A medical research assistant was about to be switched off. It had been built on a vector database, it answered questions fluently, and it was wrong more than fifteen percent of the time. In a clinical context that is not a tuning problem. The client was ready to shelve the initiative.
Aakash Angadi's team rebuilt it instead, putting graph-based reasoning alongside the retrieval layer and adding a simulation and feedback loop to validate what came out. Hallucination came under control and access to research insights roughly doubled in speed. Angadi reaches for that project whenever an executive tells him governance will slow delivery down. Responsible design, in his read, is why the thing shipped at all, which makes it a growth lever rather than a brake.
Angadi leads enterprise AI transformation and AI product strategy for insurance at LTIMindtree, one of its largest verticals. He headed the AI center of excellence at Coforge before that, as associate vice president, and carried profit-and-loss responsibility at Concentrix and at Wipro, where he owned a consumer business unit turning over around a billion dollars. By his account the work has produced more than 40 million dollars in AI revenue across those roles, and efficiency gains above 30 percent for clients running operations of up to 100 million dollars.
He organizes all of it around three tests he calls RRR. Is the approach right-fit for the business problem rather than the most fashionable option available. Is it responsible enough to survive a regulator and a reputation. And is its return measured on real outcomes instead of a pilot demonstration.
"Most programs get the first R right and stall on the other two," he says.
Where they stall is specific, and he is blunt about it. Agentic AI does not fail in regulated industries because the models are weak. It fails because "nobody has answered what an agent may decide alone, how it is tracked and how a regulator is answered." Those are governance questions. The costume is technical.
His answer is a governance standard on four control planes: identity, data boundaries, safety and observability. Agents get evaluated against outcomes rather than against a showcase. He has written on the engineering side of these problems too, including neuro-symbolic approaches to insurance submission triage.
Which leads to the advice most likely to be ignored. Most enterprises open with their most visible customer-facing use case, because that is the one that makes an announcement. Angadi starts inside the building, with use cases built around employee roles, where a mistake is cheap and the learning arrives quickly.
The responsible layer goes in before the use cases rather than getting bolted on afterwards, so everything built later inherits it. "Slower to announce, much faster to scale," he says.
The payoff tends to show up as revenue, not as a line in a cost report. One manufacturer carried more than 57,000 products with little to separate them on specification, held in formats that did not agree with each other. Sales staff could not reliably match a customer to the right item.
An agentic retrieval system let them ask in ordinary language and get a validated answer with its sources attached. He puts the fall in search effort at 90 percent and the lift in conversion at 300 percent, and the pilot became a larger engagement.
What stops that from being a single win is where the money goes next. Budget spent on reusable agents, playbooks and architecture blueprints compounds, he argues, while budget spent re-solving the same use case does not. On a recent insurance program he replaced four bespoke agent builds with three configurable patterns covering four lines of business, so each additional line costs a fraction of the first.
The organizational version of that argument is the one he has had to make internally. A center of excellence has to run as a profit center rather than a cost center, which in practice means refusing to become an innovation lab that produces demonstrations nobody buys. He treats every engagement as a delivery and an asset-building exercise at once. That is how the center he built at Coforge reached more than 30 million dollars in AI revenue, and put more than ten enterprise agentic systems into production.
He extends the same logic to the invoice. Tokens are capital rather than consumption, in his framing, so the discipline is to route, cache and reuse until the next use case costs less than the last one did and spending stops rising in step with usage. Most organizations fund strategy and delivery generously and leave nothing for the layer that sustains either, which is where governance, agent evaluation and the economics of inference live.
Angadi expects the shape of the function to change as well. The central, one-size-fits-all AI center is giving way to teams sitting close to individual businesses and their domains, with a lean central layer holding governance and shared platforms. Sovereign AI requirements are accelerating that split, as clients work out where their data and their models are permitted to live.
"The industry is moving from can we build it, to can we adopt it, scale it and show ROI," he says. The research assistant that nearly got switched off is the small version of that story. Nothing about it was rescued by a better model.