Ask a development company how much artificial intelligence it used last month and the answer will often be a blank stare, a shrug, or one number pulled from a provider invoice. Even the invoice is usually wrong because it captures only the AI calls someone deliberately built. It misses the features quietly embedded in search, code editors, design software, CRMs, analytics, meeting tools, and project platforms.
We are writing AI policies before we have learned how to measure our own use.
Most AI use was never a conscious decision
A developer searches for syntax and receives an AI-generated overview. A code editor activates a completion model with nearly every keystroke. A CRM drafts emails. A meeting platform transcribes every call. A dashboard generates narrative summaries whether anyone reads them or not.
Some of these features are genuinely useful. The problem is visibility. A team may believe it is a light AI user because it makes two intentional API calls a day while background features are performing many times that amount of inference.
Begin with an AI inventory
List every tool the organization uses. Record which ones contain AI, whether those features are enabled, whether they were enabled by default, how frequently they activate, what model they use, and where the data goes. Gather exact token and completion counts where they exist. When they do not, record a reasonable estimate and make the assumption visible.
The result is often surprising. A background summary feature nobody reads may consume more resources than the AI product a team spent months building. That is not a moral judgment. It is an operational fact—and it only becomes visible when someone counts.
The largest model is rarely required
Powerful AI does not always require the largest model or the most expensive hardware. Smaller and open models can be cheaper, faster, more private, and more efficient when they meet the acceptance criteria for the task.
The important question is not where a model was created. It is whether that model is appropriate for the job, how many resources it consumes, where it runs, what powers the infrastructure, and what happens to the data it receives.
What this means at Web Experts
We keep a current inventory of AI features in our own systems and the products we build. We normally begin with the smallest model capable of meeting the requirement and move to a larger one only when the evidence says we need it.
AI assists our developers; it does not replace their responsibility. Agent-produced code is reviewed by experienced developers before it enters a project. If a problem is better solved with conventional code, a database, or a simple form, that is what we build.
We created Tokenprint to make the environmental side of this inventory easier to understand. The free tool estimates energy, water, and carbon, includes model training and hardware manufacturing, adjusts for the power grid, and shows every assumption behind the result.