Synopsis: Cheaper tokens do not guarantee a smaller AI bill. An agent can make repeated model calls, invoke tools and delegate work before completing a task that looks straightforward to the person requesting it. As organizations automate more work, that expanding usage can outweigh the savings from falling prices. Measuring the cost of a successful business outcome requires more than checking what a model charges per token.
Jon Knisley, director of AI value management at ABBYY, joins Mike Vizard to explain the economics behind that mismatch. He describes a version of Jevons Paradox: making each unit of AI cheaper encourages organizations to consume more of it. Moving from individual chatbot interactions to agents embedded in business processes adds reasoning steps, integrations and operational complexity. The question becomes whether that additional activity delivers enough value to justify its total cost, not simply whether the latest model is less expensive than its predecessor.
Knisley points to several ways to make those choices more deliberately. Frontier models may suit demanding reasoning tasks, while routine production work can use less expensive or domain-specific alternatives. Model routing can help match the task to an appropriate level of capability. How information reaches the model matters, too: changing a document’s structure can reduce the tokens needed to process it. Across a large document workload, input preparation can become a meaningful part of cost management rather than an implementation detail.
A contract-lease extraction example brings the focus back to results. Knisley describes a project where sending complex documents to an LLM alone did not deliver the required accuracy; combining traditional machine learning with generative techniques performed better. The lesson is to examine the whole workflow, including data preparation, unnecessary steps and where people remain involved. Utilization, accuracy and operational impact belong alongside the token bill. Spending less on a model is not much of a saving if the process still produces unreliable results or leaves valuable work unfinished.
