AI’s real return is trust

This is no longer a technology question. It is one of strategy, capital, and governance. AI needs a simple test of whether the value it claims is real. Applied mechanically to AI, they mislead. First, the benefits are scattered. A manufacturer predicting equipment failure gets fewer breakdowns; the bigger prize is better asset utilisation, longer equipment life, and safer plants. Much of the value shows up as better judgement and resilience.

Second, value depends on people and process. A sophisticated tool in a badly designed process delivers little; a humbler tool, fed good data and embraced by its users, can deliver far more. The old metrics are not wrong. They are incomplete. Putting a price on proof We propose a complementary discipline: TAROI, or trust-adjusted return on intelligence. By trust, we mean two things a board can inspect: Evidence that value has been delivered, and demonstrated control over the cost and risk of delivering it.

TAROI = (Value released × evidence coverage) ÷ (compute + governance + remediation) The denominator is the total cost of intelligence, on discounted cash flows over a stated horizon. Value released is the benefit claimed against a defined baseline. Evidence coverage is the share of that value backed by credible proof, scored on a ladder the board sets in advance: An assertion, a modelled estimate, a result measured, a saving realised in the accounts, or a saving independently verified.

Compute covers building and running the intelligence; governance covers training, security, testing, and oversight; remediation is the probability-weighted cost of fixing errors. Take an AI-led procurement programme claiming Rs 100 crore of annual savings. Only Rs 60 crore is visible in the accounts; the rest is a modelled price difference. Evidence coverage is 0.6. At a full cost of Rs 30 crore, TAROI is 2, not the headline 3.3 (figures illustrative).

That gap is the case for the framework. TAROI sits alongside net present value and internal rate of return, not in place of them. Safety, privacy, and human dignity are gates cleared before scoring, not terms in the formula. It is a concept still to be tested. It brings three questions together: Is the value real? Is the evidence credible? Have we counted the cost of trust? Four questions every board should ask What problem are we solving, and how will we measure it?

Start with a business problem, a baseline, and a reason why AI beats a simpler alternative. A benefit is not real until a budget or headcount plan has moved, and the sponsor should not certify it—finance validates, internal audit reviews. What could go wrong, and who answers for it? Safeguards must match the stakes: A drafting assistant needs fewer than a system authorising payments. Consequential uses need a named executive owner, human oversight, an audit trail, and a tested way to switch the system off.

Accountability cannot be outsourced to an algorithm or a vendor. Boards should assign AI oversight to a named committee. Are we building capability or buying a tool? Success rests on data quality, redesigned processes, and skilled people, not installed software. What have we learnt, and what will we stop? Report failures alongside successes. Funding pilot after pilot without deciding to scale, redesign, or stop is poor capital allocation.

Credibility and trust An AI strategy is credible only if it starts at the top. Leaders cannot expect employees to embrace what they do not use themselves. The CEO and executive committee should use AI in their own forecasting and risk spotting. The CFO’s role is to set common definitions of value, stop double counting, and separate theoretical productivity from realised savings. Risks avoided should never be booked as savings.

Start with a handful of high-value applications, each with a pre-agreed decision to scale, modify, or stop. In AI, that is hard to sustain. Employees who distrust a system will work around it, customers will drift away, regulators will tighten controls, and boards that cannot tell results from claims will allocate capital badly. Trust is not blind faith. It is earned through reliable performance, honesty about limitations, and the ability to correct mistakes.

That makes AI literacy a leadership capability: Reskill people, and let them question AI outputs without being labelled resistant to change. AI is changing the economics of knowledge, the speed of analysis, and the scope of automation. Professionals who refuse to learn to work with intelligent systems will not necessarily be replaced by machines. They will be outpaced by colleagues who use them well.

The answer is neither uncritical adoption nor excessive caution. It is purposeful experimentation, measurable value, responsible governance, and continuous learning. The race will not be decided by how much is spent, how many pilots are launched, or how sophisticated the models are. It will be decided by whether intelligence produces better outcomes for business and society, and whether the trust needed to sustain them is deserved.

The question for every board is no longer “What can AI do for us?” It is “What value can we responsibly unlock, how will we prove it, and what will make people trust the result?” Sundar Pichai said at Davos that trust “is something easy to lose, if you are not good stewards of it”. For leaders in the intelligence age, that is not just an ethical reminder. It is a business imperative. The writers are respectively Chairman, Shailesh Haribhakti & Associates, and Non-Executive Chairman, Continuum Green Energy and Protean eGov Technologies; and Group CFO, Vedanta. Disclaimer: The views expressed are the author's own and do not reflect the official policy or position of Financial Express.