AI Risk and Reward Across Industries: Lessons From NORC, IHG and AT&T

Watch the full session · 48 min

How do AI's risks and rewards change from public health to hotels to telecom? At COLLIDE in Atlanta, leaders from NORC at the University of Chicago, IHG and AT&T discussed hallucination and bias, proving business value, governance practices and why scaling AI too fast can sacrifice quality.

AI promises a lot, but turning it into results looks different in public health than in hotels or telecom. At COLLIDE in Atlanta, a panel of leaders from three industries compared the risks and rewards of AI and advanced analytics: Amelia Burke-Garcia, Director of the Center for Health Communication Science at NORC at the University of Chicago; Giang Do, Director, Customer 360 at IHG; and Frances Boykin, Director of Advanced Analytics at AT&T, who moderated.

Boykin opened with a caveat that the panelists' comments did not reflect the views of their organizations.

How do AI risks differ in highly regulated industries?

Burke-Garcia, whose clients include federal agencies, said the starting point is basic: know the regulations, and expect them to change as AI does. In health, she said, the tools cannot hallucinate, because people may make decisions about their health based on what the tools say.

She also pointed to fairness and robustness. Some health communication tools already on the market fall short, she said, using only white avatars, working only in English or giving inaccurate information. Her team's work on what it calls health communication AI starts from inclusion to avoid that. "If we don't start from that place, we're going to regress to the mean," she said.

Measurement is unsettled too. Burke-Garcia said the field has not agreed on what accuracy means or how much risk is acceptable when 100% accuracy isn't possible.

Do added the concern of where data leaks to, and whether a person's record becomes one more piece of training data. He said he likes working with privacy and governance teams because they act as "another filter, another check" on what a data scientist is trying to do. Rules vary by state, by country under laws such as GDPR, and beyond, so his team works with privacy experts and segments by country.

What are the biggest barriers to AI adoption?

Public health is "not always the first penguin," Burke-Garcia said. Fear, a lack of knowledge and limited budgets compared with the private sector slow adoption. She cited social media as a warning: 20 years of not studying its effects left the field behind, and she called for longitudinal research now on how young people engage with AI.

For Do, the highest-impact barrier is showing value. Data scientists often can't promise a specific return before a project starts, and many struggle to translate technical gains into business terms. Telling a leader a forecast will improve from 88% to 95% invites the response "okay, so what," he said.

He also named past failures that made leaders wary, and a growing fear among workers of training the machines that will replace them. His answer is to acknowledge that fear honestly and help people keep upskilling so they are not left behind.

Boykin added data itself. At a company like AT&T, with many acquisitions and mergers, "data is everywhere," she said, and models are only as good as what goes in.

How can organizations keep AI models secure and trustworthy?

Boykin said AT&T has a generative AI governance team with leaders from several organizations. Burke-Garcia described a similar working group and policies at NORC, plus private sandboxes where staff can upload data safely without violating the terms under which the data was obtained.

Do relayed a conversation from a roundtable with a banking leader about tools that evaluate models for fairness and bias. Human oversight comes first, he said, but people also need tools, because AI now personalizes too quickly for humans alone to check. He called fairness evaluation an area where demand is "exploding."

How can AI improve health communication at scale?

Answering accurate, tailored and empathetic health questions at today's scale is impossible for humans alone, Burke-Garcia said, a gap made clear during the pandemic. Her health communication AI project aims to use peer-reviewed, publicly available information from trusted sources to answer people's questions anywhere, at any time.

Medical and health experts can help build and fact-check these tools, she said, reducing their burden and keeping answers accurate. Do described a similar role for data scientists: any AutoML tool will produce a result, but it is the data scientist's job to validate it.

How should companies scale AI across the business?

Do warned against leaders who simply want more models because AI is popular. Mass scaling sacrifices quality, he said, so teams sometimes need to say no. "Don't just build models for the sake of building models," he said.

Burke-Garcia said NORC makes its business case by showing that what is built for one use case can be adapted for others across its work, from health to transportation and education. Boykin said small teams are a common obstacle, since five people can't build, review and measure AI at scale. She also noted that data can be monetized, citing a feature store AT&T built for its own use and later offered outside the company.

In a closing lightning round, Burke-Garcia said "there's just as much risk with not doing something than with doing something." Do chose "value and human," and Boykin closed with a line she uses for everything: "Chance favors the prepared mind."

Hear more cross-industry AI lessons at COLLIDE 2027

COLLIDE, the largest AI conference in the Southern U.S., brings more than 1,000 enterprise data and AI leaders to Atlanta for keynotes and panels like this one. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center. Registration details are coming soon.

Watch: Cross-Industry Applications:The Risk & Reward of Realizing the Potential of AI & Advanced Analytics on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights