The ROI of Healthcare Agents

Back

The ROI of Healthcare Agents

Learn how to measure healthcare AI agent ROI beyond labor savings, from improved workflow completion and patient access to reduced rework, operational costs, and staff workload.

Stochastic / Blog

Oct 29, 2026

The ROI of

Healthcare Agents

Oct 29, 2026

Healthcare organizations do not need another AI tool that looks impressive in a demo. They need systems that create measurable operational value. That distinction is becoming more important as AI adoption moves faster than financial returns.

McKinsey’s 2026 global survey found that 80% of respondents said AI had improved their individual productivity, but only 37% reported a positive contribution to enterprise EBIT. Just 6% qualified as AI high performers, meaning they attributed at least 5% of EBIT to AI and described its impact as significant. [1] Research at the task level shows why the opportunity is still real. In a study of 5,179 customer-support agents, researchers from Stanford and MIT found that access to a generative AI assistant increased issues resolved per hour by 14% on average while also improving customer sentiment. [2] AI can make individual work faster and better, but those improvements do not automatically become enterprise ROI.

Healthcare makes this problem especially visible because so much administrative work happens between systems. Staff read information from one document and enter it into another system, retrieve information from a payer portal and relay it to a patient, move referrals between queues, verify data before forwarding it downstream, and reconstruct context when cases move between teams. McKinsey describes this broader problem as a “coordination tax,” estimating that coordination, verification, reconciliation, waiting and handoffs consume 35% to 60% of total work time across knowledge-intensive organizations. [3] In healthcare specifically, CAQH estimates that moving administrative processes toward fully electronic workflows could unlock another $20 billion in annual savings and save an average of 70 minutes of administrative work per patient visit. [4]

The ROI of a healthcare agent therefore cannot be measured only by asking whether the model answered correctly or whether one employee completed a task faster. It has to be measured across the workflow the agent is changing, including the labor involved, the capacity consumed, the time added to the service and the work created when something goes wrong.

The Four Costs Hidden in Healthcare Administrative Work

Before calculating the return from an agent, an organization needs to understand what the existing workflow actually costs. Those costs generally appear in four forms: direct labor, coordination and capacity, service time, and rework. Looking at all four creates a much more useful baseline than simply comparing software cost with employee salaries.

1. Direct labor

Direct labor is the easiest cost to see. It is the time employees spend answering calls, reviewing documents, entering information, navigating portals, checking eligibility, following up on requests and completing other administrative tasks. A useful baseline starts with transaction volume, average handling time and the fully loaded cost of the people performing the work.

Prior authorization illustrates the scale of that burden. CMS estimates that requesting prior authorizations costs providers between $20 and $50 per hour and takes an average of 13 hours each week. Across a year, CMS estimates approximately 700 hours and $34,000 in administrative time for each provider. [5] This is the most obvious cost to measure, but it is only one part of what the workflow consumes.

2. Coordination and capacity

Administrative work also consumes the capacity of people who could be doing something else. A staff member who spends an hour reconciling information across systems cannot use that hour to resolve an exception, help a patient or work on another operational priority.

This is where the human-as-interface problem becomes important. Many healthcare workflows still depend on people acting as the glue between systems that do not naturally communicate. A person reads information from one place, interprets it, verifies it, and then enters or communicates it somewhere else. None of these actions may look especially expensive in isolation, but repeated across thousands of transactions they consume a large amount of organizational capacity.

Prior authorization again provides a useful example. The AMA reports that physicians and staff spend an average of 13 hours each week completing prior authorization work, while 40% of physicians report having staff who work exclusively on prior authorization. [6] Even if an organization does not reduce headcount, returning some of that capacity to the team has economic value because the same people can handle more work or focus on tasks where human judgment matters more.

3. Service time and patient friction

The third cost is the time experienced by the patient and by everyone waiting downstream. A scheduling request sitting in a queue delays access to an appointment. A referral waiting for review delays the next team that needs to act on it. A patient transferred between departments spends more time trying to accomplish something that may ultimately be simple. A prior authorization that takes longer than necessary can delay care even if the administrative labor required to process it remains unchanged.

The AMA reports that 95% of surveyed physicians say prior authorization delays access to necessary care, while 79% report that patients sometimes abandon treatment because of authorization challenges. [6] These are extreme examples of a broader point: the quality and speed of the administrative service have value too. Shorter call duration, faster resolution, fewer transfers, quicker document processing, clearer follow-up and greater availability can all improve the patient experience while reducing the amount of time spent moving a request through the organization.

