Traditional software cost tracks users. Add users, cost rises; hold users flat, cost holds flat. AI breaks that intuition. An AI bill can grow steadily while your user count sits perfectly still, and the reason reveals something important about how these systems consume resources.
Cost tracks tokens, not people
The core reason is that AI cost is a function of tokens and calls, not of users. A single user can generate a little consumption or an enormous amount depending on what they do and how your system responds. Because the link between a user and their token consumption is loose and variable, a stable user base can sit on top of a rising or falling token base without any contradiction.
Agents work harder over time
As agentic features mature, they often take on more per task, more reasoning steps, more tool calls, more self-checking. An agent that did three steps at launch might do eight after a few iterations aimed at improving quality. Each added step is more calls and more tokens for the same user doing the same thing. The product got better. It also got more expensive per interaction, with no change in how many people use it.
Sessions get longer and heavier
Engaged users have longer sessions, and in conversational features longer sessions cost more because context accumulates. A flat user count with rising engagement produces a rising bill. This is not waste, it is usage deepening, but it decouples cost from user count in a way that surprises anyone expecting the two to move together.
Background work has no user at all
A growing share of AI cost comes from work that no user triggered directly: enrichment jobs, monitoring agents, scheduled analyses. This spend scales with your data and your ambitions, not your user base. A company can hold users flat and still grow this category simply by automating more.
Keeping cost tied to value
The risk in this decoupling is that cost can grow without a corresponding growth in value if nobody is watching the relationship. The discipline is to measure cost against outcomes rather than against users, cost per completed task, cost per resolved ticket, cost per unit of whatever your AI actually produces. When you track that ratio, rising cost is fine as long as the value rises with it, and a problem only when it does not.
A bill that grows while users stay flat is not inherently wrong. It is a signal to check whether the extra spend is buying extra value. In AI systems, cost was never going to track headcount. The job is to make sure it tracks results.