Putting an ROI on AI in Community Oncology: Larry Bilbrey

Community oncology practices need to move unstructured data into accessible fields before artificial intelligence (AI) can support clinical decisions, and they can judge AI purchases by whether a tool generates revenue, generates savings, or improves quality of life, according to Larry Bilbrey of OneOncology.

In an interview with The American Journal of Managed Care^®, Bilbrey discussed the data barriers that limit AI, the difficulty of projecting return on investment (ROI), and a 3-part framework for evaluating AI solutions.

Why Unstructured Data Limits AI

Bilbrey said a large share of health care data still lives in unstructured formats and that faxes remain the biggest and easiest way for data to move between organizations. Bilbrey noted involvement in a federal task force working to make interoperability more accessible but said the field still has a long way to go.

That gap matters because AI can only work with the information it can reach. "If AI is going to help a physician make a decision on a patient, but it only has access to 3 pieces of data, and there's 7 other pieces of data that are stuck in a scanned PDF somewhere, then that AI is not going to give that doctor a very good recommendation," Bilbrey said.

Bilbrey described the first step as getting unstructured data into structured fields or into another place where AI can access it so the technology has a complete picture. “There's more complicated steps that happen after that, but that's the beginning," Bilbrey said.

According to Bilbrey, it is very hard to put an ROI on AI right now, and much of the case for it rests on the promise of the technology. Practices expect it to make a difference over the long term, but that is a difficult basis for a large purchase. "It is hard for a practice to say, 'Oh, I'm going to sign a 6-figure and even a 7-figure contract to just hang it on hope,'" Bilbrey said.

A 3-Part Framework for AI Decisions

Bilbrey sorts AI solutions into 3 buckets. The first is whether a tool generates revenue. Bilbrey pointed to revenue cycle management AI that supports prior authorization for drugs or coding of evaluation and management charges, where a practice can attach actual dollar amounts to the result. That makes it the easiest sell.

The second bucket is savings. Bilbrey said this case is somewhat tougher, but a hard number can still be assigned. One example is a tool that supports reporting for the Enhancing Oncology Model and produces savings through that work.

The third bucket is intangible and, according to Bilbrey, the hardest to evaluate: quality of life. The question there is whether the improvement is large enough that the buyer is willing to pay for it. Bilbrey used AI scribes as the example. Scribes do not generate revenue or savings, but they improve life for physicians, and physicians are willing to pay for that.

The ownership structure explains why. "At the community practices, the physicians are the owners generally," Bilbrey said. "They're the ones paying the bills."