At TechSparks 2026, Intuit's Principal AI Scientist Vignesh Subrahmaniam delivered a pointed message to India's startup and technology community: the era of AI that merely responds is giving way to AI that acts, and that shift will reward builders who focus on reliability, governance, and trust rather than novelty alone.
Subrahmaniam's remarks came at a moment when the industry is moving beyond conversational interfaces and into agentic systems that can complete tasks, invoke tools, and make decisions inside enterprise workflows. That transition is already reshaping how companies think about product design, data pipelines, and risk management. In his view, the most important question is no longer whether a model can generate a fluent answer, but whether it can safely execute a real-world action with predictable outcomes.
Tools Over Prompts
Subrahmaniam framed the current AI moment as a shift in emphasis from prompt engineering to tool building. In the early wave of generative AI, much of the excitement centered on how users could coax better responses from large language models through carefully crafted prompts. But as AI systems become more operational, he suggested, the bottleneck is moving away from wording and toward infrastructure.
That infrastructure includes the tools an AI system can call, the permissions it has, the quality of the data it can access, and the checks that prevent it from taking the wrong action. In practical terms, this means enterprises will need to design AI systems that are not just intelligent in isolation, but integrated into workflows where they can retrieve information, update records, trigger processes, and escalate uncertain cases to humans.
The implication for startups is significant. Many early AI products have been built around chat interfaces and productivity shortcuts. Subrahmaniam's argument suggests the next generation of winners may be those that build the plumbing behind the interface: secure APIs, robust orchestration layers, and domain-specific tools that make AI dependable enough for business use.
Trust Needs Guardrails
A central theme in Subrahmaniam's comments was that AI acting on behalf of users raises the stakes on consent and control. When a system is only answering questions, the consequences of error are often limited to misinformation or inconvenience. When it is acting, the consequences can include financial mistakes, operational disruption, or privacy violations.
That is why, he argued, consent cannot be treated as a formality. Users must understand what an AI system is allowed to do, what data it can access, and when it needs approval before proceeding. For enterprises, this means building explicit guardrails into product architecture rather than relying on policy documents alone.
The emphasis on consent also reflects a broader industry reckoning. As AI tools become more autonomous, regulators and customers alike are asking harder questions about accountability. Who is responsible when an agent makes the wrong call? How much autonomy is appropriate in a given workflow? And how can companies prove that sensitive data is being handled responsibly? Subrahmaniam's answer, at least in spirit, was that trust must be engineered, not assumed.
Data Quality Becomes Strategy
Subrahmaniam also underscored the importance of high-quality data, a point that has become increasingly central as companies attempt to move from demos to deployment. In the consumer imagination, AI progress is often measured by model size or benchmark scores. In enterprise settings, however, performance depends heavily on the relevance, freshness, and integrity of the data feeding the system.
Poor data can undermine even the most advanced model, especially when the AI is expected to take action rather than simply summarize information. If the underlying records are outdated, incomplete, or inconsistent, the system may confidently produce the wrong result. For that reason, data quality is no longer a back-office concern; it is a strategic asset.
For India's startup ecosystem, the message lands with particular force. The country's AI builders are operating in a market where enterprise adoption is accelerating, but where buyers are also increasingly cautious about reliability, compliance, and measurable return on investment. Subrahmaniam's framing suggests that the companies most likely to endure will be those that treat data governance, permissioning, and workflow design as core product features rather than afterthoughts.
His comments also reflect a broader maturation in the AI sector. The industry is moving from experimentation to operationalization, from prototypes to production, and from novelty to utility. In that environment, the most valuable AI may not be the one that sounds the smartest, but the one that can act with precision, restraint, and accountability.
For founders and investors gathered at TechSparks 2026, the takeaway was clear: the next frontier of AI will not be won by better prompts alone. It will be won by systems that can safely do work, with the right tools, the right data, and the right permissions.