Artificial intelligence is often seen as a technology option. Which model should I adopt? Which vendor should I work with? How quickly can I get started?
Two recent case studies from London Business School suggest that these questions are the wrong starting point. Lasting advantages in AI rarely come from getting better models. It comes from designing how AI is used in real-world decision-making and under what conditions it can be trusted.
In one case, we examine UltraTech Cement, a large industrial company that is incorporating AI into plant operations across its vast manufacturing network. The other considers Crisil, a global analytics firm whose reputation depends on the reliability of its client-facing insights.
The industries could not be more different. One is to optimize the kiln and logistics network. The other supports financial decision making. Although their approaches to AI innovation are superficially different, they both converge on common insights. In other words, AI innovation is guided by a shared set of leadership design choices.
Across these organizations, three leadership design options stand out to executives and board members.
Designing for AI experiments: The relationship between cost of error and delay
Both organizations started with the same constraints of limited in-house AI capabilities, but chose very different paths.
UltraTech leaned into external partnerships and rapid experimentation. By collaborating with startups and testing solutions in real plant environments, we accelerated learning while minimizing obstacles. Internal teams focused on integration, implementation, and governance, while external partners provided specialized innovation capabilities. This approach was consistent with the organization’s operational discipline and allowed for experimentation within a culture of deliberate decision-making.
Crisil took the opposite path. Given the reputational and regulatory implications of analytical errors, we prioritized in-house development and careful iteration. The organization chose to iterate with internal teams to ensure the output met the standards required for customer-facing financial workflows.
In industrial environments, errors are often manipulable, contained, and recoverable, but the cost of delaying experimentation can be high. In financial analysis, the cost of error is measurably significant and difficult to reverse, but the slower the experiment, the much lower the risk. Therefore, UltraTech was optimized for learning speed and Crisil was optimized for learning control.
Experimentation with AI is not optional. The leader’s job is to balance the cost of making a mistake with the cost of moving too slowly.
The impact of AI beyond ROI
A second barrier to AI progress is an over-emphasis on immediate financial ROI. When every initiative has to justify itself through short-term gains, organizations delay experimentation and eliminate the very ideas that build long-term capabilities.
UltraTech addressed this issue by broadening the way impact is defined. Rather than relying solely on financial metrics, we framed our AI efforts around three outcomes: reduced costs, improved decision-making, and increased process efficiency. Use cases that demonstrated plant-wide diversity were prioritized, and early successes quickly spread and built trust in the organization.
Crisil applied the same logic in a different context. Rather than aiming for complete automation, we focused on increasing analyst productivity in core business services. Even small improvements across tasks, such as compiling data or speeding up first drafts, freed up analysts’ time to make better decisions, ultimately increasing value for our clients.
Taken together, these cases highlight the broader insight that the value of early AI is organizational before it is financial.
Initial results will show up in increased productivity, improved decision-making processes, and increased trust in AI systems. These enable learning and make future experiments more effective.
Leaders obsessed with immediate ROI risk stalling this learning cycle. Those who broaden the definition of impact create the conditions for AI innovation to emerge.
AI is a matter of trust
A common assumption in AI adoption is that better models automatically lead to better outcomes. In both cases you can see that this is incomplete.
The real challenge is not building AI systems. This allows end users to trust those systems enough to trust them in their own decisions.
This is where AI is fundamentally different from traditional digital transformation. For most systems, adoption means learning how to use the tools. In an AI system, that means deciding when to trust its judgment.
Both organizations recognized this early and involved end users from the beginning. Crisil embeds analysts directly into the development process, ensuring output aligns with real-world analytical expectations. We designed an AI solution that combines guardrails and human-involved workflows to increase trust.
UltraTech worked with factory-level champions who independently tested and validated the initial solution in real-world operating conditions. Adoption was not mandatory. Solutions are trusted only when consistent performance is demonstrated in real-world operations and trust is spread through visible results rather than training.
No adoption was imposed in either case. It was driven by trust built through experience.
Guardrails and human-involved design were seen not as constraints, but as a means to bridge the trust gap.
AI systems are adopted when users trust them in their own workflows, not when organizations deploy them at scale.
Questions Boards Should Ask Now
As investment in AI grows, surveillance must also evolve. Three questions can help anchor a productive boardroom conversation.
-
How do we foster AI experimentation? Where can we afford to move quickly and where should we prioritize control?
-
How are you defining the impact of AI in the early stages? Are you enabling learning and adoption, or are you stifling innovation with premature ROI expectations?
-
What are we doing to help users trust AI in their daily decisions? How are we reducing the cost of relying on error-prone AI output through system design, guardrails, and human oversight?
UltraTech and Crisil’s experience shows that AI success is no fluke or accident. It comes from intentional leadership and executive choices about how the organization experiments, defines impact, and builds trust.
In an era of rapidly improving models, the source of lasting advantage is not who adopts AI first. Who can design the best?.
