Why AI cost control starts with reducing unnecessary compute

Choosing cheaper models can help manage artificial intelligence costs, but it does not address the whole problem, according to Ben Canning of TechRadar. As companies shift from experimentation to deploying AI agents at scale, compute bills are rising. The source says Uber used its entire AI budget for 2026 in four months.

Model routing and lower-cost models, including Chinese models such as Kimi K3, can reduce spending for selected tasks. But Canning argues that organizations also need to decide when a large language model (LLM) is appropriate. Repeated tasks such as file reconciliation, business-rule checks and compliance work may be handled more efficiently by existing data-processing workflows.

Many LLM systems, the article says, lack a business logic layer: workflows that apply an organization’s own definitions, rules and compliance requirements. Without one, a model may repeatedly gather broad context to answer a question and still miss internal definitions, such as how a company calculates margin.

With that layer, analytics workflows can calculate recurring measures, while an LLM uses their output. The article says this can reduce unnecessary token use and duplicated compute between cloud data platforms and LLMs. It also argues that consistent business logic can support confidence in AI tools and their rollout.

Canning cautions that agents can act in the wrong direction without appropriate rules. In an enterprise with 1,000 agents, he writes, a shared source of business context can help keep answers consistent. He concludes that targeted use cases, trusted workflows and governed business logic can improve returns from AI.