Whiteboarding the Future of the HC Enterprise | AI Lessons from Other Industries

Session Focus

Gandhi framed the session deliberately against the rest of the program. A conference on innovation cannot avoid multiple AI sessions, so this one was designed differently.

His observation, drawn from conversations with the panelists and across the industry, is that most other sectors are further along in using technology to transform operations, and that healthcare delivery in particular has been slower. Every panelist came from an organization with a longer history of technology-enabled business transformation, both before generative AI and since.

He was explicit that he would not try to connect the points to healthcare, trusting the audience to do that themselves, and instead wanted a genuine conversation about how technology is transforming other industries.

One New Use Case Each

Jeff Busconi described a capability recently rolled out for junior bankers, allowing them to research companies, assemble pitch book presentations and build financial models. He noted the area is known for tough hours, and expects the tool to change how people work and fundamentally change roles in the banking organization, moving junior bankers toward higher value work.

Athina Kanioura PhD described a digital twin program launched last year in partnership with NVIDIA and Siemens, incorporating operational and engineering twins across operations. In the past four months the company has moved into a next phase around physical AI, testing what a dark factory would mean in some facilities.

Margaret Pierson PhD chose examples where generative AI outperformed existing science models, which she framed as the more interesting case.

The first is catalog accuracy. Customers viewing a couch see dimensions that are sometimes entered incorrectly, often by humans making mistakes. The previous approach was an outlier prediction model that caught the most obvious errors, which were also the easiest for customers to spot themselves, and performed poorly on everything else. Generative AI can take arm height, back height, overall height and width together, reason across them and identify subtler errors far more accurately.

The second is lead scoring for sales teams. The prior system was a coarse binary classification based on on-site signals. Now a model can research the business, enrich the information, follow a chain of reasoning and circle back, handling unstructured signals in ways a traditional ML model could not.

Her most useful observation was about what happens next. Replacing one step interrupts a process that was balanced end to end around whatever detection or escalation rate the previous model produced. Repeatedly, the result is a far better model finding far more errors, with nowhere near enough throughput in the rest of the process to act on them.

She also mentioned computer vision in warehouses, including detecting damaged boxes before shipment.

Scott Sperling described an example that changed behavior as much as it applied technology. The firm reformatted roughly 25 years of investment memos for a consistency never originally planned for, then added about 20 years of offering memos from thousands of deals it looked at and did not do, in industries and sub-sectors it still focuses on.

He called the result a data pond, probably not large enough to be a data lake, made accessible to natural language queries and analytics that would previously have kept associates up all night for two weeks. It was then attached to a general LLM for outside information, with the impurity of that external data and the resulting errors still being worked through.

The value shows in comparables. The firm can now examine at much greater granularity what genuinely makes a company comparable, rather than receiving twenty companies from an investment banker where one is a hundred times larger, which he called a garbage in, garbage out situation.

An Unexpected Side Effect

Sperling identified one negative that is easy to miss. Investment committee members now interrogate the system before meetings. Questions individuals have already raised and answered privately may not get raised again during the committee meeting itself, so that information is no longer shared with the group.

The firm now forces itself to ask questions the system has already answered, to preserve the collective discussion.

Gandhi mentioned another investment firm that put an agent on top of a similar corpus and has it participate in investment committee meetings. Sperling’s response was immediate: they have refused to do that to date, allowing that as IC co-chair he may be a biased party.

He then described what deal teams have done instead, particularly younger staff. They have created personas reflecting what they interpret as the biases of individual investment committee members, and query the AI as though it were Sperling or another member, with the tilt they believe that person holds.

Governance and the Case Against Pilots

Gandhi turned to PepsiCo’s approach of deploying resources to things that can have large impact rather than running small experimental pilots.

Kanioura explained the structural advantage first: leading strategy, technology and business operations lets her connect business strategy to the execution arm, ensuring anything implemented carries an attached value realization.

Her argument against pilots is about scale. For a company with 300,000 employees operating in every market, a pilot is too small to be meaningful either in the value it produces or in the gravitas needed to convince the organization the effort and money are worthwhile. The company did run pilots early in the generative AI period, and has had traditional AI for many years.

The complexity she described is substantial: beverages and snacks, large supply chains, manufacturing, logistics, a massive front line, a highly diversified employee base and product portfolio, and significant R&D where product lifecycle management matters for products with shelf lives ranging from a day to three months. She drew the healthcare parallel explicitly around product development.

Against that complexity, any AI implementation must satisfy three constraints. It cannot compromise product quality, safety or the value chain. It cannot compromise go-to-market, meaning speed to market and customer service. And it cannot compromise what employees expect in safety, governance and operational excellence.

The company’s response has three parts. The business strategy has five pillars, and every AI capability built must align with one of them, with nothing outside those five accepted as either a use case or a capability. A governance committee that Kanioura chairs includes other executive committee members. And the work goes to the board twice a year for validation.

Key Takeaways

1. Other industries are further ahead, which is why the session deliberately avoided healthcare examples.

2. Generative AI is replacing existing predictive models, not only filling gaps, and outperforming them on subtler cases.

3. Improving one step unbalances the whole process. A better detection model is useless without downstream throughput to act on it.

4. Institutional knowledge can be made queryable. Twenty-five years of memos plus twenty years of passed deals produced better comparables.

5. AI can quietly remove discussion from group settings when individuals resolve questions privately beforehand.

6. Pilots can be too small to matter. At PepsiCo’s scale, a pilot cannot generate the value or the credibility to drive adoption.

7. Strict alignment gates AI investment. Capabilities must map to one of five business pillars, with board validation twice a year.