The more proficient users may not recognize how much skepticism is accumulating around them. Their own experience tells them that the technology is becoming more useful. They may therefore interpret disappointing results as a temporary lack of skill rather than evidence of a broader adoption problem.

Both experiences can be valid.

AI can be extraordinarily powerful in the hands of someone who has learned how to work with it, while still proving frustrating or uneconomical for someone encountering the same system under different conditions.

Eventually AI has to produce real, repeatable value. Where that value is difficult for others to replicate, it can begin contributing to competitive advantage. Without that connection, the current level of excitement will be difficult to sustain, even if the underlying technology continues to progress.

The difficulty is deciding where durable value actually lies.

Part of what makes this wave of AI different is that we are increasingly building systems around non-deterministic behavior. We know these systems can be extraordinarily powerful, for beneficial and harmful purposes. We are still working out which capabilities are genuinely unprecedented and which applications merely use generative AI to perform something that could be done more reliably or economically through deterministic software.

Consider two hypothetical applications.

Use A relies on generative AI today, but the same outcome could eventually be reproduced using a deterministic system. Its value may be sensitive to changes in token costs, model pricing, reliability requirements or the availability of a simpler technical alternative.

Use B depends on a capability that is uniquely enabled by generative AI. Its economics may also change, but its underlying value could prove more durable because the application cannot easily be reproduced through conventional software.

This distinction is important, but it is not static.

Models improve. Architectures change. Costs move. New capabilities emerge. An application that makes economic sense today may not tomorrow, while something uneconomical today may become compelling sooner than expected.

There is a hypothesis worth investigating: as AI applications move from defensive uses toward more offensive value propositions, the distinctive characteristics of generative AI may become more important to the value being created. Defensive applications often improve or protect something that already exists. Offensive applications are more likely to depend on discovery, simulation, creation or forms of interaction that were previously impractical.

That hypothesis may prove wrong. Too many parts of the decision space are still moving for it to be treated as a settled principle.

Following every new development is not enough. What is needed is a robust mental model that helps us distinguish among three things:

• what is genuinely new;
• what remains fundamentally unchanged; and
• what consequently needs to be rethought.

Those distinctions should lead to operating principles that remain useful even as individual models, products and cost structures change.

The test is whether those principles help individuals and organizations move from experimentation toward repeatable value and, eventually, competitive advantage.

That is the territory I want to explore: bringing greater clarity to what matters, determining which operating principles need to change in the age of AI, and connecting those principles to value that can survive beyond the current noise.

There is also a subtler problem. Small discrepancies accumulate.

An intention is slightly misunderstood. Something already corrected reappears. A distinction is flattened. Context gradually drifts. Fixing one issue introduces another. The interaction begins to resemble a game of Whac-A-Mole.

No single discrepancy needs to be catastrophic. The AI may still produce something genuinely useful. But the distance between the user's intention and the work being produced increases gradually.

Maintaining complete control becomes increasingly expensive.

Eventually, the user may allow the AI to finish a cycle and accept the result as good enough. This can be entirely rational. The output may still be valuable, even though it no longer reflects every question, correction or nuance that arose along the way.

The gap between usefulness and intentional control is one of the first impediments that appears when AI moves from being a toy to becoming an actual tool.

More experienced users learn to manage this problem. They structure the task, preserve important context, divide complex work into parts, constrain the output and restore decisions when the AI begins to drift.

Managing the context can itself prove valuable. Properly framing the problem, arranging the relevant information and making the intended outcome explicit often clarifies the user's own thinking. Sometimes that clarification provides much of the value, even if the AI-assisted work is not taken through to completion.

The last mile nevertheless remains difficult.

AI can move remarkably quickly from nothing to something impressive. Moving from impressive to genuinely finished requires verification, reconciliation and continued control over the original intention. The more complex the task becomes, the more difficult that final movement can be.

This produces an apparently contradictory experience. A person may simultaneously believe that AI is extraordinarily powerful and that using it for a particular task is not worth the total effort. Both conclusions may be reasonable.

The frame expands further when AI is introduced into teams and organizations.

