Liz Boschee: the lab-to-production AI gap Enterprise teams keep failing to bridge

Enterprise AI teams are usually not short on ideas. They are short on something harder: taking an AI prototype that looks great in a controlled setting and turning it into a dependable, production-scale system. That lab-to-production gap is exactly where Liz Boschee, VP, AI Foundations at Capital One, says most efforts stall. In her view, “successful AI implementation isn’t just about adopting the latest models or tools.” It requires a disciplined R&D approach that connects foundational research to real-world systems, and holds ideas accountable as they move from concept to production.

And that accountability matters because enterprise environments are complex, fragmented, and risk-minded. AI capabilities may evolve quickly, but the practical question is what actually works for a specific workflow, user, or decision, using today’s technology and constraints. The consequence for decision-makers is straightforward: if you cannot explain why an AI system works under real latency requirements, live production data complexity, and business constraints, you are not “innovating.” You are just iterating. Boschee argues the fix is to design the R&D process so teams learn what matters early, before a project calcifies into a slow, expensive commitment.

So what does that look like in practice? First, Boschee emphasizes bridging foundational and applied research, not treating them like separate worlds. When research exists in an academic vacuum, untethered from operational reality, models that perform well offline often fall short in production. Without a tight feedback loop, teams lose sight of what moves the needle for end users. At Capital One, Boschee describes AI teams designed to span the spectrum from foundational research to highly applied problem-solving, covering friction points before they stall a project. The point is not to dampen curiosity. It is to keep research tethered to use cases so teams can accelerate learning and avoid dead ends.

She highlights an example of this kind of bridging: research into combining multi-agent architectures. Boschee says the work goes beyond simple LLM reasoning, aiming to enable specialized AI agents to coordinate across distinct tasks, such as researching customer context and preparing documentation simultaneously. She also ties that research to a real product launch, explaining that this research supported the launch of Chat Concierge, a car-buying solution that “mimics human reasoning” to not just provide information, but take action on customers’ behalf based on their requests. In other words, the lab work is not just proof of capability, it is connected to a workflow where customer interactions create measurable outcomes and real operational demands.

Second, Boschee pushes for a disciplined path from concept to production, with rigorous evaluation staged as proof of concept, pilot, and production. But she warns against a common enterprise trap: treating those stages as formalities rather than decision points. A proof of concept must be functional, not theoretical. It should not be “here’s what we could do” slide deck work; it must be an actual machine doing something measurable. A negative pilot result is not a failure. If pilots always “succeed” by definition, then they are not functioning as decision points. They become slow-motion commitments to production. Instead, a pilot should expand scope and realism, generating valuable data on whether a solution helps a human do real work.

Third, she frames moving to production as a team sport, because solving the core model problem is only part of the job. Production requires cross-functional reality: software engineering, science, product and design, technical program management, operations, and other disciplines across the enterprise. This is where many AI efforts become brittle: models get attention, but integration, monitoring, reliability, and workflow fit get treated like afterthoughts. Boschee says measurement is an important input throughout the journey. Capital One’s “ultimate ROI is a happy customer,” and that shows up in key AI performance indicators like accuracy and latency, among others, to ensure the system meets the moment for customers. Her underlying rule is blunt: if you cannot tell whether you are improving, you will not be able to manage improvement.

Finally, Boschee argues that sustainable innovation depends as much on culture as technology. Research involves uncertainty and exploring the unknown, so a healthy culture recognizes reality and creates space for informed risk-taking paired with accountability. Crucially, organizations must encourage course-correction. If admitting “this isn’t working” is treated as a disaster, teams will learn to hide problems instead of solving them. But if teams are encouraged to evaluate honestly, pivot when needed, and learn from false-starts, the organization can move faster and safer at the same time. That also means treating pilots as real decision points, with the willingness to stop, reshape, or narrow efforts based on the data, rather than pushing ahead by default.

The second-order implication for boards, CIOs, and AI leaders is that “model performance” is not enough to justify production deployment. The production bar includes operational constraints, measurable workflow value, and governance that enables honest learning. Boschee’s conclusion is that building impactful AI is not about chasing every new breakthrough. It is about guiding ideas from research to reality through evaluation, collaboration, and a culture that embraces learning, so AI can deliver lasting impact at enterprise scale in the real world.