.png){kind=logo}
Most AI pilots fail not because of weak models, but because of missing data foundations and governance frameworks that only surface at scale.
Fragmented knowledge bases, not the AI itself, are usually behind inconsistent or hallucinated answers from enterprise AI assistants.
India's Global Capability Centers are moving from execution-focused units to strategic hubs that proactively solve business problems using AI.
Enterprise AI adoption is accelerating, but the majority of AI pilots never reach full production. Industry research consistently finds that more than 80% of enterprise AI projects stall between proof-of-concept and production, a trap practitioners call "pilot purgatory."
In this exclusive Analytics Insight podcast, Priya Diyalani speaks with Anand S., LLM Psychologist and Head of Innovation at Straive, about why this failure pattern is rarely about the technology itself, and why content and knowledge infrastructure is becoming as critical as cloud and data platforms.
The shortest way to explain Straive is how our CEO puts it: we build AI and we run AI. Building is the exciting part, but running it means connecting to messy data, fitting it into real workflows, and making sure it doesn't break when other systems fail. We call that operationalizing AI.
I lead innovation, which mostly means I spend my day poking at LLMs to see what they can and can't do. I picked "LLM psychologist" as a title because that's basically what I do, though a friend once told me "LLM psychopath" might be more accurate given how much I stress-test these models.
A project has three parts: figuring out what to do, building it, and making sure it actually works. AI has made building dramatically faster, especially coding, so leaders assume the whole project speeds up equally. It doesn't.
A six-month project might compress its middle stage into two days, but the remaining four months of testing and deployment still take four months. I wouldn't even call this a misconception, since even AI experts can't reliably predict which tasks will speed up next.
Because fragmented knowledge causes the same problems for AI that it causes for humans. If a call center agent can't access the right system, they can't help you.
AI behaves the same way, except it's trained to be maximally helpful, so instead of admitting uncertainty, it sometimes invents an answer.
We've had agents scan our entire Google Drive and summarize every document automatically, creating a lightweight index other agents can use.
Agents can even review old customer conversations to spot where things went wrong and suggest better documentation, while a second agent tests whether that actually improves results.
India may not be building the top foundation models, but it's strong in services, and GCCs represent that well. We're seeing a new role emerge, the "forward deployed engineer," who works alongside business teams and proactively spots problems, like the one who caught 104,000 transactions billed at a fraction of the correct price.
A client once wanted to scrape 15,000 websites, once considered too costly to maintain. AI made building those scrapers cheap, but keeping them accurate as websites change remains an ongoing challenge.
