Notice: This article was created with AI.
What’s It About?
A piece in Computerwoche looks at the economics of AI in companies and puts “dosage” at the center of it. Not every task needs the strongest model or an agent that acts on its own, it argues. What matters is who takes on what: fixed rules, models with room for interpretation, or agents that decide on follow-up steps themselves. According to the article, this distribution affects resource consumption, the quality of results and the amount of oversight needed, and with that the ROI.
Background & Context
As the central cost driver the piece names tokens. These are small units of text such as parts of words, syllables or punctuation marks that a model processes. Both the input and the output cause computing effort and therefore costs. The more demanding the task, the more tokens arise. A short email draft is cheap. A multi-stage agent that reads documents, weighs up alternatives, writes code and corrects it, by contrast, consumes tokens at every step. Because every intermediate step reads the history so far all over again, it is above all the input tokens that drive the costs, according to the article.
The piece also points out that the overall bill can rise sharply even when individual tokens become cheaper. Particularly intensive users, often with demanding coding applications, accordingly consume a multiple of what a typical user calls upon. Without control, the ROI is at risk of suffering.
The article also clears up a widespread opposition: the dividing line does not run between rigid processes and “unpredictable” AI, because fixed workflows can contain AI steps too. Boris Krumrey of UiPath warns against letting agents derive a process anew on every run. That costs computing power, leads to fluctuating results and makes auditing and compliance harder. By this logic, fixed steps belong in a workflow, while changing content or open-ended processing tasks are better suited to models or agents.
What Does This Mean?
- Tasks should be distributed deliberately: recurring, clearly defined steps fit into fixed workflows, open-ended tasks rather with models or agents.
- Token consumption belongs in monitoring early on, because falling prices per token do not automatically mean lower overall costs.
- According to Computerwoche, several industry representatives argue for examining the deployment continuously after the rollout and taking tasks back off agents where necessary.
Sources
Zum ROI von KI: Die Kunst der Dosierung (Computerwoche)
AI use is exploding, so are the bills (J.P. Morgan Private Bank)
KI-Nutzen ohne ständig ans Limit zu laufen (ai-first.ai)
Maximizing AI ROI (SAP)
This article was created with AI and is based on the listed sources as well as the language model’s training data.
Further Reading: My Digital AI Newsroom on the Synology NAS