Ignorance Now Has a Meter

Before anyone gets offended about the title, I'm using "ignorant" in the literal sense. If you're not an expert in something, you don't know what you don't know. So, in this particular piece, ignorance is not an insult. I'll use the softer word, novice, going forward.

The AI industry is in the middle of a pricing pivot. The flat-rate, all-you-can-eat subscription era is winding down, and consumption-based billing (pay-per-token) is here. And, predictably, the media is now flooded with articles about how enterprise customers are pulling back on spend, and suddenly the CFOs want to know what all these tokens are producing.

As an educator, what I find interesting about this shift is what the meter reveals:

Experts are Cheaper to Run

I've had a strong career in business and tech, and so I have a lot of people in my network who are experts in their domains and respected by their peers. So, I get a lot of information on what expert usage looks like. But first, for the novices reading this, let me recap how LLMs work:

Everything you send to a model is measured in tokens: the prompt, any associated files, information pulled from other sources (such as web searches), and the entire conversation history. Under metered billing, you pay for it all, every single time. So, a long session is a compounding cost. And as a bonus, bloated context not only costs more, it also degrades output quality because models get demonstrably worse at using information as the context window increases.

So, back to the experts. Here's what most of the experts in my network do:

• They write tight, scoped prompts because they know exactly what to ask for (including edge cases) and exactly what the outputs should be.
• When the output is 90% correct, they can fix the last 10% themselves. They do not spend six round trips asking AI to move a button or adjust a color. They just... make the change. It only takes a few minutes.
• They can evaluate output quality, which means they can route tasks appropriately. Cheap fast model for the boilerplate, expensive frontier model for the hard problem. Not the flagship for everything.
• They know when the AI is the wrong tool, and can roll up their sleeves and do the work.

Every one of those behaviors requires expertise.

You cannot scope a detailed prompt for work you don't understand.

You cannot fix the last 10% of code you can't read.

You cannot evaluate the quality of output in a domain where you have no taste.

It's not a coincidence that some job ads that want heavy AI use have started to mention "taste" as a desirable candidate trait, this is a proxy for "expertise".

Novices are More Expensive to Run

Now, consider the inverse. I see this all the time in my training and advisory.

They write vague prompts because they don't have the expertise to be specific. The output looks ok, but the novice can't tell the difference between "done" and "looks done", so they either accept garbage or they start looping.

Every minor tweak re-sends the entire bloated conversation through the meter. They can't make a 2-minute manual fix, so it becomes a multi-hour, multi-request negotiation between the novice and the chatbot, and the company is paying for both.

They're not being lazy or stupid. They're being ignorant, in the literal sense. They don't know what good looks like, what context the model needs, or when to stop. The meter does not care that they're doing their best. The meter just runs.

Under flat-rate pricing, all of this was invisible. A confused employee flailing at a chatbot cost the same as your best engineer surgically extracting value from it. That era is ending, and ignorance is about to get taxed.

If I were a supervillain who owned an AI company, I would be doing everything I could to burrow into education and stop expertise from being gained, because a legion of ignorant users is far more profitable under metered billing.

The Executive Problem

Now I have to hit the executives who caused this mess with both barrels.

For the last two years, the standard executive playbook has been to buy AI tools for everyone, mandate their use, and measure adoption. I've watched this from the inside of corporate training. To be fair, under seat-based pricing, it was rational. Worst case, you wasted some licenses because the downside was capped.

The part where CFOs are about to lose their minds is that the reporting can't distinguish between valuable tokens and wasteful ones. Those high usage numbers might mean that your workforce has been productively transformed. It might also mean that most of your workforce is burning tokens in correction loops or low-value tasks because they don't know what they're doing.

Adoption is a vanity metric that has nothing to do with return on investment.

Expertise is More Valuable than Ever

Metered AI was supposed to be the final commoditization of expertise. Intelligence too cheap to meter, remember? Why pay a premium for senior people when the machine does the thinking?

Instead, the meter is putting a price tag on the absence of expertise. This compounds in the loops that novices often find themselves in. The cost of not knowing what you're doing is a line item on the invoice. And I haven't seen any enterprise that has the processes or tools in place to tell the difference.

I've said this before, and I'll keep saying it: the most effective way to use AI tools is to know what you're doing. The industry has spent years insisting otherwise, but its pricing model is proving my point. Because, honestly, if the outputs were superior for anyone, the industry would charge by value/task completed, not by token.

And if you're a learner thinking about using AI to shortcut your expertise, you'd best start thinking about how you're going to explain to your manager why you're burning 10x the tokens as your peer who developed expertise the right way.

The premium on human expertise is about to go up. And this time, there's a meter to prove it.