By Nithin Alkanand, Partner, Sapphire Human Capital
Capital has stopped being the binding constraint on India’s artificial intelligence ambition. Capability has taken its place. Speaking at the India AI Impact Summit in February, the Union Minister for Electronics and IT indicated that more than $200 billion of AI investment is likely over the next two years, layered onto a national compute base of 38,000 GPUs, with a further 20,000 being added. Money at that scale rarely fails loudly. It fails in the operating layer, in the unglamorous discipline of keeping live models honest, and that failure arrives as a leadership problem long before anyone calls it a technical one.
The vocabulary in the boardroom has shifted accordingly. Directors who spent two years arguing build versus buy are now being asked who owns a model once it goes live, who watches it degrade, and who answers for it when a credit decision or a customer conversation goes wrong. LLMOps, the discipline of running large language models in production rather than demonstrating them in a pilot, has become the fault line separating enterprises that report AI progress from enterprises that book AI value.
Adoption is running well ahead of expertise
The asymmetry is now measurable. Deloitte’s State of AI in the enterprise report for 2026 found 40 per cent of Indian respondents reporting significant or full AI usage against a global average near 28 per cent, with at-scale deployment strongest in product development and operations. Depth tells the opposite story. Indian organisations reported a high level of AI expertise at only 0 to 4 per cent, compared to a global range of 2 to 8 per cent, even though 94 per cent of them expect AI spending to increase over the coming year.
Spend curves are steepening faster than competence curves. That same study put the leading integration challenges at regulatory and compliance requirements, cited by 39 per cent, and resistance to change, cited by 34 per cent, with cost at 12 per cent and infrastructure at a negligible 5 per cent. Rendered into commercial language, the bottleneck sits in governance readiness and operating model change. In other words, the bottleneck lies with people, not platforms. India’s banking regulator arrived at much the same verdict, though it got there by lifting the bonnet.
The regulator saw the operating gap before the market did
Ahead of publishing its framework in August 2025, the Reserve Bank of India surveyed banks, NBFCs and fintechs. Its FREE-AI committee report records that among the 127 entities already using AI, only 21 per cent monitored for data or model drift, and a mere 14 per cent conducted real-time performance monitoring. Roughly one third had board-level oversight of AI in any form.
Those are not compliance statistics. They are operating statistics. Drift monitoring is what separates a fraud engine that works from one that stopped working three quarters ago and never mentioned it. The committee’s remedy was structural rather than technical, calling for a board-approved AI policy, model governance across the full lifecycle, and capacity building at the board, management, and workforce levels. Capacity at board level is a hiring instruction masquerading as a recommendation.
The mandate has moved out of the lab and into the P&L
Sourcing that capability is harder than sourcing a title. LLMOps readiness in a CXO is a composite competence rather than a credential, and it surfaces in specific questions. Has the candidate sustained a model in production for eighteen months, through retraining cycles, vendor changes and a challenging quarter, or only launched one? Can they read a model risk register the way a CFO reads a cash flow statement? Do they grasp the unit economics of inference well enough to kill a use case that will never earn out? Most tellingly, can they say no to a board that wants an announcement?
For the capital markets, this has already become a diligence item. Investors and lenders now probe AI governance much as they once probed cyber posture, and disclosure expectations in regulated sectors are moving the same way. Scarcity is doing what scarcity does. Compensation for proven operators has decoupled from conventional benchmarks; lift-outs of intact engineering and risk teams are running ahead of individual hires because capability of this kind lives in a group rather than a person, and retention has become the quiet risk nobody underwrites. Appointing a leader the organisation is not yet equipped to evaluate is a governance failure with a long lag and an expensive tail.
Adding a title to the organisation’s chart does not solve any of these issues. It is solved upstream in how boards define the role before the search begins, how nominations committees rebalance composition so that technical fluency sits alongside financial and regulatory judgement, and how ruthlessly the market is mapped rather than merely canvassed. The organisations that emerge from this cycle in front will not be those that have bought the most compute. They will be those whose boards treated leadership composition as infrastructure and moved with the agility to reset it when the technology reset the requirements.
A leadership appointment made well does considerably more than fill a seat. It sets the tempo at which an institution learns, and in a production AI environment, the rate of learning determines the competitive position. The benchmark now extends some distance beyond the hire itself.