The Enterprise Fight Against Runaway AI Costs

As AI moves from pilots into everyday enterprise workflows, per Inc42, three engineering levers are emerging as the primary tools for controlling costs: agent harness design, model selection routing, and inference economics. The shift marks a turning point where AI adoption is no longer judged solely on model capabilities or benchmark scores, but on whether deployments deliver measurable productivity gains without pushing project budgets higher.

The first lever — the agent harness — is the software layer that controls what information a model receives, how tasks are broken down, and how frequently the model is called. Inc42 points to Sarvam AI's late-July launch of Sarvam Code as an illustration: the coding agent divides work between planner, worker, and verifier agents and completed tasks on Terminal-Bench 2.1 at an average cost of about $2, compared with $4.1 to $27.8 for Claude Code and OpenAI's Codex. Sarvam also said its system, which routes work between its 105B model and a larger open model, reduced serving costs by roughly 40%. Kausal Malladi, CTO (investments) at INDmoney, told Inc42 that a task running 40 steps can consume millions of tokens even when the original request contained only a few thousand, because the agent repeatedly sends prior context back to the model. He estimated that prompt caching alone — reusing already-processed context rather than resending it — can cut effective costs by roughly 80%.

The second lever is model selection. Rather than routing every request to a frontier model, enterprises are sending routine tasks to smaller, cheaper models and reserving expensive ones for genuinely complex work. Krupesh Bhat, founder and CEO of Melento (the document-automation company formerly known as SignDesk), told Inc42 that the goal is to match the cost of the model to the difficulty of the task. He suggested tracking three metrics to verify the approach is working: the share of requests handled by smaller models, average cost per transaction, and task success rate. If cheaper models absorb more of the workload while success rates hold steady, overall project costs fall without sacrificing quality.

The third lever, inference economics, goes further by removing AI from workflow steps that do not require it at all. Arjun Nagulapally, CTO at AIONOS, an agentic AI startup, told Inc42 that account lookups, eligibility checks, and field validations can often be handled by a database or a business rule rather than a model call. In one project for a telecom operator processing more than a million interactions a month, AIONOS moved roughly 30% of routine steps to conventional software, used an open-weight model trained on resolved customer queries, reused outputs where possible, and shifted some compute to more cost-effective capacity. The total cost per resolved interaction fell by 35%, with the largest single saving coming from the work that stopped being an AI problem altogether.

The broader pattern Inc42 identifies is that lowering AI costs is fundamentally an engineering discipline rather than a procurement decision. Enterprises can also batch requests to keep compute utilization steady and avoid paying for idle capacity between jobs. The article notes that the distinction between productivity and savings matters: higher throughput does not automatically mean lower spending unless it translates into reduced model costs, reduced headcount on rework, or both. The design choices made before the first production bill arrives, the piece concludes, determine whether AI deployments scale economically or spiral out of control.