Almost everyone in the AI conversation builds it, studies it, or sells it. On Unscripted, I wanted the view from the person they are all selling to: the enterprise buyer inside one of the largest companies in the world.
The short answer from that conversation: a Fortune 10 company does not chase the flashiest AI. It runs a disciplined process that starts with the biggest business opportunities, treats AI as one tool among several, and de-risks bold ideas with prototypes before committing real money. The wins that matter are often unglamorous, and the vendors that get in the door are the ones who can prove impact, not the ones offering a free trial.
Zachary Elowitz runs the enterprise AI lab at McKesson, a Fortune 10 healthcare and pharmaceutical distribution company. He holds a PhD in math and an MBA, and he helped write a federal risk framework that is widely cited. He has now done this kind of work across healthcare, manufacturing, banking, and consulting, which gives him a rare cross-industry read on what actually works.
Zach's team exists to de-risk the boldest AI ideas at McKesson. Most technology leaders in the company are embedded in specific business areas and do not have the bandwidth to chase higher-risk, higher-uncertainty ideas that might not become production capabilities. His team takes those on. They build prototypes that prove feasibility and potential before the business commits resources.
The framing matters. The lab does not start with "we want AI." It starts with the biggest opportunities in the business, then asks where AI, data science, or optimization is the right tool. As Zach put it, if you cannot clearly articulate what success means, whether that is time saved, money saved, or risk reduced, you are wasting your time and will end up on a side quest.
Zach's recurring theme is that people should not focus on flash, they should focus on impact. The example that proved it had nothing to do with generative AI.
His team examined how McKesson sources inventory and asked whether they could more efficiently borrow stock from other distribution centers rather than buying net-new from manufacturers. They discovered that decisions made years earlier rested on smart people's intuition, choices that either were never mathematically optimal or no longer held. Through optimization and simulation across thousands of scenarios, they found a way to save eight figures in working capital per year, with no disruption to the business and inside the supply chain organization's risk tolerances. Pure optimization, no flash, enormous impact.
He saw the same pattern in manufacturing. During the volatile tariff period of early 2025, having AI and clean data laid out clearly let his prior organization act on nine-figure decisions quickly, because they already understood the cost implications and likely customer responses.
Beyond the keynote version, Zach's read is that to most CEOs, AI is still a buzzword, because they are not close to it. They see peers claiming transformations and they want to understand how to get the same benefits, save money, and grow revenue.
He was also direct about layoffs: leaders are not trying to get rid of people. They are serving the business, and sometimes individual roles and the needs of the business do not line up, which is where much of the tension in the headlines comes from. The more novel thing CEOs want, he said, is to fundamentally change how they serve customers. He pointed to healthcare moving from in-person visits toward telehealth, and potentially toward low-risk prescriptions handled through an app in real time, as an example of AI reshaping the customer relationship rather than just making an existing process incrementally better.
Admissible risk is the level of failure you are willing to tolerate for a given AI system. Zach's point is that pretending the acceptable level is zero is not how the world works. Every medication carries a list of risks that are already occurring. The honest move is to name and quantify the risk you will accept.
He put it starkly: if a capability could save many lives, are you willing to accept the possibility of harm in a small number of cases? He tied it to self-driving cars as well. If the technology produces meaningfully fewer deaths and injuries than human drivers, the harder question becomes whether you can justify the additional risk of keeping a human as the primary driver. These questions sound cold said out loud, but refusing to ask them does not make the risk disappear.
This section is worth the episode on its own for anyone selling AI.
First, build versus buy. Zach builds in only two situations: when an outside vendor would be exorbitantly expensive relative to what he can do internally, usually because the need is very niche, or when the capability would meaningfully distinguish the business from competitors. For anything that outside teams will keep improving forever and that will only get cheaper, he would rather buy and spend his people on what is specific to the business.
Second, how vendors earn his attention. He gets more pitches than he can answer and ignores most. Three things break through: a peer he trusts who has used the product and can describe the problem it solved, a reference customer with no incentive who will speak candidly, or a salesperson who can describe the fundamental problem so precisely that it is clear they have lived it. What does not work is "try it for free." As he put it, there is no such thing as a free trial, because a trial means pulling his people off their work for weeks and possibly exposing enterprise data. The only real exception is desperation: if you need something urgently, you will give it a shot.
While everyone stares at AI, Zach is watching quantum computing and the race to post-quantum cryptography. Timelines to be ready have moved earlier, toward the end of this decade. His concern is blunt: if cryptographic protections are not ready when the breakthroughs arrive, the first successful break may not be announced at all, and the questions get serious fast, are the banks safe, is the data safe, does HIPAA still mean anything. It is getting attention in small circles, and he thinks the broader world, distracted by AI hype, is not watching closely enough.
How do large enterprises decide where to use AI? They start with the biggest business opportunities, not with AI itself, then ask where AI or related methods are the right tool. Clear success criteria, time saved, money saved, or risk reduced, come before any technology choice.
What is admissible risk in AI? It is the level of failure an organization is willing to tolerate for a given system. Effective teams name and quantify that tolerance explicitly rather than pretending acceptable risk is zero, which is not realistic in fields like healthcare or transportation.
When should a company build AI in-house versus buy it? Build when an external option would be far more expensive for a niche need, or when the capability meaningfully differentiates the business. Buy when outside teams will keep improving the capability and costs will keep falling.
How do enterprise buyers evaluate AI vendors? Trusted peer references, candid reference customers with no incentive, and salespeople who clearly understand the underlying problem. Generic "free trial" offers rarely work, because trials consume internal time and can expose sensitive data.
Why does an AI leader care about quantum computing? Because quantum advances threaten current cryptography. If post-quantum protections are not in place in time, financial systems, personal data, and regulated health information could be exposed, possibly without immediate detection.
Listen to the full conversation with Zachary Elowitz, and catch every episode, on the Unscripted with Jeff Pedowitz page.