When Using More Became a Virtue

For a brief period, consuming AI tokens began to look like a professional virtue. Companies announced their intention to become AI-native, employees were encouraged to use artificial intelligence throughout the working day, and internal dashboards sometimes treated usage as evidence of adaptation. The more prompts people submitted and the more agents they operated, the more convincingly they appeared to belong to the future.

The enthusiasm was understandable. Generative AI had moved beyond the familiar chatbot. Tools such as Claude Code could inspect local files, understand a codebase, modify documents, execute commands, and continue working across several steps. For developers in particular, this represented a genuine change. Instead of copying fragments of code into a conversation and transferring the answer back by hand, they could invite an agent into the working environment itself.

A larger agentic ecosystem soon developed around this capability. AI could search, summarize, draft, format, classify, monitor, and coordinate. It could read messages, prepare reports, generate slide decks, revise files, and operate connected applications. The image of AI as an assistant became more literal. Users no longer interacted with one model through a sequence of questions. They began assigning work to networks of agents that appeared able to manage entire processes.

Around these tools grew an industry devoted to explaining how to use them. Every model release produced new videos, articles, courses, prompt collections, and productivity systems. Prompt engineering gave way to context engineering, agent orchestration, looping techniques, skills, plugins, memory systems, and frameworks whose names changed faster than many people could learn them. Professional anxiety reinforced the excitement. Those who did not follow each development risked feeling that they were already behind.

The result was a form of AI intoxication. The term does not suggest that the technology lacked value. Intoxication begins with something real. AI was capable of remarkable work, and many organizations had good reason to explore it. Yet excitement changed into pressure when people were expected to demonstrate AI usage regardless of whether it improved the task before them. A tool that could expand human capacity also became a symbol of loyalty to innovation.

Once usage itself acquired organizational value, asking whether AI was needed became difficult. Reading an email personally could appear old-fashioned when an agent was available. Writing a short message without assistance might feel like a failure to adopt. A report produced through direct thinking could seem less advanced than one assembled by several agents, even when the agentic process took longer and generated more material to review.

An AI-native organization was increasingly imagined as one in which AI was present everywhere. That definition contained the seeds of the reversal that followed.

The Meter That Became the Destination

AI services do not operate without cost. Every large context, extended reasoning process, tool call, rendered document, generated image, and agentic loop consumes computing capacity. A single experiment may appear inexpensive, but the economics change when thousands of employees run similar processes each day. Long-running agents can spend resources at a pace no human worker could reproduce through manual prompting.

Companies eventually saw the bills. Some discovered that employees had little awareness of how much their activities cost. Others found that the dashboards intended to encourage adoption had turned consumption into a competition. Agents could be instructed to perform unnecessary work, repeat operations, or remain active for long periods. High token usage looked impressive on a leaderboard even when it produced little that anyone needed.

Token consumption had been introduced as a proxy for AI adoption. Once it became a target, it stopped providing reliable evidence of progress. A careful employee who used a small model to complete a well-defined task could appear less advanced than someone who loaded an entire repository into a frontier model and asked it to solve an undefined problem. The less disciplined workflow received the higher score.

The reversal was swift. Organizations that had encouraged employees to maximize their usage began imposing strict allowances and asking them to minimize it. A generous monthly budget could become a tightly controlled allocation within one planning cycle. Employees who had been told that heavy AI usage demonstrated commitment were suddenly expected to ration the same activity.

The comedy lies in the continuity between the two policies. Token maximizing and token minimizing look like opposites, but both confuse an input with an outcome. One celebrates consumption, while the other treats consumption as failure. Neither tells us whether the work became more accurate, whether a decision improved, whether customers received better service, or whether employees gained time for more valuable responsibilities.

Imagine a company that evaluates delivery drivers according to the fuel they burn. During the first phase, the driver who consumes the most fuel is praised as the most active. Engines are left running, routes become longer, and large vehicles are used for small deliveries. When the fuel bill arrives, management introduces the opposite rule and rewards drivers for consuming as little as possible. Deliveries are delayed because nobody wants to use the vehicle. Throughout both phases, the company neglects the cargo, the destination, and the quality of the service.

Tokens are fuel, not delivered value. They are a necessary cost of using model intelligence, but their quantity does not reveal the worth of the result.

