In 2025, what have we learned about AI in healthcare?

It’s been one heck of a year for artificial intelligence in healthcare. In 2025, we turned the corner from cautious speculation to a full-steam-ahead inundation of tools, platforms, models, agents, and optimizers powered by the latest and greatest in AI technology.  

During 2024, AI investment started to go through the roof, with 1 in 4 venture capital dollars devoted to AI projects, leading to the healthcare industry racing past other sectors this year to become a leader in AI adoption – in terms of volume, at least, if not measurable impact on key performance indicators. 

The jury is still out, to some degree, on the latter. Despite the FOMO-driven flurry of adoption, some healthcare organizations are still finding themselves struggling to extract meaningful value from their efforts, with a huge number of pilots failing to scale appropriately in an environment replete with unresolved questions around governance, compliance risks, patient safety, and the underlying ethics of infusing AI into certain areas of the care experience.  

As we close out a year that seems to have flown by in a flash, what has the healthcare industry accomplished, what’s left to work on, and where is artificial intelligence likely to take us in 2026 and beyond?  

By making itself invisible in the clinic, AI is really starting to shine

If there’s one thing that nearly everyone can agree on, it’s that ambient listening has made an indelible mark on the care environment this year. It’s already a billion-dollar industry with about one-third of large health systems using the technology. And it’s producing measurable results where they matter most to clinicians and executives alike: reduced burnout, higher rates of productivity, better provider experiences, and improved patient satisfaction. 

In 2025, we started to see leaders trying to translate these early gains into even more ROI by measuring additional factors such as documentation quality, provider retention, and clinical capacity expansion. 

There’s still more work to do in order to accuracy and consistently measure and compare these datapoints – and to make sure that patients accept its use and have full transparency around who (or what) is communicating from their doctor’s email address.  

But generally, ambient scribes are considered an early and pretty unanimously successful use case that can be felt in the trenches without being seen by patients, potentially building evidence for additional ways to bring AI into the workflow. 

Governance, governance, governance…and more governance

Merriam-Webster chose “slop” as the 2025 Word of Year, a reference to the poor quality AI-generated content that is being overwhelmingly force fed down the throats of consumers in every facet of our lives. 

They could have just as easily chosen its opposite, “governance,” which has been the prevailing theme of 2025’s discussions about AI in digital health. 

AI needs a lot of governance, at every point in the process – and at every level of regulation and rulemaking. From algorithmic bias and hallucinations that can get patients killed to privacy and security risks that can expose the enterprise to costly and disruptive ransomware, healthcare organizations need to obsessively stay on top of governance, especially under a federal administration that is aggressively clearing the path for AI vendors to do as they please. 

Responsible governance starts with creating accountability within the organization at the highest levels, and infusing that sense of shared ownership and vigilance through every layer of the org chart. 

Leaders have to be willing to ask tough questions of their partners – and of themselves – about how tools are being developed, deployed, monitored, and continuously improved in both the clinic and the back office. 

That’s because industry research shows that investing in robust and comprehensive governance activities, guided by one or more of the dozens of available frameworks for the industry, is an even greater predictor of success than just pouring funding into AI technologies. 

There’s simply no substitute for getting it right from the get-go, especially as cybercriminals start to leverage AI as their new weapon of choice, taking advantage of organizations that haven’t been able to invest the time or resources in bulking up their defenses. 

Preserving the human touch where it matters most

AI might be living the high life, but humans have had it rough this year across the healthcare industry. Ongoing staff shortages have left the remaining workers doing the jobs of two or three people, and we haven’t yet reached the point where AI agents are capable or trustworthy enough to reliably take over the routine functions they’re predicted to. 

As a result, we might be in the most dangerous period of AI development: the workflow is no longer fully manual, and human workers are relying more and more on AI tools…but the AI is not yet mature enough to safely and consistently manage tasks without strict oversight and stringent review processes to ensure appropriate outputs. 

The tension is clearest in areas such as mental healthcare, where overworked and understaffed providers are understandably eager to embrace AI to expand capacity and make sure patients get access to critical care.   

More than 122 million people live in areas where it’s close to impossible to access human mental and behavioral healthcare providers, yet researchers are warning that AI chatbots, even if they are getting better at identifying distress and showing empathy, still pose a safety risk in therapeutic applications. 

In 2025, Illinois was among the first states to start regulating how AI can be used in a mental health context, reinforcing what is generally still believed to be true: that AI should not be allowed to make decisions on its own without oversight from licensed professionals. 

But AI’s role in the loss of the human touch in healthcare is likely to be more insidious than that.  

AI might not be replacing human providers wholesale, but it’s definitely co-opting our human capacity to think critically and communicate authentically, which could be just as bad (if not worse). 

In the frenzy to insert technology into the gaps where people used to sit at desks, leaders are inadvertently creating an environment of self-referential learning, wherein AI provides data, humans act on that data because the AI says so, and the AI therefore doubles down on believing that data to be correct.   

Humans start to lose their critical edge after being told again and again to rely on their AI companions, and mistakes start to propagate through a system that no one is really watching closely enough, especially when “efficiency” becomes the only KPI that really matters. 

The slop keeps getting sloppier, humans keep getting less human – and patients could suffer as a result. 

The solution? More of that governance stuff. Healthcare organizations must ensure that they are applying AI solutions in a strategic manner guided by iron-clad principles of safety, accuracy, equity, necessity, ethics, and patient-centeredness. If it doesn’t meet the highest bars we can set for ourselves, it shouldn’t be part of the care process.  

Strong governance, a focus on safety, and a non-negotiable commitment to keeping skilled humans actively in the loop will be the only way to capitalize on the lessons learned this year – and to navigate the increasingly hybrid AI-human environment that we’re headed towards in 2026.  

It’s still possible to get it right, if we take a moment to reflect on the successes and failures of a tumultuous 2025 and carry forward the best of what the industry has already accomplished. 

With the first wave of Medicaid cuts approaching in 2026, and an unpredictable policy environment likely to continue bringing challenges to strategic planning, the next 12 months will be a trial by fire for those who have started to feel like they’ve got a handle on what AI can do for them.  

The key to success will be staying abreast of policy changes and technology developments, ensuring that AI adoption decisions are made with an eye toward the right metrics for success, and encouraging human staff members to actively participate in the design and execution of AI rollouts to ensure that both types of workers can perform their tasks in harmony while keeping patients safe, healthy, and engaged. 

Jennifer Bresnick is a journalist and freelance content creator with a decade of experience in the health IT industry.  Her work has focused on leveraging innovative technology tools to create value, improve health equity, and achieve the promises of the learning health system.  She can be reached at [email protected].