The enterprise AI gold rush is giving way to a more disciplined phase. After two years of racing to deploy the biggest models and consume more compute, companies are asking a simpler question: is every AI rupee generating measurable business returns? According to Business Today, Uber recently acknowledged that internal AI coding budgets were exhausted much faster than expected as employee adoption surged, prompting tighter governance. Similarly, Meta has made reducing inference costs a strategic priority, while Amazon, Walmart, Cisco and Uber have introduced usage caps or routed employees to cheaper AI models to contain costs.
The shift to accountability
The push for accountability is becoming visible in India as well. An EY-CII survey found that 47% of Indian enterprises now have multiple generative AI applications in production, while more than 95% still keep AI and machine learning budgets below 20% of overall IT spending. A separate study by SAP found that Indian organisations expect AI investment to rise 45% over the next two years, even as the focus shifts towards improving returns rather than simply increasing deployments.
| Metric | Value |
|---|---|
| Indian enterprises with multiple genAI apps in production | 47% |
| Firms keeping AI/ML budgets below 20% of IT spend | >95% |
| Expected AI investment increase in India (next 2 years) | 45% |
| Mature AI adopters that reduced BPO spending | 90% |
| Cut outsourced work by >25% (among those reducing) | >33% |
Budgets and returns
“We’re seeing diminishing returns becoming more apparent,” said Sambhav Jain, managing director and partner at BCG India, as reported by Business Today. Larger context windows, multiple AI agents and expensive inference often deliver only marginal improvements in output quality without corresponding gains in productivity, revenue or customer experience, he said. Deepak Dhanak, co-founder and chief operating officer of Rocket — an AI-native platform, said the problem is not AI spending itself but spending without accountability. “The overspend is not an AI problem, it is a measurement problem,” Dhanak said. “Token consumption became a vanity metric during first wave of enterprise AI. Businesses now need to match the right model to the right task and measure outcomes, not activity.”
The measurement problem
The rethink comes even as AI investments accelerate. According to Business Today, Microsoft, Amazon, Alphabet and Meta were expected to spend roughly $320 billion on AI infrastructure in 2025. In India, a Z47-OpenAI-Zinnov study found that nearly 90% of mature AI adopters have reduced some form of BPO spending, with over one-third cutting outsourced work by more than 25%. The same study found that 86% of Indian startup founders plan to increase AI budgets this year and more than half expect to more than double spending, yet only 9% have seen a measurable increase in sales or conversions attributable to AI. “If you cannot draw a straight line from spend to value derived, you are not measuring ROI — you are measuring noise,” Dhanak said.
Operationalisation
Executives say enterprises are increasingly asking whether cheaper models can deliver 95% of the value at 20% of the cost. “We’re not experimenting with AI anymore — we’re operationalising it. That shift changes the entire cost conversation,” said Milesh J, head of strategy and operations, SAP Labs India. “The real risk isn’t overspending on tokens. It’s mistaking spend for strategy — and not having the measurement discipline to know which AI investments are actually paying back.”
Sources:
