At TechSparks 2026, Intuit Principal AI Scientist Vignesh Subrahmaniam delivered a pointed message to India's startup community: the era of treating artificial intelligence as a clever conversational layer is ending, and the harder work of building reliable systems that can act is beginning. Speaking in the context of enterprise AI adoption, he argued that the industry is moving from prompt engineering to tool building, a shift that places far greater weight on data quality, system design and user consent.
The distinction matters. Prompt engineering, once celebrated as a new technical craft, is increasingly seen as a transitional skill in a market where large language models are becoming more capable and more commoditised. What will separate durable products from fleeting demos, Subrahmaniam suggested, is not the ability to ask a model the right question, but the ability to connect it to trustworthy tools, structured data and guardrails that make it useful in production.
Tooling Over Prompts
Subrahmaniam's core argument was that AI is moving from a passive assistant model to an active agent model. In the first phase, systems answered questions, summarised documents and drafted text. In the next phase, they will book, file, reconcile, recommend and trigger actions across software environments. That transition raises the stakes dramatically. A system that merely generates an incorrect answer can be corrected; a system that takes the wrong action can create financial, operational or compliance damage.
For that reason, he emphasised that the real frontier is not prompt design but tool reliability. AI systems must be able to call functions, access approved data sources and operate within defined constraints. In enterprise settings, this means the model is only as good as the tools it can use and the workflows it is embedded in. The implication for startups is significant: product differentiation may increasingly come from orchestration layers, domain-specific integrations and workflow intelligence rather than from the base model itself.
Data And Consent Matter
Subrahmaniam also highlighted the importance of high-quality data, a theme that has become central as companies attempt to deploy AI at scale. In consumer-facing products, messy or incomplete data can lead to hallucinations or irrelevant outputs. In enterprise environments, the consequences are more serious, because the model may be making decisions or recommendations that affect money, customers or regulated processes.
He paired that concern with a strong emphasis on consent. As AI systems become more autonomous, users must know what data is being accessed, what actions are being taken and what permissions have been granted. Consent is no longer a legal afterthought; it is a product requirement and a trust mechanism. That framing reflects a broader shift in the AI market, where companies are being pushed to design for transparency rather than assuming users will tolerate opaque automation.
The point is especially relevant in India, where startups are racing to embed AI into fintech, commerce, customer support and productivity software. In these categories, the ability to act on behalf of a user can create real value, but only if the system is auditable, permissioned and dependable. Without those controls, agentic AI risks becoming a liability rather than an advantage.
What Startups Must Build
For founders and investors, the message from TechSparks was less about model hype and more about execution discipline. Subrahmaniam's remarks suggest that the next wave of AI winners may be those that build the plumbing: secure tool layers, clean data pipelines, permission frameworks and domain-specific controls. These are not the most glamorous parts of the stack, but they are the parts that determine whether AI can be trusted to operate in the real world.
That shift also changes how venture capital may evaluate opportunities. In the earlier wave of AI adoption, speed to prototype and user-facing novelty often attracted attention. In the emerging phase, investors are likely to scrutinise data moats, integration depth, governance architecture and the ability to reduce error rates in live workflows. The market is maturing from experimentation to accountability.
Subrahmaniam's remarks capture a broader inflection point for the sector. As AI becomes more capable of acting, the industry's centre of gravity is moving away from prompt tricks and toward systems engineering. The winners, he suggested, will be the builders who can make AI not just intelligent, but reliable, permissioned and operationally useful.