Enterprise AI Maturity Requires Discipline As CIOs Tighten ROI Standards

Most enterprises can launch an AI pilot. What separates mature organizations from the rest is the willingness to shut one down. As generative AI moves into core operations, CIOs are finding that the challenge is no longer technical feasibility, but deciding which initiatives actually deserve to scale.

Udit Pahwa is Chief Information Officer at Blue Star Limited, a multinational company best known for its engineering, cooling, and infrastructure businesses. A Chartered Accountant by training and a CIO-100 Awardee, Pahwa has spent his career scaling enterprise systems with a clear focus on financial outcomes. His experience has shaped a view of AI that starts with profitability, not experimentation. At the center of Pahwa’s strategy is a counterintuitive belief: progress requires a tolerance for failure.

"To foster innovation, you need to have an appetite to accept failure. If you need to innovate, you have to be okay with failing, learn from it, and deliver faster," said Pahwa. He believes that too many AI pilots survive on executive enthusiasm or novelty rather than measurable business impact, consuming budget and talent without ever earning the right to scale. The result is pilot purgatory, a growing state of organizational inertia.

• Shut it down: In an era of innovation theater, abandoning technically impressive work has become a marker of real leadership. "A tolerance for failure creates space for better judgment, because the hardest decision is knowing when a project no longer serves the business," Pahwa explained. He pointed to a predictive maintenance pilot as proof. "The pilot was a technical success, and we built an algorithm that could predict equipment failure two to four weeks in advance. But when we looked at deploying it at scale, the economics did not hold, the ROI was not favorable, and the monetization was not there, so we shut it down."

• Betting on promise: The same discipline applies to newer generative development tools. Early promise alone does not justify continued investment without clear downstream value. "We experimented with a developer GPT where you feed a specification document in and it builds out the wireframes," Pahwa recalled. "While it is positioned as a feature that can cut development time, we found it is currently useful only for building wireframes. The technology still has to evolve, and we decided to put it on hold."