Most enterprise AI investment is not failing because the technology is immature. It is failing because the investment was never connected to a business outcome in the first place. No measurable key performance indicators (KPIs), defined return on investment (ROI) target, or real problem being solved. Just enough activity to satisfy a board that asked about the AI strategy.
David Ellison, who leads Lenovo’s AI Center of Excellence, has run more than 277 AI project across industries as different as sports bodies (NASCAR, FIFA), health care organizations, island conservation, agriculture, and cell biology. He has seen both versions up close, and his read on what separates the organizations that generate real returns from those that generate impressive presentations is unambiguous. “Treat AI as a revenue engine, not a science experiment,” Ellison says. “You should be a problem in search of a solution, not a tool in search of a problem.”
Start With the Business Problem. The Technology Comes Second
The discipline Ellison applies is to identify the business problem first, anchor it to a measurable KPI, and work backward from the desired outcome to the simplest technology capable of producing it. Not the most sophisticated technology. The simplest one that solves the problem well.
Complexity is where AI projects fail. Organizations that over-engineer solutions before validating value create systems that are expensive to maintain, difficult to adapt, and impossible to scale when the time comes. Build validation into the process from the start, including benchmarking, proof of concept, and real-world testing, rather than measuring only at the end when the cost of correction is highest. The goal is incremental delivery of value throughout, not a reveal at completion that nobody can act on.
Poor data readiness sits underneath most of these failures as the foundational constraint. Organizations over-focus on models and technology when the data powering those models is siloed, structurally inconsistent, or inaccessible. “The data is what you should be focusing on, not the models,” Ellison says. Getting the data foundation right is not a prerequisite that can be deferred until the AI strategy is ready to scale. It is the AI strategy.
Solve One Problem Extremely Well Before Building a Platform
Across every industry Ellison has worked in, from route optimization at NASCAR to production planning in agriculture to elder care, the universal formula holds. Target high-friction, high-cost problems at the center of the business, not at the edges. Establish a clear baseline so the value AI adds can be quantified against what the existing system costs and delivers. Solve one problem with precision and depth rather than building a broad platform that attempts to address everything at once.
“Solve one problem really well,” Ellison says, “and then build on that and solve another problem. You can build a platform by solving one problem at a time in a really good fashion.” The organizations that try to build the platform first and work backward to individual use cases consistently overscope, under-deliver, and lose the internal credibility required to scale. The ones that solve one problem so well that the result is undeniable build organizational trust, making every subsequent initiative easier to fund and faster to deploy.
Agentic AI follows the same logic. The genuine value of autonomous, multi-step workflows is evident in complex route optimization and in intricate business processes that span multiple systems and decision points. But the organizations deploying thousands of uncontrolled agents without governance are manufacturing failure at scale. If each agent in a sequential workflow has a 5% failure rate, the cumulative probability of failure grows multiplicatively along the chain. “Prioritize reliability, observability, and control,” Ellison says. Add agents incrementally, keep humans in the loop during implementation, and never pursue full autonomy without a clear ROI justification.
The Productivity Paradox and Why Incremental AI Adoption Has a Ceiling
When personal computers arrived in the 1980s, productivity gains were invisible in macroeconomic data until the 1990s, not because the technology was weak, but because organizations integrated computers into existing workflows rather than redesigning them from the ground up. The same pattern repeated when electricity transformed manufacturing: adding electricity to an existing plant layout yielded minimal improvement, whereas redesigning the entire system around the new capability delivered transformational gains.
AI is following the identical pattern. Scattering proof of concepts across disconnected workflows, inserting AI into isolated processes, running pilots that never reach production, none of it delivers the value the technology is capable of. “It is when you rethink and plan for an AI-native application,” Ellison says, “that you produce systems that really deliver value.” The organizations that will define competitive advantage in the next decade are the ones redesigning from the ground up, not the ones adding AI incrementally to processes built for a different era. “If you think investing in AI is expensive,” Ellison says, “wait until you see the cost of being left behind.”
Follow David Ellison on LinkedIn or visit Lenovo for more insights on AI ROI, enterprise AI strategy, and building AI programs that deliver measurable business value.
