Send in the Co-Pilots

After a crazy hot summer, hot from NAMPI, and fresh from a customer trip to the Midwest, an interesting prospect kept coming up in conversations. Would introducing co-pilots into program integrity operations save time? PIUs and MFCUs have been a buzz asking how GenAI powered co-pilots can be used and trying to understand their best use cases.

What Is a Co-pilot, Again?

We covered co-pilots in another RIViR Reads. In case you missed it, co-pilots are GenAI powered chatbot assistants that automate routine tasks. Co-pilots are typically built into an application, and you can ask them to analyze data on the screen or summarize information to make it more consumable. Properly executed co-pilots can be a productivity boost. In our world, co-pilots are alluring because the average PI investigator is combing through mountains of claims and records even after analysis.

Good co-pilots have a few key aspects. First, they are trained and attuned specifically for industrial use cases. Companies integrating ChatGPT or Gemini capability raise several risks of hallucinating misinformation and leaking data because health claims, medical records, and personally identifiable information are specific types of data that require special handling. Even a custom finetuned GenAI model isn’t specific enough to co-pilot for a PI professional.

Second, co-pilots should be context aware enough to know what the human user is looking at on the screen and interested in at a particular point-in-time while being flexible enough to expand its reach and incorporate related and additional data.

Third, co-pilots must also understand categories of information. It should know the difference between medical records, claims, personally identifiable information, and supporting documentation. When I speak with PI professionals, they admit that mistakes are not only frowned upon but must be avoided. Mistakes can have devastating impacts on providers, PI professionals, and the recipients we serve.

Co-Pilot Security

After years of mainstream use, we know ChatGPT and other large language models can be susceptible to hallucination and specialized attacks that may cause information leaks. Co-pilots can mitigate these issues by layering security. First, crafting clear, concise, and thorough system instructions help the underlying LLM technology understand critical PHI while still performing analysis and summarization. Then, configuring reliable data loss prevention (DLP) products to understand prompts and prompt injection attacks can reduce chances of data leaks. Layering programmatic checks on top of LLM API calls is another mitigation tactics.

Co-pilots can examine medical records and claims faster than a human. PIUs looking at co-pilots to do more with less, and boosting productivity can reduce errors by carefully configuring and training co-pilots. PIUs can safely integrate co-pilots in their workflows by enforcing security to reduce data leaks, maximizing their understanding of context, and using them for what they’re best at: text analysis, scanning, and summarization. As PI professionals, we all face struggles in the tit-for-tat battle against fraudsters dreaming up new schemes as quickly as we find them. Properly trained and tuned AI co-pilots can be the equalizer we’ve been waiting for.

Co-pilots featuring all these capabilities can be built with the right technique, technology, and experience in the space. Co-pilot assisted applications like ClaimsVue and TENEX are specially attuned to PIU and MFCU use cases to help PI professionals quickly find important details in reams of information while being flexible, context aware, and secure. As PI professionals, we all face struggles in the tit-for-tat battle against fraudsters dreaming up new schemes as quickly as we find them. Properly trained and tuned AI co-pilots can be the equalizer we’ve been waiting for.