Anthropic’s rise illustrates the economic forces behind this period. Its focus on enterprise customers and developer tools placed it at the center of corporate AI adoption. Usage-based enterprise arrangements allowed organizations to expand without fixed individual limits, but every additional token carried a cost. This was not hidden, and Anthropic also provided tools for monitoring and controlling usage. Even so, the commercial structure rewarded expanding consumption at the same time that corporate leaders were encouraging employees to consume.

Calling the arrangement deceptive goes further than the available evidence supports. The stronger criticism concerns the interaction between vendor incentives and managerial fashion. AI companies benefited when customers used more intelligence, while corporate leaders benefited from appearing committed to transformation. Employees responded to the signals they received. Each participant acted according to an understandable incentive, but together they created an expensive form of enthusiasm.

The public drama surrounding advanced models added another layer. Claims about unprecedented intelligence, cybersecurity capability, and potential danger created an atmosphere in which access to the latest model seemed strategically urgent. The 2026 dispute over Anthropic’s Fable 5 and Mythos 5 showed how safety, politics, regulation, and competitive positioning could become entangled. Anthropic did not voluntarily withdraw the models after admitting that they were uniquely dangerous. The suspension followed a government directive that the company disputed, and access to Fable 5 was later restored. Still, the episode demonstrated how narratives of exceptional capability can magnify both fear and demand.

Safety concerns may be sincere and technically justified while also contributing to a company’s prestige. A model described as powerful enough to trouble governments becomes more desirable to businesses seeking an advantage. Ethical concern and market value do not always oppose each other. At times they grow from the same story.

AI-Native Without AI Everywhere

After the excitement and financial correction, the phrase “AI-native” needs a more durable meaning. An AI-native company is not one in which every action passes through a model. It is an organization whose information, workflows, and responsibilities have been designed with machine intelligence available as one form of infrastructure.

Electricity transformed industry, but an electricity-native factory would not measure progress by attaching a motor to every object. Digital technology changed business, but a digitally mature company does not judge employees by the number of applications they open. In the same way, artificial intelligence becomes valuable when it is placed where intelligence, interpretation, adaptation, or scale is needed.

Some tasks remain best handled by deterministic software. A calculator is more reliable than a language model for arithmetic. A database query may retrieve a known record more efficiently than an agent searching through documents. A template can generate a standard message without asking a frontier model to compose it anew. The availability of a more powerful tool does not make established tools obsolete.

AI becomes valuable under several different conditions. It can reduce cognitive burden when a person must interpret large amounts of information. It can support intellectually demanding work such as research synthesis, complex writing, programming, and analysis. It can also perform uncomplicated work at a scale that would overwhelm human workers, such as classifying thousands of messages or extracting information from a large document collection.

These uses cannot be judged by complexity alone. Reading one email may require little effort, while reviewing fifty thousand customer messages creates a problem of scale. Formatting one short document may not justify an agentic workflow, while converting hundreds of reports into a consistent structure may. The relevant variables include volume, uncertainty, repetition, risk, time, and the cost of verifying the result.

Delegating work to AI should not be treated as a moral weakness. Human beings have always created tools to reduce labor. There is no virtue in preserving tedium for its own sake. The concern begins when delegation consumes more resources than the task requires, removes human judgment from consequential decisions, or creates additional work that is mistaken for productivity.

A mature organization therefore routes work according to its needs. Conventional software handles stable and deterministic operations. Smaller models process routine transformations, classifications, and summaries. Frontier models are reserved for tasks requiring extensive context, stronger reasoning, or complex tool use. Agents manage multi-step work when the benefits of coordination exceed the cost of supervision.

This approach also changes the meaning of innovation. Claude Code was influential because it showed how an AI system could work within a real environment rather than wait inside a chat window. Other platforms soon developed comparable abilities. ChatGPT Work, Codex, and a growing range of agentic systems can now inspect files, create artifacts, operate tools, and sustain longer processes.

Once a capability spreads across providers, possession of the tool no longer creates a lasting advantage. The advantage moves into workflow design, information quality, evaluation, governance, and judgment. The central organizational question is no longer, “Do our employees have AI agents?” It becomes, “Do we know which work should be assigned to them, what information they should receive, and how their contribution will be evaluated?”

AI maturity begins when access stops being confused with competence.

