The health care industry is learning from its early successes and failures with AI adoption. Peer–reviewed studies have yielded lessons from a variety of use cases. Now that we have a little experience under our belt, what lessons can we take into 2026?
Focus on risk management
One overarching theme from the early instances of AI implementation is that problems arise when companies lose sight of the need to start small and build trust. As an industry, health care AI adopters need to anticipate malfunctions, such as what happens when an agent goes down. As users become more reliant on automations, what’s the plan when a major outage hits? This is different from traditional downtime planning, in which the human simply moves from an online to an off-line work mode. Within many teams, there are no humans doing the work of the agents who can pick up the work being automated.
As team leaders focus on planning and establishing lines of sight into AI implementations, we gain a much better sense of what we’re automating and why. Keeping a “human in the loop” helps manage risk — and also edge cases. An agent is only as good as what’s being said into it. When a new data pattern, scenario or question pops up, you need the right safeguards in place so the system understands each edge case.
This might sound intuitive. Unfortunately, vendors rarely include this in their pitch because they see it as a weak selling point. Questions around edge cases present a slippery slope amid a “panacea of automated tasks.” Meanwhile, buyers — hospitals, health systems, payers, etc. — aren’t asking questions about the limits of the agent. They don’t necessarily know to ask “what happens if this goes wrong?”
Established vendors — such as those who have been around for a while and whose claims have been verified via peer-reviewed studies — are the most qualified to answer these kinds of questions. Many newer vendors have little history to support their claims. Be wary of vendors that promise the world but lack data, experience, a plan for handling edge cases and the ability to anticipate malfunctions.
Agents are also going to need new approaches to monitoring and performance tuning than organizations have historically used for traditional applications. More than just user usage patterns, a framework must be in place to monitor the decisions and actions agents are making at a granular level to ensure they don’t go off course.
More holistic training
Both academic and general reporting has pointed to the need for top-to-bottom, broad-based AI training programs — not just instructions coming from the top. A holistic, companywide approach to AI training will be more successful than any team being handed a set of rules, then instructed to “get on board.”
When an entire organization is trained effectively, the networking effects are revealed: People from the top of the organization, to front-line employees who are well-trained on how to leverage the tools, start integrating the new AI into their everyday workflows. This is where organizations can really start unlocking value.
Quantifying the value/ROI of AI agents
Another important question to ask in 2026: How do you measure the value of an agent? Historically, CFOs have had a harder time assigning value to productivity enhancements (i.e., soft dollars), and accounting for the ROI of a technology that saves time. Those details ought to be addressed upfront, prior to any AI rollout.
In health care, the value proposition might mean using agents to help offset workflow issues caused by the ongoing labor shortage. Hospitals don’t have to worry as much about filling open positions if automating an unfilled job leads to productivity gains. The cost savings, in that case, might start with the salary of an unfilled job listing.
In any case, an accurate estimate of costs is needed to measure true ROI. Even when using an AI agent to fill a role you can’t hire, it’s not a zero-cost effort. There is a cost associated with running that agent that should be understood upfront. Here too, partnering with an experienced, proven vendor can better ensure success. An inexperienced client can lean into an experienced vendor’s expertise for the most accurate forecasting.
Forecasting costs
The next step for many organizations is developing appropriate forecasting models that account for the plusses and minuses of their AI agents to generate a net monetary cost.
Some large language models charge based on the number of tokens being processed. Some vendors charge a fixed fee (the “all-you-can-eat” model). There’s financial uncertainty either way. If you’re paying per utilization, uneven utilization of the agent will incur up-and-down expenses; if you’re paying a flat rate, the vendor is paying for that utilization, trying to figure out the right financial model so it can stay afloat.
In theory, the organizations that try to throw an agent at everything, from tasks small to large, should have the most robust forecasting models. Those who are starting small with a pilot program still need to make sure the agent adds high-value processes that justify the cost to run the agent. Cost-benefit attribution is key, and yet really difficult.
Finding balance
The US health care market is overly administrative-task driven. That won’t change in 2026. With so many moving parts, administrative overhead and bureaucracy that makes the health care system work, the industry will continue to be ripe for disruption and automation — with plenty of opportunities inside of payer and provider organizations to automate a lot of their workflows.
Organizations must balance that potential with safeguards. Risk management, change management, training and a greater emphasis on measuring ROI — both in terms of costs and benefits — are the necessary focal points for successful AI adoption in 2026.
_______________
Subscribe to SmartBrief for Health Care Leaders, one of our more than 30 health care publications.