Artificial intelligence is entering a new phase of scrutiny inside organizations. Where early adoption was often driven by curiosity and the desire to experiment, leaders are increasingly expected to justify AI initiatives in terms of measurable business value. This shift raises questions that go beyond simple budget approval. What actually constitutes value in an AI project, cost savings, revenue growth, improved customer experience, or something harder to quantify? How can organizations reliably measure whether an initiative is delivering on that value once it moves past the pilot stage? And as financial accountability grows, how much space remains for the kind of experimentation that new technologies still require, without which innovation risks stalling altogether? A related question follows closely behind, which AI capabilities should organizations build and control themselves, and which are better sourced from outside? These questions frame a broader conversation among industry leaders, one that begins with the shift away from exploration and toward AI as a rigorously ROI-driven discipline.
This was discussed at an AIM Global Council roundtable. The leaders contributing to the roundtable discussion were Kulbhooshan Patil, Head of Data Science and Analytics at TATA AIG General Insurance Company Limited; Krupa Shah, Head of Analytics at Tata Play Ltd; Dr. Saurabh Pramanick, PhD, Data Governance Officer at Bank Muscat; and Rushabh Shah, Managing Partner at STIR Advisors. The discussion was moderated by Varun Saxena, Director - AI & Analytics at ANAROCK.
Moving from Experimentation to Enterprise Accountability
Across the discussion, leaders described a marked departure from earlier approaches to artificial intelligence, when projects could be pursued largely out of curiosity about what the technology might accomplish. That era, participants agreed, has given way to a much stricter expectation: AI initiatives must now demonstrate concrete, measurable returns before they receive sustained investment. As Varun Saxena, the moderator, summarized the change, "broadly at least our thought process as compared to three, four years ago has shifted to a more ROI-driven thought process." Organizations are no longer willing to fund AI efforts on the promise of future potential alone; instead, a clear connection to business value has become the baseline requirement for any project to move forward.
This shift has produced a more structured approach to evaluation. Rather than developing a solution and searching for its application afterward, the discussion pointed to a sequence that begins with identifying an actual business need, followed by aligning relevant stakeholders around shared objectives, testing whether the initiative demonstrates financial value, and then analyzing the outcomes of that testing before scaling further. This framework distinguishes exploratory efforts, which remain inherently uncertain, from more mature use cases that are ready for broader implementation.
Yet the conversation also surfaced a genuine tension within this new discipline. While traditional, established AI approaches tend to be easier to measure and justify quickly, many Generative AI initiatives remain experimental by nature and resist fast, easily quantifiable evaluation. As Rushabh Shah put it, a promising use case "may show a good and a stronger ROI, but if it is taking some 12 to 15 months, that's going to dilute the entire purpose." Leaders acknowledged that an overly rigid focus on ROI risks squeezing out the very experimentation that emerging AI capabilities still require to mature. This tension raises a deeper question underlying the entire discussion: what actually counts as value in the first place, and how should organizations define it before they can even begin to measure it.
Defining What Value Actually Means
If the first shift was about demanding ROI, the discussion made clear that a harder problem lies underneath it: determining what genuine value actually looks like once an AI initiative is in place. A recurring point of agreement was that saving time does not, by itself, create economic value. Efficiency only translates into real financial benefit when it results in a tangible change to how capacity is used, whether that capacity is reduced, redeployed toward other productive work, or otherwise absorbed into the organization. Without that connection, hours saved or percentage efficiency gains remain abstract figures rather than demonstrated returns, and leaders were consistent in rejecting these as sufficient evidence of impact on their own.
This difficulty becomes more pronounced with softer outcomes, such as improvements to customer experience, which are widely valued but resist straightforward translation into financial terms. As Krupa Shah emphasized, "clarity of KPIs, which metric are you really touching with this particular experiment, that's super, super critical." The discussion also drew a firm distinction between real, realized savings and "notional" savings, theoretical amounts that assume a cost would have been incurred had an action not been taken, with the latter viewed as misleading in leadership reporting. To keep evaluation meaningful, leaders pointed to the value of narrowing board-level reporting to a small number of KPIs that connect directly to business and financial outcomes, rather than overwhelming decision-makers with dozens of indicators. Risk-adjusted ROI emerged as another necessary lens, ensuring that measured value also accounts for whether an AI implementation is increasing or reducing organizational risk, a point framed directly by Saurabh Pramanick's central question: "are you trying to reduce your risk, increase your risk, and have you done your risk evaluation?"
Having established how difficult it is to define and measure value, the conversation turned naturally to a related strategic question: which AI capabilities organizations should build for themselves, and which are better acquired from outside.
Deciding What to Build and What to Buy
Alongside the challenge of measuring value, the discussion turned to a related strategic question: which AI capabilities should an organization develop and control itself, and which are better obtained from outside. Rather than framing this as a binary build-versus-buy decision, leaders converged on a more layered, hybrid approach. Capabilities considered core or strategically sensitive, the kind that protect an organization's distinct advantage, were seen as worth building and retaining in-house, while everything else was better left to those who specialize in it. As Kulbhooshann Patil put it, "core capabilities, build in-house. If that's not our core capability, let an expert do it." Non-core or commoditized functions, by contrast, were viewed as better handled through external vendors, purchased solutions, or the built-in AI modules already embedded within existing platforms.
This hybrid thinking also extended to how different types of AI work together within a single workflow, with traditional AI often used to provide direction or classification and Generative AI applied to execution tasks that follow from it. Governance was raised as a necessary companion to this approach, ensuring that as AI initiatives are developed or deployed whether built internally or sourced externally, they remain overseen and coordinated rather than fragmented across the organization.
A clear tension ran through this part of the discussion: internally built solutions carry the risk of becoming obsolete as external tools evolve and mature more quickly, raising the question of whether continued investment in an in-house build still makes sense once a capable outside alternative exists. This makes the build-versus-buy decision an ongoing evaluation rather than a one-time choice, and one more input into the broader effort discussed throughout the conversation to ensure that AI investment translates into value that is real, measurable, and sustained over time.
Conclusion
Taken together, these strands of discussion point to a common underlying discipline now shaping how organizations approach artificial intelligence. The move away from experimentation for its own sake, the difficulty of separating genuine financial value from notional or unrealized savings, and the ongoing calibration between building capabilities internally and sourcing them externally are not separate concerns but interconnected parts of the same evaluation. Each requires clearly defined objectives, outcomes that can be measured against those objectives, and a willingness to weigh risk alongside potential return before scaling further. What emerges is less a fixed formula than a continuous discipline of questioning, one that leaders must revisit as technologies and use cases evolve. Ultimately, the conversation suggests that meaningful AI value is not assumed; it is something organizations must deliberately define, test, and prove for themselves.