Skip to content


Healthcare Claims: An Opportunity for AI

Are you a healthcare entity that struggles with an outsourced billing and claims processing solution? Perhaps your annual audits bring to light underlying issues with your process that you wish you could get ahead of? If either or both questions ring true for you, you are not alone; advancements in artificial intelligence (AI) could be your answer for fast, efficient, and cheaper billing and claims processing solutions down the road.

To delve into this topic, an interview was conducted with Jamison Rotz, CEO and Founder of Nearly Human Intelligent Software, a Harrisburg based company that offers artificial intelligence solutions for businesses. The following is the narrative of questions and answers.

In what ways can AI help the claims process be more efficient?

AI Models can help significantly in the claims process, especially when it comes to coding. Since there’s a lot of historical data for coded claims, models can be trained to analyze unstructured medical history, such as charts, referrals, and correspondence, to determine codes. Language models can summarize pertinent information and validate that policy requirements have been met when submitting claims.

Are there any pitfalls to the use of AI in the billing or claims process?

Quality/Reliability is the biggest concern. LLMs (Large Language Models) do tend to hallucinate, so there are a lot of methods that need to be put in place to prevent this from happening. Additionally, given the state of maturity of the technology, efficient “human in the loop” systems must be implemented alongside the AI to ensure that claims are accurate and appropriate. System design needs to provide for quick and easy collection of human feedback when anyone in the process chain detects an error, so the transaction can be flagged for use in improving future versions of the system. It’s important that the systems don’t erroneously restrict access to care or cause inappropriate claims to be paid.

Would the design and implementation of an AI application be cost effective compared to the salaries of medical coders, billers, claims processors, etc.?

Yes, but the system needs to be deployed at some level of scale to get the best return on investment. Medical claims staff are well paid, and the transaction volume is enormous, so allowing claims staff to be 3 to 10 times more productive would pay back quickly.

In what capacity would an AI application rely on human operation/direction and internal audit?

As mentioned before, AI assisted review and efficient processes to keep humans in the loop are critical to reducing error rates and maintaining compliance. This typically involves a combination of review/approval features, automatic linking to source data for verification when needed, and potentially even adversarial model ensembles (models checking models and referring potential errors to humans for review).

Would an AI application be better suited to perform the tasks and organize them into high or low-risk items to be audited? Or to audit the work or humans?

This depends on the models used to implement the solution. Generative models are excellent at processing language but terrible at assessing their own confidence in their results; this is where the addition of shallow learning statistical models in the solution could help assess the output of the generative models. Given the state of the technology, you’d want implementations to scale into AI automation – lots of human audit at first, providing feedback to the model for reinforcement learning, then moving towards automated quality assessments with humans reviewing items that the model assesses as problematic, and humans sampling the transactions for quality review.

How would a healthcare entity begin the process of implementing AI in their claims process?

The first thing to do would be to gather the data set. You’d want as many correct claims (under current standards) as you can gather, along with data from adjacent medical systems related to the claim. Identifying the authoritative sources of each piece of data will also be important when designing the system. If data elements change over time during the claims process, it’s important to collect the audit log of these fields; remember, the system will be processing the data it has available at the time of the decision. Often that differs from the final state of the data element in a closed record. I’ve seen lots of failed attempts because a model was trained on data in its final state, but then it’s deployed to a process where it’s trying to make decisions at a midpoint in the process and failing because it was trained on data that was affected by downstream decisions.

This may go without saying, but the other thing to make sure you have is a clear understanding of (and process documentation for) exactly what the humans are doing in the process that you’re trying to automate. Ensure you have data from all of the systems they are using. If there are points where people are manipulating data in ways that are not captured by the system, this needs to be taken into account.

This prep work will save a lot of discovery (read time and money) during the development phase.

Finally, work with your technical solution provider to create an agile roadmap that attacks processes in pieces (sprints), with each sprint fielding working capabilities that your team can use and evaluate. Prioritize your capability with the highest value items first and use an MVP (minimum viable product) mentality. Don’t try to implement a full system before you start testing and evaluating its components.

Stay tuned for a follow-up article diving more into the different types of artificial intelligence models and further insight on how these can be applied to your entity’s processes.

If you would like more information regarding the implementation of artificial intelligence solutions for your company, you can connect with Nearly Human Intelligent Software. Also, if you are an entity seeking services from a firm experienced in many industries and specialties: McKonly & Asbury would be happy to help. We currently offer a full suite of assurance, tax and consulting services to healthcare entities, as well as the full suite of SOC services to clients in a broad variety of industries. Learn more about McKonly & Asbury’s Healthcare Practice by visiting our website or by contacting the Healthcare Practice Director, Janice Snyder, Partner.


About the Author

Brian Doheny

Brian joined McKonly & Asbury in 2022 and is currently a Staff Accountant with the firm. He is a member of the SOC & Internal Audit Segment, auditing Service Organization clients in completion of SOC reports.

Related Industries

Subscribe to Our Newsletter

Contact Us