- AI value is strongest when tied to time savings, decision quality, revenue impact, cost avoidance, or risk reduction.
- Generative AI pilots are often easy; production requires governance, engineering discipline, observability, and adoption planning.
- Cost control depends on model routing, caching, token monitoring, and knowing when not to use AI.
Those themes shaped Mistral AI’s panel Scaling AI That Works: Proving Value, Managing Cost, and Driving Impact, at the ALIGN AI Executive Summit, Atlanta. The session was moderated by Frances Boykin, Director Advanced Analytics at AT&T, and featured Shashank Kadetotad, Global Sr. Director of Enterprise Data Science and AI at Mars; Ian Matthew, Solutions Architect at Mistral AI; Vida Williams, Chief Data Officer at Stride; and Nisha Jaisinghani, Director IT (Data & Analytics – Corporate Capabilities) at Georgia-Pacific.
The panelists came from telecommunications, education, consumer goods, manufacturing, and AI infrastructure. They converged on one point: AI at scale is not a side project. It is an enterprise operating discipline.
What does AI value mean once pilots are over?
For Nisha, the most measurable value today is time savings. AI-enabled development environments have accelerated work that might previously have taken months. She also pointed to ServiceNow-style resolution agents that search historical tickets and recommend fixes. In those cases, Georgia-Pacific has seen savings claims around 40 percent.
Ian pushed the value conversation beyond basic productivity math. Time savings matter, but the strategic question is what an organization does with the reclaimed time. The value may be shorter delivery cycles, more ambitious product roadmaps, improved customer experience, or work that previously sat below the priority line.
Vida offered a different lens. At Stride, the strongest near-term return has come from using AI to advance data governance, data maturity, and self-service. In education, she emphasized, teams need to “tame a probabilistic technology” into systems reliable enough for high-stakes environments. Shashank added decision complexity: at Mars, AI can reduce the time and friction involved in decisions such as pricing, but the cost can exceed the savings unless the use case is carefully designed.
Why AI pilots are easy and production AI is hard
Several panelists made the same point: generative AI has inverted the traditional AI delivery curve. With LLMs and agents, Shashank said, pilots can be easy. Production is where teams discover reliability issues, ambiguous problem statements, infrastructure gaps, and the operational risks of probabilistic systems.
That is why Mars has doubled down on crisp problem statements and business partnership. Before scaling a use case, the team defines the problem clearly and connects it to shared infrastructure that can support more than one project. Vida described a similar need for modularity: teams should be able to observe, replace, decommission, or recommission a model without disrupting the broader platform.
Ian put it bluntly: “It’s not magic. It’s a piece of software.” Enterprise AI still depends on software engineering fundamentals: data governance, security, deployment practices, A/B testing, monitoring, and deep involvement from engineers and cloud specialists.
How should teams prioritize AI use cases?
With limited resources, the panelists argued for prioritization frameworks that go beyond executive excitement or demo appeal. At Mars, Shashank’s team looks at data availability, cleanliness, development time, and the ability to create a viable MVP. Nisha echoed the importance of quick wins, especially for teams still building maturity. She described a Georgia-Pacific use case where an agent analyzes engineering drawings, compares them with master data, and links maintenance parts in minutes, with a human in the loop for validation.
Ian added that quick wins should also have a path to scale. A narrow chatbot over parts data may be useful, but the stronger investment is a reusable capability that can later support reconciliation, process automation, or more complex workflows.
Vida introduced one of the session’s most important leadership cautions: teams should consider the long-term workforce impact of cognitive offloading. Automating junior-level work can remove drudgery, but it can also weaken apprenticeship paths. Ian summarized the distinction well: “You can offload your thinking, but you can’t offload your understanding.”
What cost discipline looks like in enterprise AI
The cost conversation moved beyond budget anxiety into design practice. Nisha separated development cost from production cost and encouraged teams to be proactive. Unlimited token access may help early adoption, but usage growth requires limits, contract discipline, and careful decisions about which workloads justify expensive model calls.
Ian described a shift already underway: sending every task to the largest available model is powerful but unsustainable. Many tasks can be routed to smaller, cheaper models optimized for specific jobs. Hosting choices, VPC or on-prem requirements, and task complexity all affect the economics.
Shashank described composable architecture as a cost-control mechanism. Mars focuses on routing the right prompts to the right model, caching where appropriate, and monitoring agent behavior. As employees gain access to tools like Copilot, centralized registries and observability become essential.
Vida widened the definition of cost to include environmental impact, asking engineers whether what they built was worth the water consumed to run it. That framing moves cost consciousness from a monthly slide into engineering culture.
Why governance should make AI easier, not just slower
The panel closed on governance, but not as a compliance afterthought. Shashank described Mars’ enterprise data science and AI team as a delivery team that has become deeply focused on governance. The aim is to reduce shadow IT by making the governed path more useful than the ungoverned one.
Vida described governance as a cross-functional operating model that includes architects, engineers, analysts, information security, and legal. Nisha emphasized the people side: a central AI team cannot solve every AI problem. Smaller pods with AI specialists and business experts are often more effective, especially when they track real adoption.
Shashank made adoption part of the delivery model from the start. His team talks directly with the people who will use the product, not only the executive sponsor, and allocates funding to training. They also define abandonment criteria: if a project misses a defined milestone by a defined time, the team stops, documents the learning, and moves on.
Questions answered in this session
- How should enterprises define measurable AI value beyond a successful demo?
- Why do generative AI pilots often fail to translate into production impact?
- What role do data governance and data maturity play in scaling AI?
- How should organizations prioritize AI use cases across business functions?
- When does the cost of using AI outweigh the value of automation?
- How can model routing, caching, smaller models, and observability reduce AI operating costs?
- When should an enterprise abandon an AI project and capture the learning?
The main takeaway: scale depends on discipline
The panel’s clearest message was that enterprise AI does not scale because a model is impressive. It scales when the organization can connect the work to a real problem, validate outcomes, manage cost, observe behavior, reuse infrastructure, and support the people expected to adopt it. That is less flashy than a demo, but far more durable: AI that is not just functional, but economically and operationally viable.
Event link: ALIGN AI Executive Summit ATL
