# What remains after the token? > We can buy the intelligence. The learning curve we have to keep. Why more has to survive an AI run than its output. - **Author:** Dominic Asche - **Date:** 09 / 2026 - **Reading time:** 5 min - **URL:** https://fse-group.de/en/think-tank/what-remains-after-the-token Our token bill has been rising ever since agents joined the team. The interesting question is not how high it runs. It is what remains once the token is spent. Investment in AI is rising, for us as much as anywhere else. It can pay off. In our view it only does when something more than output comes out of it: a system that gets better with every use. Since we started using AI seriously at FSE, our daily token bill has risen along with our productivity. We have brought agents into the team that do not just answer but work: they read information, decide, use tools, check results and start over where needed. That costs design, development and, yes, tokens. The bill is more likely to grow than shrink. At OpenAI, Codex already accounted for 64 percent of the combined enterprise output tokens from ChatGPT and Codex in June. Companies in the top decile of usage produced 8.3 times as many output tokens per active user as typical companies. OpenAI adds the caveat itself: token volume does not measure business value. It shows how intensively AI is used and how much work is handed to it. [1] Data from OpenRouter shows the same effect from another angle. Agentic applications now consume almost five times as many tokens as human users, and more than 85 percent of that is cached context. That makes sense: a chat follows the pattern of question and answer. An agent iterates towards a goal. [2] What interests us about these numbers is not the volume. It is what remains once the tokens are spent. ![Infographic titled Turning usage into capability on the question of what remains after the token. At the top a chain of four steps: first a mistake becomes an eval, second an instruction becomes a rule, third a proven approach becomes a tool, fourth a guardrail becomes a technical limit. Below a line chart titled Agents are using far more tokens than people, a 7-day average of token usage on OpenRouter by type. The line for agentic usage stays flat for a long time and then rises steeply, while human and mixed usage grow only slightly. Agentic applications consume almost five times as many tokens as human users, and more than 85 percent of that is cached context. Source OpenRouter, analysis by a16z, 21 August 2026. At the foot three consequences: evals instead of experience, so improvement becomes reproducible; rules in the harness rather than the prompt; stay model-agnostic and keep the learning curve.](https://fse-group.de/media/think-pieces/20260901_thinktank-token-en.png) ### Paying twice for the same insight An agent makes a mistake. We analyse it, sharpen its prompt, and next time it works. Good. But if that insight sits only in this one prompt, or in a colleague’s head, little has been gained. With the next agent, the next model, in a different process, the work starts over. So we buy intelligence a second time to find out something we had already learned. The single token is not what makes this expensive. What makes it expensive is producing the same insight again and again. ### What has to remain after the run More has to survive an AI run than its result. A mistake can become an eval. A recurring instruction can become a fixed rule. A proven approach can become a tool. And when a guardrail really matters, no model should have to interpret it; the system should enforce it. The token is gone. The capability that moved into the system stays. This has changed how we build our own agent platform. In the beginning a lot of logic lived in standardised agent descriptions: permitted tools, rules, limits, checks. The model was told as precisely as possible what it may and may not do. That worked surprisingly well. With every additional agent, two problems grew. The context got larger, and the same rules appeared in several places. Inconsistencies piled up, we kept correcting them, and the rule files kept growing. At some point a simple question came up: if a rule really matters to us, why are we merely asking the model to follow it? ### Rules belong in the harness, not in the prompt That question turned into a small but effective rebuild. Important guardrails, rules and tools moved out of the markdown files and prompts and into our own harness. “Use only these tools” becomes permissions. “Stop after X steps” becomes a technical limit. “Check your result” becomes defined gates and readbacks. And “the agent seems to be running better now” becomes evals. The reason is not primarily governance. It is the only way we know whether our system is actually improving. A mistake we correct today is experience. A mistake we turn into an eval changes how we measure quality from then on. When we switch models, we can check whether the new one performs better or worse. Improvement becomes reproducible. Our real investment is therefore not the individual agent. It is the ability to make the whole system measurably better. ### The intelligence may change, the learning curve may not A second development fits this picture. The Stanford Institute for Human-Centered AI compared inference costs for model performance at GPT-3.5 level: from 20 dollars per million tokens in November 2022 to 0.07 dollars in October 2024. That is roughly a factor of 280. [3] It does not follow that all models become equal. Our conclusion is a smaller one: plain access to capable general model intelligence is less and less of a differentiator. Which model leads on a given task will keep changing. Claude today, GPT, Gemini or a smaller specialised model tomorrow. That is exactly why we tie our differentiation to a single vendor as little as we can. Model and harness have to be interchangeable. What has to stay are our processes, tools, rules, evals, traces and the knowledge of what a good result even looks like. Satya Nadella argues in a similar direction with his “reverse information paradox”: companies should keep control of their evals, their feedback and their learning environment, and keep the choice between models open. His fifth principle is called “compound”. Usage should widen your own lead. [4] We would put it more simply. We can buy the intelligence. The learning curve we have to keep. ### What this means for utilities This applies to us as an IT service provider just as it does to companies in the energy sector. If everyone ends up with access to the same capable models, the difference will not come from being able to operate Claude, GPT or Gemini. It comes from turning that intelligence into a system that runs real processes better, stops repeating the same mistakes and can prove its quality is rising. For a utility this gets very concrete in one place. Anyone building an agent for clarification cases should not only count how many cases it resolves. The harder questions are different ones. Do the error patterns from three months ago still occur? Does a rule that applies in billing also hold in a supplier switch? And when the model is swapped, does the result get better or worse, and how do we know? Anyone who cannot answer those questions has not built an agent. They have bought an expensive habit. That leaves a rather unspectacular question, one that leads from the hype back to economics: what are we getting for our investment? More tokens mean more usage. More output can mean more productivity. It only gets interesting once something remains after the usage that works better the next time round. Then we are not just consuming intelligence. Then we are building a capability, and a capability is something you can create value with. #### Sources - [1] OpenAI: [*Enterprise Signals: What frontier firms are doing differently*](https://openai.com/business/enterprise-signals/), updated 12 August 2026. - [2] Andreessen Horowitz: [*Charts of the Week: Agents Are Early, But They’re Already Changing Things*](https://www.a16z.news/p/charts-of-the-week-agents-are-early), 21 August 2026. Analysis of OpenRouter data. - [3] Stanford Institute for Human-Centered AI: [*AI Index Report 2025*](https://hai.stanford.edu/ai-index/2025-ai-index-report), Research and Development chapter. - [4] Cloudmagazin: [*Nadellas Paradox: Wer KI nutzt, zahlt zweimal*](https://cloudmagazin.com/2026/07/13/nadellas-paradox-wer-ki-nutzt-zahlt-zweimal/), 13 July 2026 (in German).