AI can accelerate analysis, produce options or complete part of a process. It does not follow that the surrounding organization will act at the same speed.

People deliberate. They persuade others. They seek alignment, exercise judgment, hesitate and sometimes procrastinate. They may wait because something does not feel right, even when an apparently good answer is already available.

These activities did not newly appear because AI accelerated the preceding work. They were already part of the system. Faster production simply makes their independent pace more visible.

Some of this is friction. Some of it is necessary.

Decisions carry accountability. Organizational action often requires several people to develop sufficient confidence in a course of action. Gut feeling, reflection and hesitation can prevent mistakes as well as cause delay.

The objective cannot therefore be to force every part of the human system to operate at machine speed.

The more relevant question is whether the time released in one activity can become usable capacity somewhere else.

An employee may finish an analysis earlier without the organization making the corresponding decision earlier. A team may gain access to more options without having the attention needed to consider them. A process may accelerate without the rules, incentives or authority around it changing.

The potential gain exists, but it cannot automatically be claimed.

This is where the idea of a simple bottleneck becomes inadequate. AI does not merely remove constraints. It changes the relationship among them, exposes constraints that were previously less visible and may create new forms of friction around understanding, intention and control.

There is also a deeper human question.

As more cognitive work is delegated, what happens to the capabilities through which people learn to exercise judgment?

The concern is particularly visible for people entering a profession. If AI increasingly produces the work and the human role becomes approval or criticism, the person may lose access to the experiences through which practical expertise was traditionally developed.

The issue is not limited to inexperienced people. Experts can also stop exercising capabilities that they increasingly delegate. Decision-making muscles can atrophy even when the quality or volume of output increases.

Beyond capability, there is the question of agency and accomplishment.

We may become able to accomplish far more while experiencing less of what it means to have accomplished something ourselves.

These impediments rarely arrive as one dramatic failure. They accumulate more like small debts:

• loss of intention;
• context-management effort;
• verification burden;
• unfinished last-mile work;
• organizational hesitation;
• unrealized saved time; and
• uncertainty about learning, judgment and agency.

Together, they weaken trust.

People become more skeptical of every new development. They experiment less. They are less willing to delegate work that matters. The value they receive remains limited, which reinforces the original skepticism.

The result is a vicious cycle in which AI's theoretical ability to release bandwidth does not become actual bandwidth for individuals or organizations.

This suggests a different way to assess AI-enabled work. Instead of beginning with where AI can save the most time, begin with what the individual or organization most needs to accomplish now.

A company can be viewed as performing two activities in a continuing cycle. It identifies the right issue or opportunity, then solves it through a product or service. The outcome produces new information, which shapes the next issue to be examined.

Within that cycle, three modes of activity run across the value streams of the company: business decision-making, customer value creation and operational optimization.

The most valuable use of AI will depend on which of these activities deserves attention at a particular moment. A company may need to reconsider where it should compete in the age of AI. It may need to develop a solution and bring it to market before competitors. It may need to remove an operational constraint that is preventing an otherwise sound proposition from reaching customers.

The relevant unit of value is therefore not always an hour saved. It may be a better question examined, a decision made with stronger evidence, a product developed sooner, or an operating model changed before the need becomes obvious to everyone else.

This website is itself a small experiment in that idea. Building the site requires specification, copy, layout, coding and repeated quality control. Not all of those activities require the same contribution from me.

When the collaboration works, AI carries more of the mechanical work of producing and maintaining the prototype. That allows me to concentrate on the part that cannot simply be delegated: examining my experience, deciding what I believe and finding language for the ideas that are distinctly mine.

The value is that a greater share of my attention can be directed toward the work for which my judgment is most important.

That question connects AI capability to decision quality. It also creates a bridge from individual productivity toward customer value, organizational effectiveness and competitive advantage.

Although the list is not exhaustive, six value propositions help make this landscape visible:

1. Protect the business from the expanded risk surface and vulnerabilities introduced by AI.
2. Manage AI economics across employee effort, token consumption, models and infrastructure.
3. Extract relevant and trustworthy signal from growing volumes of synthetic, false and low-value information.
4. Use AI agents to discover third-party actors and collaborate with them on the organization's behalf.
5. Use AI to simulate possible outcomes and improve the basis for a decision.
6. Use AI to reveal new relationships in knowledge that may lead to product and business innovation.

