Before modern AI agents, "expert systems" were in vogue in Japan and the United States. Peaking in popularity around the late 1980s through early 1990s, expert systems mostly relied on rigid, rule-based logic, like sprawling decision trees of “if-then” statements. To create medicine-related expert systems, for example, clinicians might codify hundreds or thousands of medical scenarios, much like pre-AlphaZero chess programs that combined systematic searching with hand-tuned heuristics. While more sophisticated than simple brute force approaches, early medical expert systems shared a fundamental limitation with early chess engines: they couldn't adapt or learn on their own and could only apply whatever rules were explicitly written into their code.
One of the most well-known examples was MYCIN, a medical expert system developed at Stanford in the 1970s. It used around 600 hardcoded rules to diagnose and treat bacterial blood infections and meningitis. After asking questions about a patient and their symptoms and labs, it applied its rules to suggest diagnoses and treatments. MYCIN impressed many at the time with its ability to explain its step-by-step reasoning—a major feat in its day—but it was fragile: novel pathogens stumped it because it couldn’t update itself with new knowledge or handle situations beyond what its creators anticipated.
Hardcoding "expertise” is tough for a few reasons:
- Assuming knowledge: Much of medicine’s 'common sense' is so fundamental that medical experts often take it for granted and fail to formally encode it into expert systems—despite it being vital in medicine. If a patient tested positive for Vibrio cholerae, for example, MYCIN would prescribe two weeks of tetracycline to eliminate the bacteria while overlooking the far more urgent need for rehydration. This potentially deadly oversight arose because immediate fluid replacement is, to any doctor treating cholera, so obviously the top priority that experts simply forgot to encode it into MYCIN’s rules.
- Evolving knowledge: Medical knowledge evolves over time, but updating rule-based systems is expensive and labor-intensive—requiring experienced (and costly) human physicians to manually revise their rules. If a newly discovered disease like Legionnaires' suddenly appeared, for example, MYCIN remained oblivious to it unless doctors developed new rules for that novel disease.
Though it performed well in well-trodden scenarios, MYCIN's rule-bound expert system approach was ultimately too inflexible for medicine's messiness; it couldn't adapt, learn from new data, or apply common sense. As its limitations became clear, researchers moved away from purely rule-based systems in favor of more adaptive, learning-based approaches—developments that, skipping many advances between the 1990s and 2020s, eventually set the stage for today’s AI agents.
Note: To see just how far we’ve come in the development of medical AI, check out this medical transcription system. The Nova-3 Medical model, specifically trained on medical jargon, directly addresses MYCIN's inability to handle the complexities of medical terminology. Features like Keyterm Prompting further allow for customization to specific medical specialties, enhancing accuracy for terms that MYCIN would have no knowledge of.
How AI Agents Differ from Old-School Medical AI
AI agents go beyond fixed rules: they learn from experience, adapt to new discoveries, and even recognize when they’re out of their depth—something rule-based expert systems like MYCIN never quite managed. Instead of relying on rigid, pre-programmed logic trees, modern AI agents can pull in live data, current research, and updated clinical guidelines to stay aligned with the latest medical standards. They can also use external tools—calling APIs or even writing their own scripts—expanding their ability to act dynamically.
A major breakthrough in AI agents is how they handle uncertainty and decision-making. Earlier systems like MYCIN assigned crude confidence scores to their rules, forcing a "best match" diagnosis even when uncertain. AI agents, by contrast, can flag ambiguous cases for review rather than guessing. Techniques like prompt engineering and RAG can help AI agents to recognize when they lack confidence, escalating cases to humans or more specialized systems instead of producing misleading answers (not perfectly, though—AI agents still hallucinate).
This shift from passive tools to active partners matters. Most traditional healthcare software was designed to assist with documentation and diagnosis but remained largely reactive—more of a data repository than an intelligent assistant. AI agents, by contrast, proactively analyze data, plan next steps, and learn on the fly. As Yage.ai put it, non-agentic automation feels "like working with a tool," whereas agentic automation feels "like working with a person."
This shift from passive tools to active partners matters because today’s electronic healthcare systems need renovation. Clinicians don’t need yet another piece of software to wrestle with—they need intelligent assistants that anticipate their needs and handle complex workflows autonomously. While AI agents making independent medical decisions like human doctors is still a distant dream, there are already immediate ways that agents can transform healthcare.
AI Agents Immediate Potential in Healthcare
Modern healthcare is drowning in digital 'paperwork,' administrative bloat, and fragmented data. But why? Wasn’t software like EHRs and clinical decision support systems supposed to make medicine more efficient?
EHRs’ biggest flaw? Their clunky user interfaces. Wading through labyrinths of fields, forms, dropdowns, windows, and inboxes forces clinicians to click, scroll, and move their cursors so much that the very digital tools engineered to help them are instead draining their time, adding stress, and stealing focus from patient care.
Start with Admin
AI agents might break this paradox. By parsing unstructured input—like clinical notes or verbal instructions—they could eventually take over record-keeping grunt work, operating with little oversight.
With patient lives at stake, AI agents won’t be making high-stakes clinical decisions anytime soon. Their first proving grounds are administrative tasks like insurance claims, scheduling, referral management, coding, and other time-devouring tasks that burden physicians but pose minimal risk to patients.
Here, AI agents may, for example:
- Pre-authorization checks: Ensure insurance approval before potentially costly procedures
- Error detection: Fix insurance claims before submission and resolve insurance rejections
- Claims tracking: Monitor the entire process, flagging delays or missing info
- Data integration: Merge patient records from multiple sources for a fuller picture
By taking over this type of tedious admin grind—parsing insurance forms, verifying coverage, and nudging staff to address incomplete claims—AI agents can save hours of physicians’ time. Multiple specialized agents can work together—one verifying eligibility, another compiling patient data—so humans only need to step in when something goes sideways.
Beyond Pure Paperwork
In the near future, AI agents may even move beyond simple documentation, assuming more active roles like:
- Virtual medical scribes: AI agents that record visits, suggest next steps, and highlight key details in EHRs
- Remote specialists: AI agents that "play" the role of medical specialists for routine issues, escalating to humans only when needed
- Smart patient communication: AI agents that send timely reminders, adapt their advice based on user feedback, and present data in intuitive ways.
AI agents could transform healthcare software from a rickety, monolithic record system into a responsive, collaborative assistant—one that understands context, anticipates needs, and gets things done—like a human.
In the next part of this series, we’ll discuss the impact AI agents have on healthcare in the future—from doctor selection to pre-visit preparation, and more!