This is the fourth and final part in a four-part series examining the critical strategic dilemmas healthcare executives must navigate in 2025. Read part one, part two, and part three.
In healthcare, we've historically demanded that our technologies be all but perfect. Radiation therapy equipment must deliver precise doses within tight tolerances. Implantable cardiac devices must function flawlessly for years. Surgical robots must behave exactly as surgeons expect, every single time.
But AI tools are fundamentally different. They operate on statistical rather than deterministic logic, they’re designed to learn and evolve, and they often work with unpredictable inputs such as patients’ spoken words. As such, “perfect” reliability in AI remains largely unachievable.
This creates healthcare's fourth fundamental AI dilemma in 2025: Should you wait for highly validated AI solutions that meet healthcare's exacting standards for safety and reliability? Or should you expect that early AI tools will be imperfect — just as medical residents are imperfect — and focus on managing risk while capturing benefits?
As with all dilemmas presented in this series, this is to some extent a false binary; different AI use cases will demand different degrees of reliability. Still, it’s important to have a clear vision for how much imperfection you’ll accept from your AI tools, and how you’ll manage AI’s inevitable errors.
The case for 'perfect': Flawed AI is just too risky
Unlike an incorrect Amazon recommendation or a flawed Google search result, AI mistakes in healthcare could theoretically harm or even kill patients (although, thankfully, there have so far been no public reports of patient deaths directly attributable to AI). Consider these risk factors:
News reports have revealed AI tools hallucinating false and even dangerous content. An Associated Press investigation raised alarms about OpenAI's Whisper transcription tool, which is used in some healthcare AI products, “making up chunks of text or even entire sentences.” In one case an innocuous statement about an umbrella was transformed into disturbing text about a “terror knife.”
Law enforcement officials allege AI vendors have misled hospitals about their products’ capabilities. In late 2024, Texas's attorney general secured a settlement with Pieces Technologies, a healthcare AI company that had deployed its summarization tools at several major Texas hospitals. The attorney general alleged that Pieces made “deceptive claims about the accuracy of its healthcare AI products, putting the public interest at risk.”
Patient safety experts rank AI risks as healthcare’s biggest technology hazard. In December 2024, ECRI — a leading patient safety organization — ranked “risks with AI-enabled health technologies” as the number one health technology hazard for 2025, ahead of concerns like cybersecurity threats and medical device fires. “AI is only as good as the data it is given and the guardrails that govern its use,” noted ECRI’s CEO Marcus Schabacker.
Healthcare's strict regulatory environment may also make perfection non-negotiable in certain contexts, leaving little room for the “move fast and break things” ethos common in other industries.
The case for ‘good enough’: Imperfect AI still beats the alternatives
Still, there’s a compelling counterargument: If we refuse to use “imperfect” AI in healthcare, the result won’t automatically be that patients receive “perfect” care instead. At best, they’ll receive care from all-too-imperfect humans – and, at worst, they simply won’t receive care at all.
Consider these arguments for “good-enough” AI:
Health systems report that AI is already enabling care that wouldn’t otherwise happen. When WellSpan Health deployed Hippocratic AI's virtual assistant for colonoscopy preparation outreach, nurses supported the initiative despite the tool's limitations. Why? They recognized that many of these calls simply wouldn't happen otherwise due to staffing constraints.
Research demonstrates that “imperfect” AI already outperforms human clinicians in some diagnostic tasks. A 2024 JAMA Open Network study comparing diagnostic reasoning between AI and physicians found that human doctors answered only 74% of diagnostic questions correctly. By comparison, GPT-4 scored 92%.
Real-world experience shows that, even when AI’s outputs are imperfect, humans can — and will — make them better. At UW Health, where AI drafts patient messages, only 1% of AI-generated texts were sent as-is. For the rest, staff incorporated significant changes to make AI outputs work better for patients.
If you wait for “perfect” AI to arrive, you’ll miss years of learning along the way. AI requires significant technical, clinical, and cultural changes — a multi-year journey for most health systems. If you delay starting this journey until "perfect" AI tools emerge, you'll lack the infrastructure and expertise needed to deploy them effectively.
Which should you choose: 'Perfect' or 'good enough'?
Every health system needs both perfectionism in some areas and a tolerance for "good enough" in others — but your emphasis should reflect your organization’s philosophy, circumstances, and risk tolerance.
Our guidance: Consider prioritizing a "perfect" approach if:
You're highly risk-averse in AI implementation due to past technology failures or a conservative organizational culture.
You're limiting your AI to only a few pilots and can afford to be selective about which tools you adopt.
Your governance structures aren't yet ready to manage AI risks effectively, creating additional vulnerability.
In that case:
Start with behind-the-scenes applications first. Begin with proven back-office tools, such as revenue cycle applications, before venturing into clinical decision support.
Build comprehensive governance before you need it. If you wait to create your AI governance structures and rules until AI achieves some standard of “perfection,” you’ll be ill-prepared to embrace these tools when they arrive.
Stay vigilant for rapid improvements. Today's flawed AI tools may improve rapidly, and you'll want to recognize when previously rejected technologies become viable.
On the other hand, consider prioritizing a "good enough" approach if:
You have the institutional credibility to take smart risks and the cultural comfort to acknowledge and learn from mistakes.
Staffing shortages mean imperfect AI beats none at all in specific use cases where you simply cannot meet patient needs otherwise.
Your monitoring processes, including "humans in the loop," can catch and correct AI errors effectively.
In that case:
Consider being transparent about AI use with patients. For instance, when Wellspan’s semi-autonomous AI “agents,” which run on Hippocratic AI’s platform, call patients to check in over the phone, they begin by self-identifying as AI. Similarly, many AI-generated messages to patients identify as “generated by AI” and “reviewed by” a human.
Equip staff to spot and correct AI errors. This can include sending senior team members to AI Catalyst’s “AI Bootcamp,” and making use of our other training materials, such as our AI Glossary and our FUMBLE Framework for cutting through AI hype.
Test your safety nets regularly. Consider "AI fire drills": Deliberately plant AI errors to test your risk management practices and ensure your safety nets actually work when needed.
So how are health system leaders navigating this dilemma? When we posed the "perfect vs. good enough" question to health system CEOs and board members at The Health Management Academy's recent Trustee Summit, 87% favored the "good enough" approach while only 13% leaned toward "perfect."
Today's leaders, perhaps recognizing both AI's potential and the competitive imperative to adopt it, appear willing to accept some imperfection in their AI tools — as long as the risks are properly managed.
Questions to consider:
Outside of AI implementation, where does your organization’s philosophy currently fall on the spectrum between perfectionism and pragmatism in technology adoption?
What specific AI use cases in your organization would benefit most from a "perfect" approach, and which might be better served by "good enough"?
How will you communicate your approach to perfection vs. "good enough" to patients, clinicians, and other stakeholders in ways that build trust rather than undermining it?