Why Agentic AI Is Not Fairly Prepared for Primetime in Healthcare

Agentic AI is a scorching matter throughout quite a few industries together with healthcare. Everybody appears to be buzzing about AGI, envisioning Jarvis from Iron Man the place AI techniques are able to human-level intelligence and decision-making capabilities.

The joy is comprehensible – all of the sudden there’s a actual alternative to construct fashions and robots that mimic human duties and may work together and interact with people.  Companies are obsessed with this functionality, particularly in industries with labor shortages to switch human staff with AI machines who can automate duties extra effectively, with out errors or monetary compensation.

However in terms of healthcare, I believe we have to take a minute to evaluate at a deeper degree. Generally, healthcare is a extremely regulated trade for good cause and can’t tackle among the dangers and failures inherent with AI that different industries can. We have to consider agentic AI by means of the precise lens of the healthcare trade and its workflows, evaluating the wants of suppliers, healthcare organizations in addition to sufferers, recognizing that healthcare will not be monolithic and may have workflows which are extra acceptable than others for implementing AI.

To know this requires being sincere concerning the strengths and weaknesses of agentic AI. LLMs excel at artistic processes – for instance, ChatGPT can immediately create a brand new tune about Taylor Swift’s newest album within the model of Dr. Seuss with minimal prompts. Comparatively, rules-based engines are good at structured output. Take into account a immediate for a self-driving automotive: “When a visitors mild is crimson, THEN cease.” What’s fascinating about agentic AI is that it’s going to possible fall someplace within the center – partly rules-based and partly artistic, and navigating that center floor and the suitable danger and reward of when it ought to be utilized in healthcare workflows is the massive problem.

Essentially the most helpful heuristic I’ve discovered on this matter is trying on the idea of danger versus consequence. I outline danger as the proportion that one thing would fail and consequence as the end result of that failure.

In eventualities the place workflow dangers are excessive, you don’t need the agentic course of to personal it – the truth is that each AI mannequin will fail sooner or later and the price of that failure could possibly be too excessive in terms of healthcare outcomes.

Listed here are two examples the place agentic workflows wouldn’t work: 

• Authoring or defining superior directives (end-of-life planning). There’s clearly a artistic and interpretive workflow right here that requires empathy, human expertise, and judgement on when to information vs. when to hear; as a result of the supply of knowledge (folks and their conditions) should not all equal. It’s additionally a scenario with excessive penalties that you simply don’t wish to get unsuitable in any means form or type.
• Managing triage in an ER — a chaotic, fast-moving atmosphere. Individuals are finest suited to make fast selections — there’s no time to enter information into an agent.

Nonetheless, listed below are two examples of the place agentic AI would work in healthcare:

• Unlocking EHR information utilizing an agent to automate a sequence of duties that require it to navigate a person interface. Enterprise degree software program has achieved this earlier than. It was once referred to as RPA or Robotic Course of Automation, however now agent-powered processes can do that with rather more resilience.
• Reviewing affected person charts to verify emergent persistent circumstances weren’t missed by clinicians.

Proper now, I believe agentic AI is at a stage the place it’s going to fail if it begins telling docs what to do and takes over resolution making — and worse, damage the belief of everybody (by its affiliation) with synthetic intelligence in drugs. When affected person security, empathy and human judgment take priority over value financial savings and potential effectivity enhancements, a human should be within the loop. However automating duties, word taking, extracting information — the tedious, guide processes in healthcare that don’t require human intervention in each step — is a good place to start implementing AI. The healthcare trade ought to be taking a look at how AI can discover and current related information with context to allow human clinicians to make knowledgeable selections, liberating up suppliers’ time to cut back burnout and permit them to supply extra personalised affected person care to enhance affected person outcomes.

Picture: filo, Getty Photographs

This publish seems by means of the MedCity Influencers program. Anybody can publish their perspective on enterprise and innovation in healthcare on MedCity Information by means of MedCity Influencers. Click on right here to learn how.