AI-tokenomics: more AI does not automatically mean more value

Generative AI is becoming increasingly accessible. Sending a prompt, analyzing a document, or having an AI agent perform a task often seems like a single, simple action to the user. Behind the scenes, however, the reality can be quite different. Multiple model calls, validation steps, tool interactions, and agents may be required to achieve a single result. And every interaction consumes tokens. This raises an increasingly relevant question: when does AI actually create more value, and when does it mainly add cost and complexity?

 

What is AI-tokenomics?

Tokens are the small units in which an AI model processes information. Both the input the model receives and the output it generates consist of tokens. The more extensive the context, the longer the response, and the more often a model is called, the greater the token consumption. This forms the basis of what we can call AI tokenomics: the economic model behind AI usage, in which consumption and costs are closely linked.

Compare it to a consulting request. You ask a single question, but behind the scenes, one consultant interprets the request, a colleague gathers information, a specialist prepares an answer, and a reviewer checks the result. It may then be revised once more. Each step can be useful, but every additional handover and review also costs time and money.

The same can happen with Agentic AI. What appears to be a single action to the user may involve multiple model interactions, tool calls, validation steps, and exchanges between different agents.

 

When AI usage and AI value start to diverge

Token-based pricing is understandable in itself. AI models require significant computing power, infrastructure, and energy. A pay-per-use model also makes advanced AI accessible without requiring organizations to build and maintain the entire infrastructure themselves. However, this also creates an interesting tension.

Under a consumption-based model, an AI provider earns more when more tokens are used. For the organization using AI, however, the value lies in solving a problem as efficiently as possible. This difference becomes increasingly important when AI becomes part of business processes.

An automated action, for example, may involve multiple AI agents that gather information, analyze it, review each other’s results, and ultimately reach a decision. Every additional step increases token consumption. But does that extra step actually improve the outcome? A larger model is not necessarily better. Neither is a longer response. And five collaborating agents do not automatically create more value than one clearly defined AI task.

 

From token consumption to business value: how Pega approaches this

The focus therefore shifts from how much AI are we using? to how do we use AI in a predictable and manageable way? Pega addresses this with Predictable AI. During the design and development phase, AI can help create structured, deterministic workflows. This means the process does not need to be determined by AI again every time it is executed. Instead, the workflow defines the process flow, ensuring repeatability and predictability.

A workflow can still deliberately include an AI step, for example for interpretation or analysis. That step then performs a specific task, while the workflow remains in control. This prevents AI from determining the process from scratch with every execution, which could otherwise result in varying process paths, additional model interactions, and unpredictable token consumption.

This approach is also reflected in the pricing of Pega Infinity 26: Pega opts for predictable AI costs rather than separately metered token costs for Pega-managed AI.

 

Conclusion on AI-tokenomics: focus more on outcomes, less on costs

AI tokenomics is not just about costs, but about finding the right balance between AI usage, complexity, and business value. The goal is not to maximize the use of AI, but to solve problems effectively. The best solutions use AI where it adds value while remaining as predictable as possible. Within Pega, this can be seen in the combination of AI with deterministic workflows and predictable outcomes.