ROI From Your AI Agents Isn't About Costs

A client recently shared a screenshot of their monthly AI model bill showing three agents that had been live for just three months. The agents were performing useful work, yet no one could explain how the costs had compounded to such an extent. This scenario is becoming increasingly common as organizations deploy AI agents without proper cost management strategies.

After implementing over 100 AI agents, Siva Surendira, Founder and CEO of Lyzr AI, has observed monthly frontier model spending ranging from $50,000 at smaller companies to $4 million at large retailers. The bigger question isn't just about these costs, but what companies are asking these models to do on such a scale.

Model Misallocation

When new tasks arrive, they typically go to the most capable model regardless of complexity. An agent meant to review contract risks and one meant to build marketing strategy end up running the same way in most companies. Because model quality has improved and inference costs have fallen, tweaking every AI agent call feels unnecessary.

At scale, this approach creates an allocation nightmare, akin to assigning a principal engineer to reset a system password or a law firm partner to check spelling in a contract. The solution is to match the model to the difficulty of the task, applying the same principles we've learned about workforce expertise allocation.

Cost-Effective Implementation

Much enterprise work involves classification, extraction, lookup, formatting or validation—tasks that don't necessarily require advanced reasoning. A simple routing layer can assess the task and send it to the least expensive model that meets quality requirements. For complex cases, breaking jobs into smaller pieces and comparing independent runs can determine if more compute power is needed.

Companies can test smaller models alongside production traffic, comparing their answers against stronger models on repetitive tasks. When the smaller model consistently meets quality standards, traffic can gradually be shifted to it. This approach recognizes that the model best suited for a task today may not be optimal six months from now.

Measuring True ROI

A token bill is not an ROI metric. McKinsey's 2026 enterprise AI FinOps survey found that 93% of respondents had exceeded their AI budgets. Research shows token usage for the same coding task can vary by as much as 30 times depending on how an agent executes it.

The meaningful metric is closer to cost per completed outcome—whether that's a resolved ticket, processed document, cleared transaction or merged pull request. This requires connecting consumption to teams, workflows and results. Instead of asking what models cost, companies should ask why a particular task cost what it did.

Consider customer refund requests as five distinct tasks: email classification, order lookup, policy verification, exception handling and response generation. Only the exception handling requires a premium model; the other steps can be handled by smaller models or even deterministic logic. Running all tasks on a premium model means paying top rates for steps that don't require advanced intelligence.