A new frontier of AI is approaching as artificial intelligence moves from assistants to agents. AI assistants—such as chatbots—can respond to questions and create content, but they rely on people to initiate and advance each step of a process. AI agents, on the other hand, can be created to handle entire complex workflows end to end.
The healthcare industry has been grappling with soaring costs and a labor shortage. AI agents can now be used to manage many of the complex workflows that often bog down staff. These intelligent agents can involve people when necessary to clear roadblocks and ensure oversight. This frees up valuable time to focus on what truly matters: For clinicians, that’s providing exceptional patient care, and for those at health insurance companies, that’s ensuring a seamless and cost-effective experience for their members.
How healthcare can put AI agents to work
AI agents are akin to a workforce built on code. They’re powered by generative AI, coupling predictive and creative capabilities with reasoning to perform complicated workflows. Think of agents as AI virtual workers that can work independently once they’re given a specific goal, details on what tasks to perform and how, additional contexts to consider, guardrails to work within, and existing online tools to help them implement tasks.
What’s not to like? One common fear is that agents can make healthcare decisions without any human oversight, which would lead to substantial risk. To be clear: Agentic AI allows a spectrum of autonomy. In high-stakes contexts such as healthcare, a strategically placed human in the loop can be a critical safeguard. The good news is that with appropriate governance and oversight, agents can be used across the full gamut of front-office to back-office activities.
Both clinical-care organizations and health insurance companies can benefit from agentic AI across all phases of a patient encounter. For example, agents could work with patients to identify appropriate sites of care for nonemergency needs and to help them prepare for upcoming appointments. During care, agents could coordinate case management tasks within and across sites of care. Agents could also streamline the discharge experience by offering patients tailored information about their appointment and any follow-up required. Additionally, agents can accelerate billing and reimbursement by submitting claims, working with physicians to fix documentation gaps, and drafting appeals.
What’s more, multiagent systems can be set up to handle various tasks and to coordinate between care organizations and insurers. While some administrative processes in the healthcare ecosystem are quite complex, an agentic system can help streamline how work is done and relieve some of the burden on healthcare workers. For example, a multiagent system can optimize provider networks and contracting by flagging key considerations for human review during negotiation processes. It could also collaborate with call center representatives at providers and payers by retrieving answers and completing tasks like scheduling and sending documents.
Another area that a multiagent system can be set up for is the claims process. Within this process, claims that payers consider to be high cost can trigger underpayments to the providers. As a result, handling these claims can often be long and laborious and requires coordination between clinical-care organizations and health insurers. A fleet of agents can help both the provider and payer sides. Starting at the care-delivery organization, after a patient appointment, the provider will need to generate an itemized bill and submit a claim to the insurer. Agents can verify insurance details, figure out the right associated codes for reimbursement, review policy compliance, and compile a claim for a clinician to review for accuracy before submission. On the health insurer side, an agent can do automated checks to verify coding accuracy and contract terms. Another agent can retrieve required documents and synthesize the provider’s submission to support the reviewer checking for clinical and medical necessity. Another agent can calculate the provider’s payment. Finally, an agent can create the 835 form that’s sent to providers and used to create the Explanation of Benefits letter that’s sent to members. Once the provider receives the 835 form, a set of agents can identify underpayments or discrepancies and generate an appeal letter when needed for a final reviewer to check before sending to the payer.
How to get started with implementing agentic AI
Implementing agentic AI can power a fundamental rethink of how work is done. However, organizations tend to dip their toes into just a few applications at a time that span different aspects of the business. While this can lead to some limited benefits, the true benefit of agentic AI comes from the flywheel effect of scaling use cases. Achieving this expansion can, however, be tough: Resources are spread thin, there’s an inability to build on the right reusable foundations, and tracking the value and impact is difficult. We suggest six key considerations to ensure agentic AI succeeds.1
Ascertain if agentic AI is the optimal solution. For example, for a well-defined problem that can be solved with a one-time piece of code, traditional software is the way to go (such as to reset customers’ passwords). Traditional AI can also solve many problems with explicit user instructions (for instance, to forecast time in surgery to assess operating room capacity). Agentic AI would be best suited to implement end-to-end workflows or to solve more complex problems—that is, processes that are marked by continuously evolving inputs and that require using different tools dynamically, connecting with different sources, gathering information, and using the information to communicate, reason, and solve problems (for example, to streamline how patients transition among various care settings).
