The Enterprise Fight Against Runaway AI Costs

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.

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?

Making AI Do More With Less

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. Sarvam said the system 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 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.

Kausal Malladi, chief technology officer (investments) at INDmoney, says much of an AI agent’s cost comes from the amount of information it sends to the model.

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.

Frontier Intelligence Needs A Filter

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.

“The idea is not to replace capable models indiscriminately, but to match the cost of the model to the difficulty of the task,” said Bhat. 

To determine whether this approach is lowering project costs, he suggested that enterprises should track:

• Requests Handled By Smaller Models: A rising share indicates that frontier models are being reserved for genuinely complex work.
• Average Cost Per Transaction: This shows whether the overall cost of completing a task is falling.
• Task Success Rate: This confirms that lower costs are not coming at the expense of quality or increased human intervention.
  

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.

Using Inference Economics To Cut The Bill

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.

Arjun Nagulapally, CTO at AIONOS, an agentic AI startup, said hefty savings can come by removing AI from parts of a workflow where it is not needed. A customer account lookup, eligibility check or field validation can often be handled by a database or a simple business rule instead of a model. 

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%.

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. They can also use more efficient versions of models that require less computing power while still handling routine tasks effectively.

The broader lesson is that lowering AI costs depends on how enterprises design, run and scale their AI systems. It is an engineering discipline that begins before the first production bill arrives.  

Top Stories From India & Around The World

• Peak XV Leads $45 Mn Bet On Blacksmith: The VC major led a $45 Mn Series B in US-based AI code testing startup Blacksmith at a $550 Mn valuation. GV and Y Combinator also participated, taking its total funding to $58.5 Mn as AI-generated code drives demand for testing infrastructure.
• Razorpay’s ChatGPT Gambit: Razorpay is building infrastructure for merchants to create ChatGPT storefronts and accept payments inside conversations. It is working with around 50 brands and 200 businesses on agentic commerce and Agent Studio, using UPI Circle and UPI Reserve Pay for controlled checkout.
• L&T’s Data Centre Bet: L&T is consolidating its data centre and cloud businesses under Vyoma.AI after transferring them for ₹1,400 Cr. Vyoma plans a 10,000-GPU NVIDIA B300 AI factory in Chennai by November, while evaluating a 100 MW campus in Maharashtra.
• Grok 4.6 Targets Long-Running AI Agents: xAI has released Grok 4.6 with a focus on multi-step agents, coding and interactive applications. The model is available through Grok Build, Cursor and its API, with pricing starting at $2 per million input tokens and $6 per million output tokens.
• Cursor To Open First India Office Soon: The US-based AI coding company, recently acquired by SpaceX, is preparing to open its first India office by December. India is its third-largest market, with over 3 Mn developers and the highest concentration of power users, intensifying competition with OpenAI, Anthropic and Sarvam.
• Anthropic’s Watermarks Move: The frontier AI model developer will embed machine-readable watermarks and metadata in content generated across Claude’s API, chatbot, Claude Code and enterprise deployments worldwide. The rollout is applicable to models launched in the EU after August 2, 2026.

The Weekly Buzz: Apple’s China Coup

Apple has trained a proprietary large language model specifically for the Chinese market with technical support from Alibaba, according to a report. Beijing has cleared the model, making Apple the first foreign company approved to offer its own generative AI system to users in mainland China. The China-focused model marks a clear departure from Apple’s previous reliance on third-party systems. 

It will power elements of Apple Intelligence on devices sold in the country and is expected to roll out in the coming months alongside an iOS update. The approach uses a dual-track setup that pairs Apple’s own model with Alibaba’s Qwen technology for generation and Baidu components for search, giving the company greater control over the AI experience in one of its most competitive and tightly regulated markets.

The breakthrough demonstrates that even under China’s strict cybersecurity and content rules, a US tech giant can still build and ship proprietary frontier AI capabilities on the ground. It opens a promising path for Apple to close the feature gap with local rivals such as Huawei and to retain tighter influence over how its devices behave for hundreds of millions of users. At the same time it raises sharp questions about dual standards in AI governance, data flows across geopolitical lines, and whether other Western firms will now push for similar bespoke arrangements or face even higher barriers.

 Startup In The Spotlight: Curium Life 

Founded in 2023 by Vinayak S Rengan and Balachandran Seetharam, Bengaluru-based Curium Life is building a surgical intelligence platform designed to bring real-time AI into laparoscopic and robotic procedures. As surgical technology advances, many intraoperative decisions still depend heavily on what the surgeon can see and interpret in the moment.

Instead of building another standalone surgical device, Curium is positioning its technology as an AI layer for existing surgical hardware. The startup is working with laparoscopic and robotic surgery OEM partners, giving them access to intraoperative perception and guidance that can be integrated into their platforms.

The startup is targeting the growing AI-in-healthcare market, estimated at $50 Bn globally, alongside the surgical AI assistance software market, projected to reach $11.8 Bn by 2034.

As AI moves deeper into clinical workflows, Curium believes real-time surgical intelligence could help transform operating rooms from environments driven primarily by the surgeon’s visual judgement into systems where AI continuously interprets what is happening and provides relevant guidance during procedures.

Prompt Of The Week

What prompts and hacks are CTOs, CEOs and cofounders using these days to streamline their work? 

Here’s the prompt founder and CEO Siddhartha Chandurkar of Shephertz, uses to get advise on scaling a SaaS company:

“You are an organisational design advisor to a rapidly scaling SaaS company. 

Here is our situation: [Describe headcount growth, leadership team, current layers, span of control, geography, and problems emerging as we scale.] 

I need you to: 

• Identify which problems are caused by insufficient structure versus poor execution 
• Assess our current management layers and spans of control 
• Identify where leadership capacity is becoming a bottleneck 
• Recommend where to add management layers and where NOT to 
• Define which decisions should move closer to teams as we scale 
• Identify processes that are becoming unnecessary bureaucracy 
• Propose an operating model that preserves speed while increasing accountability 
• Define the organizational signals we should monitor over the next 6–12 months 

Do not assume that adding managers or processes automatically improves scalability.”

Editor’s Note: Some prompts may need to be adjusted by users for best results or may not work as intended for certain users.

[Edited by Shishir Parasher]
[Creatives by Varshita Srivastava]

The post The Enterprise Fight Against Runaway AI Costs appeared first on Inc42 Media.