When a company evaluates an AI project, it compares usage prices. That is the most visible figure and the least decisive. In a real setup, model consumption is rarely the main share of the spend.

Cost 1: implementation

Describing the task, gathering the material, defining what a good result looks like, connecting the tools: this work is human, it is measured in days, and everything else depends on it. It is also the cost people underestimate most, because it looks like organizational work rather than technology.

Cost 2: oversight

A setup left alone drifts: edge cases pile up, instructions go stale, quality drops without any alert. Planning regular review time is not a luxury, it is what keeps the gain going beyond the first quarter.

An unmonitored system doesn't stay stable. It slowly degrades, until someone declares that it doesn't work.

Cost 3: undetected errors

This is the invisible cost, and potentially the heaviest. A wrong answer sent to a customer, a false figure carried into a document, an invented piece of information that travels down the chain: the cost isn't in the calculation, it's in the repair. Hence the common-sense rule: human review wherever an error carries consequences.

Cost 4: poorly sized usage

  • Using a very powerful model for a simple task, out of convenience, multiplies the spend without improving the result.
  • Sending an entire context with every call instead of the relevant excerpt quietly inflates consumption.
  • Rerunning an identical job because the previous result wasn't saved is the most common source of waste.

These adjustments are not a technical detail: they routinely cut the usage bill by a factor of several, with identical results.

What to compare

Not the model's price, but the full cost of one unit of work: what it costs, all in, to handle one request, yesterday by hand, today with the setup. It is the only figure that lets you decide whether to continue, expand or stop.

Demand three figures before scaling up

Demand three figures before scaling up: the implementation cost, the monthly running cost including oversight, and the cost per unit handled compared with the existing process. A project that can't provide these three figures isn't ready to be expanded.

AI isn't what's expensive. What's expensive is AI set up with nobody to tune it.