Get the strategic fundamentals right. Instead of investing in dozens of agents without regard to priorities, organizations should prioritize applications that have the biggest business impact. In our experience, many organizations experimenting with gen AI find themselves in pilot purgatory even one to one and a half years after kicking off the initial tests. With agentic AI, it is important to be a “focused transformer,” picking a few domains to start with.
Make the correct architectural choices to suit the required complexity. Simpler individual agents can be used for straightforward tasks, like directing requests to the right resources. For complex workflows, multiagent systems are ideal, since agents are assigned different roles and cooperate to complete a task sequentially or in tandem. Some agent roles include the following:
- Orchestration agents act as supervisors; they direct task agents and involve other agents as needed.
- Task agents perform one specific task and provide outputs to other agents and/or human reviewers.
- Review agents check other agents’ outputs to ensure accuracy and flag discrepancies.
- Planning agents anticipate future scenarios, instead of being reactive, and create a plan to achieve a goal.
Manage risks and ensure governance. The potential autonomy of AI agents could be a catalyst to quickly implement reimagined workflows. But with increased autonomy comes the need for increased governance to build trust in an agentic system. Given the unique context and risks inherent to healthcare, keeping a human in the loop will always be core. With agents automating portions or entire workflows, setting the right checkpoints in the workflow for a person to validate is critical. AI agent tools, functions that extend an agent’s capabilities to do specific tasks, are also important because they prevent agents from doing anything they should not, such as deleting records from a database.
Invest in change management and talent development to support process redesign. The day-to-day lives of those who work at healthcare organizations that employ agentic workflows will be altered, making change management vital. In fact, agentic AI can eliminate components of the workflow that have high manual labor and require repetitive steps. Organizations will thus need to both hire people with specialized expertise and upskill existing employees. While it’s important to acknowledge workforce fears about job security, it is more critical to shift mindsets and, where necessary, encourage talent to embrace agentic AI and transition employees’ operating models so that it is easier for them to understand how AI can augment and enhance what they do.
Consider the big picture on build-versus-buy decisions. Organizations will need to evaluate whether to invest in building their own agentic systems or partner with the many start-ups offering solutions. The enterprise architect’s role will increase in importance, as they have to evaluate the enterprise tech stack and keep tabs on where, how, and why agentic AI is being considered. Another key factor for care organizations, health insurers, and tech vendors alike will be to decide whether to build agentic systems as “open” or “closed” technical systems (for long-standing legacy claims systems or existing call center software, for example, or to collaborate flexibly across multiple technologies and vendors).
AI agents can incorporate new information, make decisions, carry out tasks, and coordinate with other AI agents and people as needed. In an industry facing skyrocketing costs and a shortage of workers, agentic AI has the power to tackle time-consuming administrative work and coordination among patients, clinicians and clinical staff, and health insurers. Healthcare organizations stand to benefit from being at the forefront of adopting agentic AI to enhance efficiencies, create new value, and improve how patients experience the healthcare ecosystem.
Carlos Pardo Martin and Jessica Lamb are both partners in McKinsey’s New York office. Amine Dahab is a consultant in the Boston office, and Saumya Singh is an associate partner in the Bay Area office.
The authors wish to thank Eduardo Coronado, Jack Eastburn, John R. Jones, Mark Daggett, Sameer Chowdhary, and Sanjiv Baxi for their contributions to the article.
This article was edited by Querida Anderson, a senior editor in the New York office.
1. Eric Lamarre, Kate Smaje, and Rodney Zemmel, “Rewired to outcompete,” McKinsey Quarterly, June 20, 2023.