with the resources immediately available to me.

That does not mean my own available hours, attention or judgment expand at the same rate.

The difference matters because “capacity” is often used casually to describe whatever output becomes possible after a new tool arrives. But the tool may be carrying part of that capability externally. I can produce more without having more hours in the day. I can explore more options without becoming proportionally better at deciding which ones matter. I can generate another version after my judgment has already become less useful.

None of this diminishes the capability gain. It explains why the gain can create a new allocation problem.

A technology can increase what is possible faster than a person or organization improves its ability to decide what is worth doing.

---

Human-Centered After The Interface

The language of human-centered AI appeared repeatedly at DES, and rightly so. Better usability, appropriate oversight, transparency, safety, explainability where it matters, meaningful control and alignment with human goals are all legitimate concerns. But work introduces a second question that interface design cannot answer.

Suppose the AI is intuitive, accurate, trusted and genuinely useful. It cuts a recurring task from two hours to forty minutes. From a product perspective, that may be an excellent result. The question of what happens to the remaining eighty minutes belongs somewhere else.

If the worker leaves earlier, the gain has one distribution. If another task fills the space, it has another. If the team handles more customers with the same staff, another. If staffing is cut because the same output now requires fewer hours, another again.

The technology can be human-centered in its design while the work system captures the benefit in a way the worker experiences as intensification. Those are separate layers of the problem.

This is why “human in the loop” is useful but insufficient as a description of good work. A human can remain technically present while the surrounding process changes faster than their authority over it. Human oversight matters. So does who can contest a decision, how much context they receive, whether responsibility matches actual control, and what performance assumptions change after the tool is adopted.

That is not an argument for preserving inefficient work to protect people from technology. Removing pointless effort is good. The question is whether removing effort automatically improves the job.

Sometimes it will. Sometimes the same gain will be converted into greater throughput, lower cost or fewer people. The technology does not resolve the trade.

---

The Service Example

Service work makes the allocation problem easy to see.

Imagine a hotel where AI removes much of the administrative work around scheduling, guest requests and routine communication. That could give employees more time with guests and reduce repetitive tasks. It could also become a reason to cover more rooms with fewer people.

Both outcomes can be described as productivity improvements. They are not the same experience of work.

The same distinction applies in customer support, healthcare administration, sales operations and professional services. Automation can remove work nobody valued doing. It can also move the performance target as soon as the old friction disappears.

This is why employee experience cannot be inferred from process efficiency. A faster system may genuinely reduce aggravation. It may also increase the amount of activity expected during the same shift. You have to look at what changed around the person, not only at what changed inside the workflow.

Skills do not settle the allocation question

Future-of-work conversations often move quickly from AI exposure to reskilling. There is good reason for that. Roles are changing, some tasks are becoming easier to automate, and workers will need to learn new tools and move across functions more often. Training is part of the response. It is not the whole response.

A person can learn the new tool and still face a job whose expectations changed faster than its rewards. A company can invest seriously in reskilling while using the resulting productivity gain primarily to increase output. A worker can become more employable while also absorbing more uncertainty about how long the current role will exist.

These are not arguments against learning. They are reasons to separate skill acquisition from the distribution of technological gains.

The ILO’s current evidence review is useful here because it does not collapse the transition into either mass displacement or effortless augmentation. Productivity effects vary by context, and questions of autonomy, coordination and job quality remain open. That is a more credible picture than either extreme.

The relevant workplace question is therefore not simply whether employees can use AI. It is what changes in the job after they can.

---

Health Tech Reveals The Same Allocation Problem

Digital health conversations at DES offered a different version of the same structure. AI can make signals easier to detect, clinical information easier to organize and workflows easier to coordinate. Those are meaningful capabilities. But a better signal still enters an existing institution and an existing life.

A risk flag may lead to earlier treatment. It may reduce clinician workload. It may create additional follow-up. It may generate anxiety without a clear next step. It may help a patient understand something important. It may become one more item in a dashboard that nobody has time to interpret well.

Those possibilities are not interchangeable, and the technology alone does not decide among them.

That is as far as I want to take the health argument here. Earlier detection and prevention are different questions, and they deserve their own treatment. What matters for this essay is the same allocation problem: information, time and capability become available, then a human system decides what can actually be done with them.

Coherence is not another productivity metric

The title of this essay uses a word that can become vague quickly, so I mean something modest by coherence here.

Does an improvement in one part of the arrangement make the whole easier to inhabit, or does it mostly move the cost somewhere else?

AI can make a task faster while increasing expected task volume. It can make communication easier while raising expectations of responsiveness. It can make research cheaper while multiplying how many questions now seem worth pursuing. It can reduce administrative work while making a leaner staffing model financially attractive.

The capability gain is real in every case. The larger arrangement does not automatically improve with it.

That is why I am cautious when productivity and progress are treated as synonyms. Productivity is a relationship between inputs and outputs. It can tell us that fewer resources were required to produce something, or that more output became possible with the same resources.

It cannot tell us by itself whether the saved resource improved the life or work of the person who helped create the gain.

This is also where the Human OS becomes relevant without needing to turn AI into a biological theory. The Manual’s broader concern is what happens when improvements in one part of a life or system transfer cost somewhere else. AI gives us a particularly visible contemporary example because the capability gain is arriving so quickly.

The technology can be excellent and the allocation can still be poor.

---

What DES Clarified

DES did not convince me that technology companies are ignoring the human side of AI. The conference was explicitly trying to put people, leadership, talent, trust and health inside the transformation conversation. The unresolved part is harder because it is not primarily a technology problem.

If AI saves ten hours across a team, there is no neutral answer to what those ten hours should become. More output may be the right answer. Better service may be the right answer. Lower cost may keep a company alive. Shorter hours may improve retention. More time for judgment may reduce expensive mistakes. A smaller team may be economically unavoidable.

Those are management, labor, market and ownership questions. They involve incentives and trade-offs that cannot be solved by making the model more capable.

The easiest way to evaluate AI is to look at what became possible after it arrived. The harder question is what happened to the gain.

Did a worker finish earlier? Did the customer receive better service? Did the company produce more with the same team? Did the team shrink? Did quality improve? Did another project appear because there was suddenly room for one? Did the benchmark quietly move? None of those outcomes is contained in the model.

AI changes the cost of doing the work. People, organizations and markets decide what to do with the difference. That is why capability is only part of the story.

A technology can be extraordinarily useful while leaving the larger arrangement no easier to inhabit. It can even make an old incentive structure more efficient.

---

The question is not only what AI lets us do. It is who gets the time, attention and possibility that the tool gives back.See how the Human OS fits together →If this helped you name something you have been feeling but had not quite put into words, please share it with someone who might need the same language.Share

---

The Human OS Manual is now . If this work resonates, you can reserve your copy and follow the ideas here as they continue to develop.

Feel free to leave a comment. I read and respond to every thoughtful note.

---

Notes and sources

Some observations in this essay come from my field notes at DES Málaga 2026, including sessions and hallway conversations. External references below verify event context and support the broader claims around AI productivity, work organization, future skills, and digital health.

• Digital Enterprise Show, “Official event page” (2026).
• Digital Enterprise Show, “The future of work in the age of AI: Talent, business, and transformation” (2026).
• Digital Enterprise Show, “Digital Health: Innovate, heal, transform” (2026).
• Dillon, Jaffe, Immorlica, and Stanton, “Shifting Work Patterns with Generative AI” (NBER Working Paper 33795, revised 2025).
• International Labour Organization, “The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence” (2026).