Perception vs. Reality: The Real Cost of AI

One of the easiest assumptions to make about AI is that it is cheap.

You drop a prompt into ChatGPT, get back a polished answer in seconds, and it’s not hard to see why so many businesses have started treating it like a productivity hack with very little visible cost attached.

A few subscriptions for the team can look like a fair trade for faster output and extra capacity.

But that sense of low cost is not quite as straightforward as it seems.

Where’s the real cost?

Whilst it is easy to think that AI subscriptions are just another figure on the bottom line, most business leaders are not especially aware of how those costs are actually measured.

A big part of the answer lies in AI tokens.

In simple terms, tokens are the chunks of language AI platforms use to process and generate text. They sit behind every prompt you type and every response you get back. Anthropic describes them as the smallest units of a language model, and notes that for Claude, a token is approximately 3.5 English characters, though the exact number varies depending on the language used.

And this is where the topic stops being technical and starts affecting cost: tokens are often what pricing is built around. For example, Anthropic’s pricing model separates input tokens from output tokens, and also accounts for factors like cache writes, cache hits, and batch processing.

So the economics of AI are not just about whether you use it. They are about how much language the system has to process, generate and keep carrying with it as usage grows.

That’s why AI can feel cheap on the surface while the underlying model is telling a different story: the more language the system has to deal with, the more cost starts to build.

What tokens actually point to

Each token represents something much heavier underneath: infrastructure, computing power and energy.

Every time a tool generates a response, it’s doing work in the background. And that work is not the same every time. A short task does not place the same demand on a model as a long prompt, a complex analysis, or another round of “make it better” because the first answer was not quite right.

For a lot of AI usage today, especially at API level, pricing is not built around a simple flat subscription - it’s built mainly around the consumption of input and output tokens. That all points to the same reality: providers are trying to price AI in a way that reflects the amount of work the system is doing underneath.

This is where the illusion of cheap AI starts to wobble.

When businesses think of AI as a low-cost productivity layer, they are usually thinking about what they see on screen: a prompt goes in, a response comes back, and the monthly spend looks manageable. What they are not always thinking about is how usage expands over time, how long conversations quietly get more expensive, or how repeated regenerating and heavier models turn convenience into consumption.

When casual use becomes normal use

Once AI becomes part of day-to-day work, the risk goes beyond cost creep.

It’s the casual use becoming normal use, built quietly into business processes. A bit of prompting here, some rewording there, a few workflows added into the tech stack. Before long, the team has developed a habit without ever really defining a standard.

As a result, AI starts being treated less like infrastructure and more like background noise, or a slightly unreliable thermostat: always on, constantly available, rarely challenged.

And that’s where I think the next phase of AI becomes a leadership matter.

If teams are using AI casually, repeatedly and at scale, but nobody is really watching how that usage behaves, then control is already slipping. Read my AI Joyriding blog to learn more about that.

And once cost becomes more visible, whether through token bills, usage reviews, more granular charging models or internal governance, a lot of businesses are going to realise they’re not quite as on top of it as they thought.

So what should business leaders do now?

Not panic, for a start.

This is not a call to retreat from AI or start counting every token like you are monitoring the office printer or fax machine (do you remember those days?!).

But it is a prompt to get a bit more focused on how AI is being used inside the business.

If token-based pricing is already the reality for a lot of AI usage, and providers are making cost models more granular, then casual use will not stay casual forever. Anthropic’s pricing model already separates input, output, caching and batch processing, which tells you exactly where this is heading.

For me, the practical questions are fairly simple.

• Do we know which tools are being used most heavily?
• Do we understand where usage is creating real value, and where it is just creating volume?
• Are teams defaulting to premium models for routine tasks?
• And are repeated prompts, long context windows and messy workflows quietly driving cost without improving the outcome?

The businesses that handle this well will not necessarily be the ones using AI less. They will be the ones using it with more visibility, more intention and better standards around what good actually looks like.

And in my experience, that’s usually where better decisions start.

If you’re stuck with determining the real cost of AI in your business, book your AI marketing healthcheck today.