The Enterprise Fight Against Runaway AI Costs

What happened

As enterprises increasingly turn to AI to get work done, three new weapons are emerging in their fight to control spiralling AI adoption costs: agent harness, model selection and inference economics. Simply put, AI adoption is no longer being judged solely on model capabilities, benchmarks or impressive agent demos, but on whether these systems can deliver measurable productivity gains without pushing project costs higher.

The next lever is model selection. Enterprises do not need their most powerful and expensive model for every task. By routing routine requests to smaller models and reserving frontier models for complex work, they can bring down the cost of each AI-powered transaction without compromising the quality of the outcome.

According to Krupesh Bhat, the founder and CEO of Melento, a document automation company formerly known as SignDesk, the goal is to use expensive AI only where it is needed. Melento’s systems route routine requests to smaller, faster models and send only genuinely complex cases to frontier models.

To determine whether this approach is lowering project costs, he suggested that enterprises should track: If cheaper models handle more of the workload while the success rate remains stable, the enterprise can reduce project costs without compromising outcomes. This is also where the distinction between productivity and savings becomes important. An AI system may complete more work, but that does not automatically mean a project is cheaper.

The savings appears only when higher throughput translates into lower spending on models, people, or rework. Even after removing unnecessary AI calls and limiting the use of expensive models, enterprises still need to reduce the cost of running the AI that remains. This is where inference economics comes into picture.

In one project for a telecom operator processing more than a million interactions a month, AIONOS moved roughly 30% of routine steps to ordinary software. It also used an open-weight model trained on resolved customer queries, reused information where possible, and shifted some work to more cost-effective computing capacity. The company said the total cost of resolving each interaction fell by 35%.

Why it matters

As AI moves from pilots into everyday enterprise workflows, the cost of every model call, token, inference and computing resource begins to add up. The bigger question, then, is not simply which model an enterprise uses, but how the entire AI system is designed around it.

Companies are increasingly looking at ways to reduce unnecessary model calls, route simpler tasks to cheaper models and remove AI from workflows where conventional software can do the job just as well. This prompted us to ask a sharper question: what actually goes into cutting the cost of an AI project, and which engineering decisions can keep the bill from rising as deployments scale?

The first lever is the agent harness, the software layer that manages how an AI agent uses a model. It decides what information the model receives, how a task is broken down and how often the model needs to be called. This was illustrated in late July when Sarvam AI launched Sarvam Code, a coding agent that divides engineering tasks between planner, worker and verifier agents.

8 for Claude Code and OpenAI’s Codex. Sarvam also said its system, which routes work between its Sarvam 105B model and a larger open model, reduced serving costs by about 40%. The point is not that the harness makes the model itself cheaper. Instead, it helps prevent the model from doing unnecessary work.

As an agent takes more steps, it can repeatedly send parts of the earlier conversation, instructions and other context back to the model. A task that takes 40 steps can therefore consume millions of tokens, even if the original request contained only a few thousand. The job of the harness is to limit this repetition.

It can ensure that the model receives only the information it needs for the next step and can store information that is likely to be needed again. This is where caching comes in. Instead of sending the same information to the model and paying for it to process that information again, the system can reuse what it has already processed.

Malladi estimates that this kind of caching can reduce the effective cost by roughly 80%. In other words, one of the easiest ways to lower an AI bill is to avoid sending unnecessary information to the model in the first place. This makes the harness an important cost lever. Rather than simply using the model more efficiently, enterprises can use the same model budget to complete more work.

What to watch

The largest saving came from the first step, the work that stopped being an AI problem altogether. Other measures help reduce the cost of the AI that remains. Smaller models can handle repetitive tasks, while training them on a company’s own past cases can reduce mistakes and the human effort needed to correct them. Reusing previous answers can also prevent the same question from being sent to the model repeatedly.

Enterprises can lower costs further by processing several requests together instead of handling each one separately. This keeps their computing capacity working steadily, rather than paying for expensive equipment that sits idle between requests.