E-Paper
When AI acts on bad data, someone gets hurt
This article is authored by Roberto Sigona, chief operating officer, Qlik.
Picture a bank that hands credit decisions to an AI model. The model is state-of-the-art. The data feeding it isn't. Its fragments pulled from siloed systems that haven't been reconciled in months.
India's AI ambition is real. The country has moved past pilots into scaled deployment, and speed is the priority: 91% of leaders name deployment speed as the deciding factor in build-versus-buy decisions, according to EY's Agentic AI in India report. But speed and readiness are not the same thing. BCG's Widening AI Value Gap found that only 5% of organisations worldwide are generating meaningful value from AI at scale. The bottleneck was never the model. It's everything underneath it.
For two years, the AI risk conversation was about what models say hallucinations, bias, an embarrassing chatbot reply. Agentic AI ends that conversation. EY finds 24% of Indian leaders are already deploying it, and these systems don't wait for a human to check their work. They act triggering transactions and workflows on their own, at scale. The danger isn't AI saying the wrong thing anymore. It's AI doing the wrong thing, repeatedly, before anyone is watching. The margin for bad data has shrunk to zero.
Most enterprise data infrastructure was never built for this. It was built for reporting: periodic, structured, checked by a person before it mattered. Bolting agentic AI onto that foundation doesn't upgrade it; it exposes it. Indian enterprises grew through acquisition and multi-cloud sprawl, so fragmentation isn't an accident; it's the architecture. EY reports that 78% of Indian enterprises struggle to integrate their systems, and 64.5% call data governance and security their top barrier to scaling AI. Finance, HR, supply chain, customer data scattered across systems with no single source of truth. Every AI initiative starts as a clean-up job in disguise.
A pilot works because the environment is controlled: Clean data, narrow scope, a human watching closely. Scaling it into live systems is a different test, and most projects fail it. EY found nearly half of Indian enterprises push over 21% of their proofs-of-concept into production, yet most still stall before they scale. When that happens, the model isn't the problem. The infrastructure underneath it never was built to carry the weight.
The instinct is to choose sides: A multi-year overhaul before touching AI, or quick, siloed pipelines to meet immediate needs. ISG's State of Enterprise AI Adoption Report 2025 is blunt: Neither works. What works is treating data readiness as a parallel track, not a prerequisite that delays AI, and not an afterthought that undermines it.
The regulatory clock is also running. India's Digital Personal Data Protection Act reaches full enforcement in May 2027, with penalties up to INR 250 crore per violation. The RBI's FREE-AI framework already requires financial institutions to have board-approved AI policies. Companies with fragmented data are the ones least able to show regulators where a decision came from and that gap will get expensive.
The enterprises actually scaling AI successfully share three traits: Connected data instead of silos, governance built end-to-end, and open, interoperable architecture. Many have stopped moving data to the AI and started bringing the AI to the data, using data mesh and fabric models that cut out duplication. None of this requires infrastructure perfection. It requires treating data readiness as a priority, not a formality.
India has the ambition, the investment, the talent; the country ranks third globally in AI competitiveness, per the Stanford AI Index. Whether its data can carry that ambition is a separate question, and the answer isn't yet yes. The winners of the next five years won't be the companies that deployed AI first. They'll be the ones that made their data trustworthy enough to act on.
The models are ready. The agents are already live. The data underneath them is not and every quarter that gap stays open, it gets more expensive to close.
(The views expressed are personal)
This article is authored by Roberto Sigona, chief operating officer, Qlik.
