Your company switched on an AI feature. It works, everyone loves it, and then the invoice arrives — bigger than expected, and nobody can explain exactly what made it grow. Welcome to the most common blind spot of AI adoption: the cost, which almost nobody measures.
The good news is that this cost is no mystery. It has a unit, that unit is countable, and the simple act of counting it changes how you decide. Let's take it step by step.
01The meter nobody turns on
Traditional software has a relatively stable cost: once built, running it costs roughly the same no matter who uses it. Generative AI is different — every call has a variable cost that depends on how much text goes in and comes out. You pay per use, all the time.
The problem is that almost nobody turns on the meter per feature. You know the invoice total at the end of the month, but not how much each AI feature consumed — which one is cheap, which one is getting expensive, which one spiked after a change. Without that, optimizing is shooting in the dark.
02The unit is the token
The token is the piece of text a language model processes — roughly a word (sometimes less). Everything that goes into the model (the prompt, the context, your documents) and everything that comes out (the answer) is counted in tokens. And it is in tokens that AI cost and limits are measured.
That changes how you think: a "simple" feature that injects an entire document of context on every call can cost far more than a lean "complex" one. What matters is not how the feature looks — it is how many tokens it moves.
03Why estimating is not enough
The temptation is to estimate: "it should cost a few cents per use." The problem is that token estimates miss badly — high and low — because consumption depends on invisible details: the size of the context, how many times the model is called per operation, whether there are retries, whether the conversation history grows with every message.
A single operation that looks like one question can, under the hood, fire dozens of calls to the model. Without measuring, you find out on the invoice — late, and with no idea where to act.
04What changes when you measure
When token cost is visible per feature, three things become possible. You cut the waste (that giant context that never needed to ride along). You turn on caching — reusing repeated parts of prompts slashes a huge slice of the bill. And you decide with numbers: is that AI feature worth its real cost, or does a deterministic path solve it for less?
None of that is possible in the dark. That is why measuring tokens is not an engineer's whim — it is the difference between an AI that pays for itself and an AI that quietly bleeds money.
05How we do it
At ZDZCloud, AI cost is not an estimate — it is telemetry. When we build with our agent factory, we measure consumption token by story, not by guesswork.
A real example: a sprint of 8 stories from our Pet Planner product consumed 385 LLM calls and 3.28 million input tokens — with a 71% cache hit rate. That 71% did not show up by luck: it showed up because the bill was being measured, and what gets measured gets optimized. It is the same discipline we bring to the AI cost of every product we build.
TAKE THIS WITH YOU Cost questions to ask before scaling your AI
- Do you measure per feature? Do you know how many tokens each AI feature consumes — or only the invoice total?
- How much context rides along? Does each call carry only what it needs, or an entire document out of habit?
- How many calls per operation? One question = one call, or dozens under the hood?
- Is there caching? Are repeated parts of prompts reused — or paid for again every time?
- Would deterministic do? Does this feature really need an LLM, or would a rule be cheaper?
Good AI is not the one that uses the most model — it is the one that delivers the result at the lowest token cost the job requires. And that starts with something banal and rare: turning on the meter.
Want to know what your AI really costs?
We run a free diagnosis of where AI shows up in your operation and what it consumes — and map the path to an AI that pays for itself. No sales pressure.
Request a free diagnosis →