Giving Intelligence the Right Material

Many discussions about AI efficiency concentrate on prompts. Users are taught to refine instructions, compress context, divide tasks into stages, or select a less expensive model. These techniques can help, but they address only one part of the system. The structure of the underlying information often has a greater effect on cost and reliability.

Most workplace knowledge lives inside containers designed for human interaction. Word documents present pages. PowerPoint organizes slides. Excel displays cells, charts, formulas, and visual relationships. Email arranges communication around senders, recipients, threads, and inboxes. PDF preserves a finished appearance across devices.

These containers serve legitimate purposes. People need documents they can review, presentations they can deliver, spreadsheets they can calculate with, and reports whose layout remains stable. Trouble arises when the finished container becomes the only available form of the knowledge it contains.

A language model working with plain text can read headings, paragraphs, lists, and links directly. Markdown makes much of that structure visible without requiring a specialized application. CSV and structured data formats expose relationships that might otherwise be buried inside a visual table. A well-organized repository allows both humans and machines to locate sources, understand versions, and trace how a result was produced.

Word, Excel, and PowerPoint files are not automatically expensive because they are binary applications. Modern Office files contain structured components that software can often extract efficiently. The larger expense appears when the task requires the AI system to preserve complex formatting, interpret charts, examine embedded objects, render the file visually, compare the rendered result with the source, and repeat the process after each revision.

The difference becomes clear in document work. Revising the wording of a report may require little more than extracting its text and returning the edited content. Producing a publication-ready PDF in another language can involve changed line breaks, expanded headings, altered table dimensions, font substitution, image placement, page rendering, and repeated visual inspection. The intellectual content may remain stable while the container demands most of the labor.

A mature workflow separates these layers where possible. Ideas, evidence, structure, and canonical wording remain in portable, machine-readable forms. Human-facing formats are generated when the work reaches the stage at which presentation matters. The content can then survive beyond a particular application, while the container performs its proper role as an interface.

This principle should not become another absolute rule. A spreadsheet may be the correct source of truth when formulas, dependencies, and recalculation are part of the knowledge. Comments and tracked changes in a Word document may carry essential editorial history. A slide deck may contain visual reasoning that cannot be reduced to prose without loss. Maturity requires recognizing what belongs to the content and what belongs to its presentation.

Email presents a similar question. Asking a frontier model to monitor a modest personal inbox may consume resources without producing much benefit. A person can often recognize the relevant message faster than an agent can retrieve, interpret, summarize, and explain it. In a large operational mailbox, however, AI may identify urgent cases, group related requests, detect suspicious patterns, and reduce hours of repetitive review.

No container is inherently appropriate or inappropriate for AI. The decision depends on the amount of material, the structure available to the model, the consequence of mistakes, and the value of the assistance. Efficient AI use begins before the prompt. It begins with the architecture of knowledge.

Maturity Learned Through Excess

It is tempting to remember token maxing as an embarrassing episode that sensible organizations should have avoided. That judgment overlooks the role of experimentation in learning. The boundaries of a new technology cannot be understood entirely through planning. People must use it, misuse it, overestimate it, encounter its limitations, and discover where it produces value.

The early internet followed a comparable pattern. Organizations rushed to create websites before they understood what a website should accomplish. Companies placed “digital” in their strategies without changing how they worked. New platforms inspired exaggerated predictions, failed experiments, inflated valuations, and costly infrastructure. Yet the excess also trained a generation to think differently about communication, publishing, commerce, and information.

AI intoxication has already produced useful knowledge. Organizations have learned that agents can complete substantial work but require supervision. They have discovered that long contexts and repeated tool calls carry real costs. Employees have learned that a polished output may conceal weak reasoning, while an unremarkable text file may preserve far more reusable value. Developers have seen that an agent can accelerate coding while also introducing errors across a codebase with remarkable speed.

Such lessons are expensive, but they are not worthless. They become valuable when individuals and organizations are willing to remember them accurately. Without reflection, disappointment produces only another swing in fashion. Token maximizing becomes token minimizing, unrestricted experimentation becomes rigid control, and enthusiasm becomes cynicism.

Recording the experience interrupts that movement. What did we believe the technology would accomplish? Which expectations came from direct evidence, and which came from marketing or social pressure? What did we attempt? How much human review remained necessary? Which workflows continued after the excitement faded? What knowledge became reusable?

These questions create institutional memory. They also protect individuals from rewriting their own histories. Once an experiment fails, people often remember that they were skeptical from the beginning. When a technology succeeds, they may remember that they recognized its importance before everyone else. A written record preserves the uncertainty, enthusiasm, resistance, and error that accompanied the actual learning process.

Innovation will continue to disturb whatever balance we achieve. A new model will perform a task that previous systems could not handle. A new agent will cross another boundary between conversation and action. A new interface will make yesterday’s complicated workflow appear unnecessary. Each advance will renew the feeling that everything has changed and that immediate adoption is essential.

AI maturity cannot provide permanent immunity from that feeling. Mature users can still become excited, and mature organizations can still make mistakes. Their strength lies in recognizing the pattern earlier. They ask better questions, establish baselines, limit the scale of initial experiments, and define what evidence would justify expansion.

Maturity is therefore less like reaching a stable destination and more like developing the capacity to recover. The intoxication may return, but it does not have to govern the organization for as long. Experience shortens the distance between enthusiasm and reflection.

Intelligence Looking Back at Its Own Use

Artificial intelligence can participate in that reflection. The same system that helps us write, analyze, code, or organize information can also help us examine how we are using it. This may be called AI-assisted metacognition, but “reflexive AI practice” describes the broader discipline. We use AI, then turn it back upon the relationship and ask what is happening.

An AI system can examine a workflow and identify unnecessary complexity. It can compare the cost of a frontier agent with a smaller model or conventional program. It can ask whether the task requires interpretation, whether its output can be verified, and whether the process is generating more material than anyone will use. It can also notice patterns across conversations that an individual may overlook.

This creates a form of personalized guidance that generic AI education cannot provide. A video promising ten essential prompts must speak to an imagined average user. It does not know whether the viewer writes research papers, manages customer requests, develops software, translates technical content, or prepares financial reports. Advice designed for everyone often encourages people to adopt techniques unrelated to their actual needs.

AI can respond to the work in front of us. A user can describe a recurring task, show the current workflow, explain the constraints, and ask whether a different approach would be more efficient. The model can recommend a structured source format, a smaller context, a staged process, or no AI at all. It can explain why an agentic system is justified in one case and excessive in another.

The reflective questions should extend beyond financial efficiency. Has AI reduced cognitive burden or weakened an ability that the user still wants to preserve? Is it supporting thought or replacing the moment when thought would otherwise occur? Does the final work express the user’s judgment, or has the user become an editor of material whose direction was chosen elsewhere? Are more reports and presentations being produced because they are useful, or because production has become easier?

AI can help articulate these questions, but it cannot serve as the sole judge of its own value. Models are built within the same commercial and cultural environment that promotes their adoption. They may accommodate the user’s framing, repeat current industry assumptions, or provide an elegant justification for an unnecessarily complicated process. If asked whether a sophisticated AI workflow is possible, they are more likely to explain how to build it than to question whether it should exist.

Reflexive AI practice therefore needs a third reference point. We use AI to support the work. We use AI again to examine that partnership. We then compare both with external reality.

Reality includes the time saved after human review, the quality of the accepted output, the financial cost, the errors introduced, and the knowledge that remains reusable. It includes feedback from colleagues, customers, readers, and specialists who understand the consequences of the work. It also includes our condition when we occasionally work without AI. Can we still formulate a difficult thought, write a personal message, recognize a weak argument, and remain present during an unresolved problem?

The mature use of AI does not require choosing between complete delegation and complete independence. Human beings have always thought through relationships, languages, books, institutions, and tools. AI becomes another participant in that history, though one with an unusual capacity to respond, adapt, and act.

Its power gives us reason for excitement. Its speed gives us reason for discipline. Its apparent intelligence gives us reason to examine our own.

An AI-native company does not maximize AI use. It develops the judgment to know where AI belongs, prepares knowledge in forms that both humans and machines can work with, and measures the value created rather than the intelligence consumed. An AI-mature person follows a related practice: use AI to think, use AI to examine how that thinking is changing, and allow reality to evaluate both.

The movement from token maxing to AI maturity will not be completed by a new policy, model, or productivity method. It continues each time we resist the pressure to adopt an extreme, remember what earlier enthusiasm taught us, and choose the form of intelligence that genuinely serves the work.

Photo by Igor Omilaev on Unsplash