Carriers Re-evaluate AI Strategies for Medical Record Review
Following a surge of interest in late 2025, insurance carriers are re-evaluating their strategies for implementing AI in medical record reviews. While initial development of these tools has become more accessible for small teams, many organizations are discovering that the true challenge lies in maintaining accuracy and managing costs after the first version is deployed. With adjusters typically spending between 2 and 6 hours daily reviewing records, the pressure to automate remains high, yet the operational realities of these systems have become clearer by the second half of 2026.
The decision to build an in-house tool offers carriers control over workflows and roadmaps, but it introduces significant, recurring responsibilities. Amir Tavafi, Chief AI Officer at SiftMed, notes that while early demos often show 80 to 85 percent accuracy, the investment required to maintain that performance is substantial. Building requires managing complex data, such as handwritten notes and scanned documents, and ensuring the system can handle the structural nuances of medical files. Furthermore, carriers must account for the rapid pace of model updates, which necessitates constant regression testing to prevent accuracy drops.
Beyond technical maintenance, carriers face substantial governance and security hurdles. Managing Protected Health Information (PHI) is a critical concern, particularly given that the average breach cost is $7.42 million. Insurers must also comply with evolving regulations, such as the NAIC model bulletin and OSFI Guideline E-23, which becomes effective for federally regulated insurers on May 1, 2027. These regulatory requirements place the burden of model governance squarely on the carrier, regardless of whether they build or buy.
Buying a specialized platform shifts the day-to-day maintenance, such as model testing and integration support, to a vendor. While this approach requires relinquishing some control over the roadmap, it allows carriers to leverage vendor expertise in handling edge cases. However, Holly Hill, CEO and co-founder of SiftMed, emphasizes that accuracy must be defined before summarization begins. She notes that effective systems should sort, de-duplicate, categorize, and date records before any data is summarized. Regardless of the chosen path, carriers must ensure that output accuracy is measured and re-tested over time.
General AI assistants, while useful for drafting emails or answering policy questions, often fall short for core medical review tasks. These tools may lack the capacity to handle large scanned files, struggle with page-level citations, and require significant development to integrate with claims systems like Guidewire or OnBase. Furthermore, organizations using consumer-grade tools may face limitations regarding the control and security of sensitive PHI.
For insurance professionals, the shift from initial AI implementation to long-term operational management highlights the need for rigorous vendor vetting and internal testing protocols. As regulatory scrutiny increases and model turnover accelerates, the focus is moving away from the speed of deployment toward the sustainability and security of these automated workflows.
Why it matters
As AI adoption matures, insurers must shift their focus from the speed of initial deployment to the long-term costs of maintenance, regulatory compliance, and data security.
FAQ
What are the primary regulatory requirements for AI model governance?
Insurers must navigate the NAIC model bulletin, adopted by about half of US states, and OSFI Guideline E-23, which is effective for federally regulated insurers starting May 1, 2027.
Why is building an in-house AI tool considered a high-cost option?
Building involves recurring costs for tokens, maintenance, and constant re-testing as models change, alongside the need for infrastructure and engineering resources to handle messy medical data.
What is the main difference between buying and building AI tools?
Building provides full control over the roadmap and workflows but requires the carrier to own all maintenance and security. Buying shifts day-to-day maintenance and model testing to a vendor, though the carrier retains the ultimate regulatory responsibility.
Source: Claims Journal. See the original source for full details.
Scan with your phone
Scan with your phone