Two competent practitioners working on token economics may face a difficult problem, but once the operating conditions and constraints are defined, their approaches may converge. The work remains important, while the response is comparatively more standardized.

Knowledge discovery is different. AI may reveal many relationships, but a person with relevant experience and context must still recognize which one is surprising, commercially important or worth pursuing. The technology surfaces possibilities. Human expertise and sense determine their meaning.

The same is true of simulation. AI may generate scenarios or estimate outcomes, but the user determines which variables deserve inclusion, which assumptions are credible, what patterns to examine and how the result should affect a decision.

Signal extraction is particularly interesting. It sits closer to the defensive side because it responds to information complexity created or intensified by AI. Yet deciding what constitutes signal can require substantial context, judgment and sense. The propositions do not therefore fall neatly on a single line.

A more useful representation uses two axes. The horizontal axis moves from responding to conditions created by AI toward expanding business possibility. The vertical axis moves from a more standardized response toward greater dependence on human context, experience and sense.

The position of each proposition remains a hypothesis. It will move according to the organization, application, available data, economics and maturity of the technology.

This also connects back to the concern about skill atrophy. If consequential AI applications depend on strong human judgment, preserving and developing that judgment becomes an economic requirement.

AI advantage is relative and transient.

An organization may create substantial value by introducing a new capability, improving a process or reaching the last mile more reliably. But if competitors can acquire the same models, adopt similar practices and achieve comparable results, that capability gradually stops being a source of differentiation. It becomes part of the expected baseline.

This is likely to happen repeatedly. AI proficiency will spread. Techniques that currently require specialist knowledge will become easier to use. Tools will absorb practices that users now have to perform manually. Workflows that appear advanced today will eventually be packaged, standardized and made widely available.

Organizations therefore face two related questions: What creates advantage now? What will remain differentiating as the baseline rises?

The second question is harder because it cannot be answered by looking only at current AI capabilities.

A promising place to look is in the interaction between technology and the organization using it.

What we notice influences which problems and opportunities receive attention before they become obvious to everyone else.

What we know includes the experience, context, relationships and accumulated understanding that determine what an organization can see in information available to many.

How well we judge affects which possibilities are credible, which trade-offs are acceptable and which decisions are worth acting upon.

What we learn determines whether experience compounds or repeatedly disappears into individuals, documents and disconnected systems.

The environment we create for people influences whether individuals develop judgment, exercise agency, contribute their distinctive capabilities and choose to remain within the organization.

None of these is automatically an advantage. They matter when they translate into better outcomes, including products that customers prefer, decisions that create greater economic value, capabilities that competitors struggle to reproduce, and an organization in which capable people want to work and develop.

The sources of advantage may also reinforce one another. Better attention can lead to better learning. Better learning can strengthen judgment. Better judgment can improve decisions. An organization that develops people well may become better at noticing, learning and adapting.

This does not produce a fixed formula for competitive advantage. The relevant combination will vary by organization, market and moment.

That leads to the final question of the thesis: How should organizations combine evolving AI capability with the human qualities on which differentiated value increasingly depends?

AI will change more than the speed at which work gets done. As capabilities improve, organizations will need to rethink how people and AI work together, how decisions are made, how knowledge accumulates, and how people continue to develop experience and judgment.

The answers will differ across organizations. But several questions are becoming increasingly important:

• How can AI improve decision-making without weakening the judgment needed to make good decisions?
• How can an organization turn its experience and knowledge into learning that accumulates?
• How should operating models change as the division of work between people and AI changes?
• Where can AI enable genuinely new products, services and business possibilities?
• How should organizations manage the economics, risk and governance that accompany greater AI use?
• How do we design human + AI work so that people continue to develop capability, agency and meaning?

These questions are connected. Progress in one area will often expose questions in another, and some of our current assumptions will change as the technology and our experience develop.

The thesis establishes a point of view. The landscape is where the thinking continues.

Explore the landscape →