Enterprises are slowly but steadily moving beyond AI pilots and into real-world production, but the journey is far from uniform across sectors. However, Prashanth Kaddi, Partner, Deloitte India, says AI adoption today is shaped as much by organisational readiness as by technology maturity.
In an exclusive interaction with AIM, he shares a grounded view of how generative and agentic AI are being deployed across industries, the bottlenecks holding enterprises back, and why AI should be seen as a growth engine rather than just a cost-cutting tool.
On whether there’s an AI bubble, Kaddi is unequivocal. “AI is not a bubble. It’s real, and it’s going to change the world.”
That said, he separates technological impact from financial exuberance. “The bubble narrative is driven by valuations,” he says, adding that whether they are justified is another matter altogether.
He views investments in data centres and infrastructure in India a net positive. “Those facilities are here to stay and will be used by India. If we want to grow from a $4-trillion economy to $10 trillion and beyond, we will need foreign capital, and AI infrastructure can be one of the means of enabling that. ”
Production is Catching Up
While experimentation remains widespread, a meaningful share of enterprises have begun production deployments, especially with generative AI.
According to a 2026 report by Deloitte, only 25% of organisations have moved 40% or more of their AI experiments into production, but that figure is expected to rise to 54% within the next three to six months.
Kaddi notes that the adoption of agentic AI is higher largely because of a more selective approach to use-case selection. That selectiveness, he says, comes from lessons learned in previous AI cycles, where business cases often failed to hold up once pilots scaled to production.
“After the pilot, when we go into production, there is a significant deviation. So we try to plug that upfront and make the right choices. Then the path to production is relatively easy,” he says.
Why AI Adoption Still Stalls in Enterprises
Despite growing momentum, adoption remains challenging. Kaddi points to multiple layers of friction that organisations must navigate before AI can meaningfully show up on the balance sheet.
“The first challenge is identifying the right use case. It should have enough meat to justify both the value creation and the running cost,” he says. Too many organisations, he adds, start with lightweight applications that fail to move the needle.
The second challenge lies in embedding AI into core business processes. “Unless the business signs off on the post-agentic process and is willing to adopt it, the ROI never really comes to the P&L. That buy-in upfront is critical.”
Technology choices form the third bottleneck. While experimentation has been democratised across enterprises, many proofs of concept are unscaleable. “In the organisation, everybody’s trying to create something which could be quick and dirty to create, but then actually deploying it in a reliable, scalable, actual productionised environment is a different thing.”
Finally, skill and talent gaps persist. “People do not understand how AI is going to function in the organisation, what role it should have to play,” Kaddi says, adding that AI needs to be embedded into core business problems and everyday processes to deliver real impact.
Which Sectors Are Moving Fastest on Agentic AI
Rather than resisting AI, BFSI institutions are adopting it with a strong risk lens, prioritising areas where experimentation does not compromise trust or compliance. According to a Scielo Brazil study of BFSI firms in India, 60–70% of financial institutions actively use AI.
“BFSI has been quite open, especially for internal processes. Customer-facing use cases are more constrained, but internal operations are definitely seen as sources of competitive advantage,” Kaddi notes.
He cites relationship management as one such area. “If a relationship manager has 100 or 200 clients, there’s no way he’s going to sort of be able to be personal about each one of them. AI sales assistants can be more effective here.”
Consumer products and retail have also been early movers, driven by complex supply chains, distribution networks, and marketing needs. Media, telecom, and technology companies, he highlights, are particularly receptive due to their digital-first nature.
Traditional industries are moving too, but with different objectives. “In life sciences, it’s about shortening the discovery-to-production lifecycle. You may not crash research time, but regulatory reporting and data synthesis can be significantly accelerated,” Kaddi says. In manufacturing and automotive, AI is increasingly seen as essential to compress design cycles and remain competitive.
More Than Cost-Cutting
One of the biggest misconceptions around enterprise AI, Kaddi argues, is its framing as a cost-saving tool.
He asserts that AI creates value across multiple dimensions, with revenue growth being the most critical. It also drives new business models, large-scale customer personalisation, sharper market segmentation, and smarter capital allocation decisions.
Beyond resource optimisation, the Deloitte executive says bigger savings in many industries come from improving procurement, raw materials, and manufacturing costs, where AI can make operations more efficient.
Asset-heavy industries also stand to benefit through improved asset utilisation and lifecycle management.
India’s AI Spending Gap
Compared to the US and China, India’s AI spending remains modest. But Kaddi sees this less as a weakness and more as a strategic posture.
He says India is a major consumer of AI outputs and prefers a value-driven approach over heavy upfront investment. “We are significant consumers of AI outputs,” he says, adding that the Indian mindset focuses on jugaad-style solutions that leverage globally available technologies to extract maximum value for money.
However, he disagrees with the view that India should completely avoid model development, arguing that India at least needs significant fine-tuning because its cultural and language context is unique, and models trained predominantly on Western data cannot simply be lifted and used.
Rather than chasing frontier models immediately, Kaddi suggests a more pragmatic path. “We can start small. There are a number of people who have taken open-source models and customised them for local languages, and maybe we can start on that path, build some SLMs, for instance. Actually building frontier models may be more of a distant roadmap.”
Can Enterprises Keep Up?
With new models and frameworks emerging every few months, enterprises often worry about sunk costs. Kaddi acknowledges the concern but dismisses the idea that current investments will become useless.
“Some fundamental concepts don’t change,” he said. “The smartness lies in building modular, agile architectures where models can be swapped out when better options emerge.”
For Kaddi, the future of enterprise AI is less about hype and more about disciplined execution. “Everything does not require AI,” he says. “But the problems that do require it need to be solved in the right way—scalable, enterprise-wide, and sustainable.”
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.