We spend a lot of time talking about how to use AI more effectively.

But what if one of the most valuable AI skills turns out to be knowing when not to use it?

As AI becomes better at generating text, images, code, and analysis, the temptation is to use it for everything. The harder skill may be recognizing the moments when speed should give way to human judgment, firsthand experience, creativity, or genuine connection.

AI literacy isn’t just learning better prompts. It’s learning where AI belongs in your process and where it doesn’t.

AI Felt Unlimited, But The Limits Are Starting To Show

For many users, AI has felt frictionless. You type a prompt, and the answer appears. Maybe you pay a monthly subscription. Maybe your university or company provides access. Maybe you use a free version and only occasionally hit a limit. The meter was mostly invisible.

But the meter was always running. Free users are seeing more, “You’ve reached your free messages.” Businesses are blowing through AI budgets faster than expected.

The real economics behind the helpfull chat boxes are becoming harder to ignore.

Every prompt uses tokens. Every long conversation, file upload, agentic workflow, image request, or coding session uses compute. For now, many of those costs have been hidden, subsidized, bundled, or absorbed by companies trying to grow adoption. But that may not last long.

In his newsletter, Christopher S. Penn put token economics in blunt terms. He estimated that a $200/month Claude Max 20 plan can provide roughly $8,000 worth of token output or a 97.5% discount. No business can sell its product at that discount forever.

AI has felt unlimited because someone else has been absorbing much of the cost. But that may beginning to change.

A TechCrunch article described companies starting to balk at the price of AI as token usage grows. Uber reportedly burned through its 2026 AI coding budget by April. Microsoft pulled back some developer access to Claude Code months after enabling them.

There is an environmental side to this, too. The compute behind AI has energy, water, and infrastructure costs. Using AI more efficiently is not only better for budgets. It is also part of using the technology more responsibly.

The Hidden Cost of “Use AI”

Need ideas? Use AI. Need a summary? Use AI. Need an email? Use AI.

Need a spreadsheet formula, campaign plan, research question, presentation outline, or first draft? Use AI. Sometimes that’s the right choice. But “use AI” can also become too frictionless.

The danger is losing the pause and questions that used to happen before work began.

What am I trying to do? Do I understand the problem? Or is AI even needed? What should I think through myself first?

If the person using AI doesn’t understand the underlying work, will a more powerful model help produce better work, or simply produce a polished version of weak thinking they’re not able to judge anyway?

That last question matters for higher education. Students need to learn how to use AI, but they also need to learn how to work when AI is not available, not allowed, too expensive, or not right for the task. If student learning depends too much on AI they may price themselves out of the jobs they’re preparing for.

When Skills Depend on a High Token Budget

A student who can only produce good work with unlimited access to the most powerful AI is not as AI-ready as they may think. They may be AI-dependent.

This applies to basic communication, too. If students never learn how to write a clear email, organize a simple slide deck, summarize a source, or explain a recommendation, they may have to spend tokens every time they communicate.

They’re not more efficient because they know AI. They become more expensive because they depend upon AI.

This will matter more as organizations start managing AI costs more carefully. Employers may not always give every employee unlimited access to every model. They may expect employees to know when a task requires, simple AI, advanced AI or no AI.

Business Insider reported that Coinbase CEO Brian Armstrong is trying to keep AI costs roughly flat while token usage grows by routing prompts to cheaper models to match the task to the right level of AI. That’s not only a technical skill. It requires understanding the work.

Someone who knows the subject can often get a useful answer from a smaller, cheaper model because they know what to ask, what context matters, and how to judge the output. Someone who doesn’t understand the work may need a more powerful model, more attempts, and more tokens just to get close. In that sense, human skill becomes part of AI efficiency.

Tokenmaxxing was short lived, the next metric will be tokenminning.

Prompt skills may matter more when AI is no longer treated as free and unlimited.

Prompting Is Not a Substitute for Knowing

On the other hand just knowing prompting skills isn’t enough. Shallow AI literacy says: “I know how to prompt.”

A deeper version says: “I understand the work well enough to know when AI can help, how to ask for that help, what kind of model is needed, what the output is worth, and when I should do the thinking myself.” That’s the skill employers will be seeking and higher education should be developing.

In marketing, that might mean knowing when AI can help generate alternative versions of a positioning statement but not replace the hard work of the human insight into the market and target audience.

In advertising, it might mean using AI to explore headline territories but still knowing which line has a real idea behind it or will resonate with how people really think or talk.

In research, it might mean using AI to summarize possible sources while still opening the article, reading the methodology, and deciding whether the evidence supports the claim.

In design, it might mean using AI to produce quick visual directions while still grounding the work in human observation, user needs, constraints, and context.

The important skill is not using AI. Its knowing what kind of human judgment the task requires before deciding what kind of AI assistance is useful.

The Most Expensive Answer Is Not Always the Best Answer

One reason this matters is that more AI does not always mean better work.

A study, “How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks”, found AI agents can consume vastly more tokens than simpler AI workflows. Agentic tasks can use up to 1,000 times more tokens than code reasoning and code chat, and that higher token use doesn’t always translate into higher accuracy.

It’s easy to assume more AI means more capability. But more is not always better.

Sometimes a smaller model with a better prompt is enough. Sometimes a search engine is enough. Sometimes a conversation with a colleague is better. Sometimes sitting with the problem for ten minutes is the fastest path to a better idea.

That may sound old-fashioned, but it’s also practical. Burning AI tokens on a task a simple Excel formula could handle is not maxing your value to your employer.

If AI becomes part of the cost structure of work, using the most powerful model for every task may become like ordering Surf and Turf for every meal. Eventually someone notices the bill.

A Guide To Using The Level Of AI V2

Human Plus AI, Not AI for Everything

AI should amplify what students and professionals can do, not become the only way they can do it. That means the best AI users won’t be the ones who use the most AI. They’ll be the ones who know when to think, when to prompt, when a smaller model is enough, and when the work should stay human.

The same caution I offered writers and readers applies here. Don’t abdicate your human discernment to AI companies that have incentive to make AI feel frictionless, indispensable, and unlimited.

As AI becomes more powerful, organizations will measure it more closely. The people paying for these tools will want to know whether AI is saving time, improving quality, reducing risk, or simply moving cost from payroll to tokens.

That is why AI literacy needs to be taught in the classroom and on the job. Not just how to use AI, but when to use it, how much to use, and when not to use it. And that takes human judgment.

---

AI use disclosure: The central idea for this post was mine. It came from thinking about AI writing, AI detection, and the growing cost of token-based AI use. I used ChatGPT as a thought partner to develop the structure, test the higher-education angle, and identify possible supporting sources on AI token costs, model routing, and agentic AI workflows. I opened and checked sources before using them. I then rewrote, reorganized, added, cut, and edited the draft to reflect my own experiences, views, and voice. The stories, examples, and final decisions are my own.