Key Insights
The Model Is 10% of the Equation — Engineering Carries the Rest
Enterprises rarely fail at AI because they picked the wrong model. They fail because of stale data, unreliable inputs, no traceability, no observability, and no clear measure of outcome. As models become commoditized, the differentiation has shifted from the model itself to the engineering around it. A useful rule of thumb: roughly 10% of success comes from selecting the right model, 20% from infrastructure, and 70% from the engineering layer that makes deployment safe and outcome-generating. This reframes where investment, talent, and architectural attention should actually go: not on chasing models, but on the systems that hold them up.
Observability Is Where AI Fails Silently and Expensively
Traditional observability was built to optimize a single metric — uptime — and watch application or infrastructure health. AI breaks that model. Dashboards can stay green while the system produces wrong outputs, hallucinates, or generates results that look confident but aren’t reliable. Of the seven skills required to operationalize AI, observability is the most underestimated because its failures are quiet, not loud. In regulated environments like banking, this gap is paired with resiliency: AI systems must clearly answer when to escalate, when to say no, and when to fall back, with circuit breakers preventing cascading failures. Observability and resiliency function as the eyes and hands of any production AI system.
Weak Data Becomes More Dangerous in the Hands of AI
Bad data used to cause a bad report or slow down an analyst. Inside an AI system, that same weak data shapes a risk decision, a fraud alert, or an automated customer response, and AI makes it sound convincing. AI and data quality are inseparable, which is why data must be treated as a production-grade product, not a one-time import. That means clear ownership, definitions, quality standards, lineage, freshness, versioning, and access controls. If the definition of income, exposure, or customer segment is ambiguous, the AI inherits that ambiguity and presents it with confidence. Data governance isn’t paperwork but the infrastructure for trust.
Episode Highlights
The Real Question Isn’t Whether AI Works
When Praveen took over a team stuck running pilots for over a year, the breakthrough came from reframing the entire conversation. The problem was never whether AI could deliver — it was whether the data was fresh and reliable, and whether an operating model existed to put models into production. He cites that only 11% of organizations are actually reaping the benefits of moving from pilot to production, which underscores how much ground remains.
“It’s not about asking if the AI is useful, but ask the question about, can we run AI in a most secure and a reliable fashion so that we can create an outcome?”
A Formula That Reorders AI Priorities
In one of the sharpest moments of the conversation, Praveen offers a blunt allocation that flips how most leaders think about AI investment. Models still matter – reasoning, coding, and general-purpose models aren’t interchangeable — but the differentiation has shifted to the engineering around them. The numbers make the case plainly and give senior leaders a clear lens for evaluating where their teams are actually spending time.
“Only 10% of the time it matters to kind of have the right model. 20% of the time, it’s having the right infrastructure. 70% of them is basically building that engineering layer around it.”
Green Dashboards, Failing AI
Of the seven skills required to move AI from pilot to production, observability is the one Praveen flags as most underestimated. Traditional observability was built around uptime and infrastructure health, which means a system can look perfectly healthy while quietly producing wrong, unreliable, or hallucinated output. The key implication here is that the failure mode for AI isn’t loud, it’s silent and expensive.
“The dashboards might look green, but the AI is failing because it might produce wrong results. It might hallucinate.”
The Chef Analogy for Data Quality
When asked about hard lessons in data governance, Praveen offers an analogy that lands instantly with executives wrestling with data debt. Where bad data once meant a bad report, inside an AI system that same weak data shapes risk decisions, fraud alerts, and automated customer responses — with confidence. Data governance, in his framing, isn’t paperwork. It’s the infrastructure for trust.
“If the model is the chef and data is the ingredient, a great chef cannot cook with the spoiled ingredients.”
Slowing Down Is the Risky Bet
The riskiest decision Praveen has made in transformations isn’t moving fast but deliberately slowing down to build the foundation when everyone else wants speed. Pausing to put a platform, governance, observability, and an operating model in place can look like delay and invite hard questions about why more use cases aren’t shipping. In a previous role, that discipline let his team deploy more than 70 models at scale with over 99% adoption.
“The win is not about rushing the idea into the production. The win is about proving the value in the focused area so that you can learn from it, build a repeatable pattern, and then scale from there on.”
