For the past few years, artificial intelligence has lived in a strange duality. On the surface, it looked magical — conversational systems, image generators, copilots that seem to reason. Underneath, enterprises have struggled with a more sober reality: hallucinations, brittle workflows, governance risks, and systems that work impressively in demos but falter at scale.
In my recent conversation with Dwarak Rajagopal, Head of AI Research at Snowflake, what stood out was not optimism or scepticism, but maturity. AI, he suggested, is finally moving out of its adolescence. And the changes ahead, in 2026, will be less about bigger models and more about how intelligence is structured, verified, and distributed.
One of the most consequential shifts underway is the quiet erosion of monopoly power in foundation models. For much of the current AI cycle, a handful of companies have dominated the narrative, capital, and capabilities around large models. But Mr. Rajagopal pointed to a critical inflection: the centre of gravity is moving from pre-training to post-training.
The biggest breakthroughs are no longer coming from simply scaling parameters, but from how models are refined with specialized, high-quality data. This is opening the door for open source foundation models that can be customized, fine-tuned, and deployed for specific enterprise needs.
Weakening grip of a few AI giants
In 2026, this will fundamentally weaken the grip of a few AI giants and enable a more distributed innovation ecosystem where startups, researchers, and enterprises build on shared foundations rather than proprietary silos.
This shift has deep implications for how intelligence is deployed inside organisations. Enterprises do not need general intelligence; they need contextual intelligence. They need systems that understand their data, their workflows, and their constraints. Open source models, combined with enterprise data platforms, make that possible. AI is no longer being treated as an oracle. It is being treated as infrastructure.
The next major frontier, according to Mr. Rajagopal, will not be another leap in raw model capability but advances in agentic AI, driven by improvements in context windows and memory.
Today’s models remain constrained by working memory. They are powerful in single interactions but fragile across long, multi-step tasks. In 2026, innovation will focus on giving agents persistent memory that will let them retain longer contexts, learn from prior actions, and operate across extended timelines.
This will allow AI systems to move beyond reactive assistance and toward sustained problem-solving, supporting complex business processes rather than isolated queries.
Automation won’t be that easy
Yet autonomy brings risk. One of the biggest bottlenecks in deploying agentic systems today is error accumulation. Small mistakes compound across multi-step workflows, forcing enterprises to rely on human oversight as a safety net.
Mr. Rajagopal’s view is that this dependency will not scale. The breakthrough will come from self-verification. Instead of inserting humans into every loop, AI systems will develop internal feedback mechanisms. They will be able to judge, validate, and correct their own outputs.
These “auto-judging” agents will mark a shift from supervised automation to self-regulating systems, making large-scale, reliable agentic workflows viable for enterprise use.This evolution in AI capability will inevitably reshape the human side of technology as well.
Plumbing organisational hierarchies
One of the most striking implications Mr. Rajagopal outlined is how AI will redefine the software engineering hierarchy itself. As agents take over routine coding and repetitive tasks, the traditional ladder, in which juniors handle low-level implementation and seniors design systems, will blur.
Junior engineers will be pushed up the stack, working earlier on infrastructure, integration, and system-level thinking. Senior engineers, meanwhile, will become orchestrators who design architectures, mentor teams, and ensure humans and AI systems collaborate effectively.
This will force a rethink of computer science education, career paths, and what it means to be “experienced” in a world where execution is increasingly automated.
Threaded through all of this is a geopolitical undercurrent. The U.S. style of AI development has become increasingly corporatised and closed, driven by scale, capital intensity, and proprietary advantage. China, by contrast, is leaning more heavily into academic and open-source approaches, integrating research and state-backed ecosystems.
India at crossroads
India, meanwhile, sits at a crossroads. It has the talent, the data, and the scale. But its position in the global AI race will depend on whether it invests in open foundations, enterprise-grade data systems, and skills that go beyond prompt engineering, and toward systems thinking.
The picture that emerges is not one of runaway intelligence, but of disciplined intelligence. Enterprise AI in 2026 will be less about spectacle and more about reliability. Less about isolated models and more about systems. Less about human-in-the-loop as a crutch, and more about self-verifying agents designed for trust.
If the last phase of AI was about proving what machines can do, the next phase will be about proving what organiations can build with them. That, more than model size or benchmark scores, will determine who truly leads in the age of artificial intelligence.
Published – December 24, 2025 09:21 am IST