4. Rework and exception handling

The fourth cost is the work created when the process does not succeed the first time. Missing information causes another call. An incorrect field requires another review. A document routed to the wrong place has to be retrieved and sent again. A handoff without context forces the next person to reconstruct what already happened.

This work is easy to miss because it is usually absorbed into normal operations rather than appearing as a separate budget line. Every repeated touch, however, increases the true cost of completing the workflow. Taken together, direct labor, coordination, service time and rework provide a more complete picture of what an organization is actually paying before an agent is introduced.

Why Turning Those Costs Into ROI Is Harder Than It Looks

Once the baseline is clear, it can be tempting to compare it directly with the price of an AI system. That usually produces an incomplete ROI calculation because the difficult part is not whether AI can perform an isolated task. The difficult part is whether the surrounding workflow actually changes.

The first challenge is partial automation. An assistant may extract information from a PDF correctly, but if a person still has to copy that information into another system, rename the file, determine where it goes and send it manually, much of the coordination cost remains. The same is true of a scheduling assistant that answers a patient’s question but cannot check the relevant system, complete the booking or preserve context when staff need to intervene. The model has performed useful work without necessarily removing the expensive handoffs around it.

The second challenge is integration. Integration matters because it determines whether technical capability can become operational return. The agent needs access to the systems, tools and actions required to move the workflow forward. That may mean querying an EHR, navigating a payer portal, updating a record, sending a document, creating a task for staff or triggering the next step through an API. If those actions remain manual, the organization continues to pay much of the coordination cost even if the AI performs the reasoning correctly.

The third challenge is exception handling. No production healthcare system should assume that every case will be automated end to end. Missing information, ambiguous requests, unusual insurance requirements, difficult documents and cases requiring judgment will continue to reach people. The relevant question is what the human receives when that happens. If staff have to restart the process, the agent has created limited value. If they receive the history, extracted information, completed work and a clearly defined exception, human review can still be dramatically cheaper than the original workflow.

Organizations also need to account for the total cost of operating the agent. Implementation, integration, model and infrastructure costs, monitoring, support, human review, security, compliance and ongoing maintenance all belong in the business case. A credible ROI model includes those costs rather than treating the technology as free once deployed.

The Four Returns Healthcare Agents Can Create

The returns should map directly to the costs identified at the beginning. If the workflow has four major sources of cost, a successful agent should create value by reducing those same four burdens: lowering the labor required to complete work, increasing capacity, improving the speed and quality of patient service, and reducing rework.

1. Lower labor per completed workflow

The most direct return is reducing the amount of staff time required to complete routine administrative work. The important unit is not simply minutes saved on one AI-assisted step. It is the amount of human time required for a successfully completed workflow.

If an agent collects information, checks the relevant systems, performs the required action and records the result, the reduction in labor is much greater than if it only helps a person draft a response. In revenue cycle management, for example, McKinsey estimates that AI enablement could reduce cost to collect by 30% to 60%, while allowing staff to focus more heavily on higher-value expertise and patient experience. [7] The exact return will differ by workflow and organization, but the principle is the same: measure how much human work remains after the agent has done its part.

2. More capacity and throughput from the same team

The second return is what the organization can do with the capacity it gets back. A person handles a limited number of cases at once, while an agent can work across many routine cases concurrently. That changes the economics even when an individual transaction still takes time to process.

In document intake, for example, an agent may need time to read, classify and organize a referral, but a staff member does not need to sit with the document while that happens. In scheduling, multiple patient conversations can be handled simultaneously instead of being constrained by the number of people available to answer the phone. The return may appear as avoided hiring, smaller backlogs, more transactions processed per employee or additional patient volume handled by the same team. Organizations should decide which outcome they expect before deployment so they can measure the one that actually occurs.

3. Faster and better patient service

The third return is improvement in the service itself. For patient-facing workflows, efficiency should not mean simply making the back office cheaper. A better system should also reduce the effort required from the patient through shorter waits, fewer transfers, fewer repeated questions, faster answers, quicker resolution and clearer follow-up when something cannot be completed immediately.

This is particularly visible in scheduling. In our work on new patient scheduling, Stochastic built the agent around the full interaction rather than only the conversation. A virtual concierge replaces rigid phone-tree navigation, the agent can continue working on tasks while maintaining the conversation, unresolved cases are placed into a structured human worklist, and the patient can be contacted again once the issue is resolved. [8] These improvements affect both operational efficiency and the experience of the patient using the service.

