Enterprises face a reckoning in their AI costs as budgets come under strain

As AI adoption accelerates, companies are turning to model orchestration, smaller language models and token monitoring to control rising costs and improve returns on AI investments.

In April, Uber exhausted its entire 2026 budget for artificial intelligence (AI) after pushing engineers to adopt the technology faster. Adoption of Anthropic’s Claude Code — first rolled out in December — jumped to 84 per cent in March, up from 32 per cent just a month earlier. What the company seemingly failed to anticipate was a growing disconnect between the push for higher token usage and the teams controlling the spending, triggering an emergency slowdown.

 

Similarly, Microsoft cut employee spending on tokens in August and began monitoring consumption to ensure usage delivers meaningful outcomes. Meanwhile, Bloomberg reported that Apple is lowering daily employee spending limits from $300 to $200, with plans for further cuts to $100.

 

Expense cap

 

High costs are prompting enterprises to adopt model orchestration — the process of assigning tasks to the right AI model to achieve desired outcomes at a reasonable cost. This shift comes as massive AI investments fail to yield adequate returns, adoption remains low, and boards and chief financial officers are wary of expenses.

 

“Many enterprises are currently capping AI spend, given elevated AI costs,” said Goldman Sachs in its Connecting Dots report, drawing on examples from companies across technology, retail, telecom and financial services.

 

Model orchestration is an automated control layer that coordinates and routes tasks across multiple AI models, tools, application programming interfaces (APIs) and data sources. Instead of relying on one generalist model for every step, it separates high-level planning from specialised execution.

 

Enterprises woke up to smell the coffee in 2026: Corporate exuberance about AI is fine only as long as it justifies the cost. But when millions spent fail to adequately boost the top or bottom line, companies must return to the drawing board, reassess their gaps and fix them.

 

Model orchestration has emerged as a key priority in the AI world as enterprises discover that unguided token usage does little for efficiency or productivity — it simply burns through annual AI budgets faster than expected. In response, companies are turning to a mix of large, small and open-source models tailored to specific tasks, driving better results while reining in costs.

 

An AI model works best for an enterprise when it understands company context, pulls data from multiple sources and applies that knowledge to complete its assigned tasks. “These models are very knowledgeable about the world but they have no idea about your context, your company, your terminologies, and your acronyms,” says Srikanth Velamakanni, cofounder and group chief executive officer (CEO) of enterprise AI firm Fractal Analytics.

 

“So, bringing your context knowledge effectively to work with an AI model, that’s the first thing that you have to do in order to make them very effective. Frontier models can be extremely expensive. If you don’t use them correctly, it’s like using a bazooka to kill a fly. It won't be a good use of the ammunition. So you have to use the right ones for the right problem,” he explains.

 

Experts say companies should reserve frontier models for work requiring extreme speed and quality, and where cost is no concern. Ultimately, enterprises must balance the economics so that the value AI generates far outweighs its cost. The question remains: How do companies choose the right model for the job?

 

Using open-weight or small language models trained on proprietary, contextual data can slash costs by tenfold while improving accuracy, according to experts.

 

While experts advocate smaller, specialised models to cut costs, Google argues that some of its models are so capable that customers won't need to upgrade to a higher tier. 

 

 

“The best model is the smallest model that drives the outcome you need and if you look at models that we have like Gemini Flash, it’s an extremely capable model compared to the cost of it. Most of our customers don’t even need to use Gemini Pro because Gemini Flash is so capable,” Andi Gutmans, vice-president and general manager, data cloud, Google Cloud, told Business Standard recently.

 

Context matters

 

“Context will matter a lot. The less context you have or the less good your context is, the more reasoning loops the agent will do and the more tokens it’s going to consume and the more compute it is going to consume,” he said.

 

Google calls this concept a “knowledge catalogue” — a system providing differentiated context so agents receive only the minimum information needed to reach an outcome. By minimising reasoning loops, it cuts down compute time and lowers overall costs.

 

Velamakanni points to an Australian bank — a client of his firm, Fractal — as an example. “They launched a product which was extremely successful with their customers but they launched it for about 30,000 of their customers and they wanted to take it to about 2 million. Because they were bringing in a phenomenal model, which is very good, but the model was not fully trained in the context of the organisation itself. What we suggested was to bring in a potentially open model, post-train it with their data, so that it becomes a faster and cheaper model.”

 

Since many services for large language models (LLMs) — AI systems trained on vast datasets — charge by token usage, tracking tokens helps enterprises understand the true financial and operational cost behind every prompt, response, retrieval and automated workflow. Experts recommend implementing dynamic routing: sending lower-complexity tasks to low-cost models, while reserving advanced, token-heavy workloads for frontier LLMs.

 

Model orchestration is becoming a critical part of every client conversation, according to LTM, an IT services company. Infosys, another IT firm, offers a similar solution for its clients: based on query complexity, the system automatically selects the right backend model to optimise costs.

 

“We create a skill to make sure that tokens are worth the transaction automation. At Infosys, we give a specific budget to everybody so that people are more AI-aware and they are believers in AI,” said Sajit Vijayakumar, CEO of Infosys Finacle.

 

“For clients, we have a hierarchy of models. There are open source, locally deployable models where we just need a GPU as a capital investment. If a certain task cannot be performed with that model, that framework escalates automatically to the next model. This is what the whole industry is doing to optimise cost,” he said.