Every day, health plans are faced with the daunting task of reviewing massive electronic health records (EHRs) and scanned documents to find the critical clinical information they need to correctly pay a claim.
A single hospital encounter, however, can generate hundreds of pages of records, with vital signs, lab values, and physician notes scattered throughout. For human reviewers, piecing together a patient’s story—such as how their oxygen saturation (SpO2) changed over their hospital stay—means searching, scrolling, and cross-referencing across endless pages. It’s a process that is not only slow and tedious, but also vulnerable to human error.
At Machinify, we’re using a new class of artificial intelligence—agentic AI—to fundamentally change this workflow.
What is Agentic AI?
To set the scene as we begin to explore agentic AI, let’s revisit the evolution of AI. AI development has proceeded through several phases:
- Symbolic AI (1950s–1980s): Rule-based systems used logic and pre-defined instructions to perform narrow tasks (e.g., expert systems)
- Machine Learning (1990s–2010s): Algorithms trained on data to make predictions or classifications without explicit programming (e.g., decision trees, neural networks)
- Deep Learning (2010s–present): Advanced neural networks with multiple layers capable of pattern recognition in unstructured data, leading to breakthroughs in image, speech, and language processing
That brings us to the current frontier: agentic AI. Unlike traditional AI models that answer a single question or extract a piece of data in isolation, agentic AI systems are designed to act, reason, and adapt over multiple steps. Think of agentic AI as a digital teammate that doesn’t just “read” a page, but proactively plans, asks follow-up questions, seeks missing information, and reflects on its own outputs—much like a human reviewer would.
The Table Extraction Challenge in Healthcare
Consider the challenge of extracting a patient’s lab values, such as SpO2, from a 500-page hospital record. The data is rarely in one tidy table. Instead:
- Values may be scattered across different sections or buried in free-text, handwritten notes
- Table headers and column labels (e.g., “Date”, “Time”, “SpO2 (%)”) might appear pages away from the values they belong to
- Dates and times may be inconsistently formatted, missing, or listed only once at the top of a section
- LLMs (large language models) often “hallucinate” information—making up values or misaligning dates/times—if they can’t find the answer directly
For compliance, clinical review, or payment integrity teams, missing or misattributing just a few of these values can impact audit outcomes, reimbursement, or even patient care
Machinify’s Agentic AI in Action
At Machinify, we built an agentic AI workflow that acts more like a diligent clinical auditor than a basic extraction tool.
Here’s how it would work for the above SpO2 example:
- Proactive Search: The AI scans the entire document for any mention of SpO2, intelligently following the trail even when the value, date, and time are on different pages.
- Context Gathering: If the AI can’t find a date or time for a value on the current page, it automatically “looks back” at previous pages—just as a human would—to find missing headers or context.
- Reflection and Self-Check: The AI reviews its own work, checking for mistakes like hallucinated data or incorrect time associations. If something doesn’t add up, it goes back and tries again, learning from its previous attempt.
- Time Series Construction: It pieces together the scattered values into a clean, structured timeline—turning a messy 500-page chart into a single, reviewer-friendly table.
The Business Impact
This evolution isn’t just about implementing cool technology. The real impact is on the speed, accuracy, and efficiency of healthcare operations:
- Reviewers no longer have to manually search and collate lab values. With a single click, the AI serves up the patient’s SpO2 (or any other value) in an ordered timeline—saving hours of manual work.
- Error rates drop. The AI’s self-reflection step means it’s constantly checking for mistakes, leading to fewer audit errors and less risk of missing critical information.
- Faster, more consistent patient reviews. By automating the grunt work, clinical reviewers can focus on what they do best—clinical decision making, not wrangling data.
- Scalable across use cases. This approach applies to any numeric value—labs, vitals, medications—unlocking similar value for every chart type.
Looking Forward
Agentic AI is redefining what’s possible in clinical document review. By building systems that act, adapt, and self-correct, we’re taking a major step toward a future where human reviewers are empowered by AI, not replaced by it. The end result? More accurate reviews, faster patient care, and ultimately, better outcomes for everyone.
Interested in how agentic AI can transform your clinical review workflow? Reach out to learn more about our solutions.