The financial value of patient experience should not be invented. An organization should not claim that every shorter call or more satisfied patient directly creates a specific amount of revenue unless it can measure that relationship. Service metrics such as abandonment, transfers, time to resolution, scheduling completion and follow-up completion should still be part of the ROI scorecard because they show whether automation improved the service rather than simply lowering its cost.

4. Less rework and higher completion

The fourth return comes from getting more workflows right the first time and making exceptions cheaper to resolve. A useful agent can validate information before an action is taken, preserve context across steps and make incomplete work visible instead of allowing it to disappear between systems. When human review is required, staff can begin from prepared work rather than reconstructing the entire case.

This is also why automation rate alone is a weak measure of ROI. A workflow that is 70% fully automated but requires the remaining 30% to be restarted manually may create less value than one where the same 30% reaches staff with almost all of the work already completed. Healthcare organizations should therefore measure both straight-through completion and assisted completion. The first shows how often the agent finishes without human involvement, while the second shows how much human work remains when it does not.

The Challenges in Measuring the Return

Even with clear costs and returns, several measurement mistakes can make an AI deployment look better or worse than it actually is. The first is double counting. If an agent saves 1,000 staff hours and those hours are then used to process additional patients, an organization should not automatically count the full labor value of those hours and the full value of the additional throughput as separate benefits. The same released capacity may be producing both numbers, so the ROI model should reflect how the organization actually used it.

The second mistake is measuring activity instead of outcomes. Calls handled, documents classified and tasks generated can be useful operational metrics, but they do not prove return on their own. Completed appointments, successful submissions, correction rates, backlog reduction, staff time per completed case and cost per completed workflow are much closer to the economic outcome the organization is trying to improve.

The third is ignoring quality and rework. Speed without reliability can create negative ROI. If an agent processes twice as many cases but generates enough errors that staff have to verify or redo its work, the apparent productivity gain disappears. The percentage of agent outputs accepted without correction, the types of corrections required and the amount of human review time therefore belong inside the ROI calculation rather than being treated as separate technical metrics.

The final mistake is measuring a deployment without a baseline. Before the agent goes live, the organization should understand current volume, handling time, touch count, queue time, completion rate, error or rework rate and relevant service metrics. Not every workflow will have perfect historical data, but without some baseline it becomes difficult to separate real improvement from a convincing story about improvement.

Building Toward a Practical ROI Equation

Once the workflow costs, returns and measurement rules are clear, the equation becomes much simpler. A useful annual model starts by calculating gross benefit as realized labor savings + value of additional throughput + avoided rework or outsourcing costs + quantified service-level benefits. The organization can then compare that with total cost of ownership, including implementation and integration + model and infrastructure costs + monitoring and support + human review + security, compliance and ongoing change management.

The final calculation is:

ROI = (Gross benefit - Total cost of ownership) / Total cost of ownership

The important part is not the arithmetic. It is what goes into each term. “Realized labor savings” should represent savings the organization can actually identify, such as avoided hiring, reduced overtime or lower outsourced work. If staff time is redeployed rather than removed, its value may be better represented through additional throughput instead.

“Additional throughput” should also be tied to a measurable outcome, such as more completed appointments, more referrals processed or faster cash collection, rather than a theoretical maximum. “Avoided rework” should come from observed changes in corrections, repeated touches, failed handoffs or external processing costs. Service-level benefits should only be monetized when the organization has a credible way to connect them to financial outcomes. Otherwise, measures such as wait time, abandonment and patient satisfaction should remain important operational outcomes without being forced into an artificial dollar value.

The goal is not to produce the largest possible ROI percentage. It is to produce a number that finance, operations and the people doing the work all recognize as credible.

The Longer-Term Return

There is one additional return that is harder to put into the first-year equation: having the operational foundation in place as AI improves. Models are continuing to become more capable and less expensive to run, while new tools are expanding the range of actions agents can perform. Organizations that already have agent workflows connected to their systems, clear human escalation paths, evaluation processes and operational ownership will be in a much stronger position to benefit from those improvements.

A better model can be introduced into an existing agent architecture, new tools can be added and individual steps can be automated more deeply without rebuilding the entire operational workflow from scratch. The same applies to employees. Working effectively with agents is an organizational capability, and teams need experience deciding when an agent can act autonomously, when its work should be reviewed, how exceptions should be handled and where human judgment remains essential.

This does not mean automating everything immediately. It means starting with high-volume workflows where the cost is visible, the outcome can be measured and the boundary between agent and human responsibility can be clearly defined. As the technology improves, both the technical infrastructure and the people using it are already prepared to take advantage of what comes next.

The ROI Question Healthcare Leaders Should Ask

The question is no longer simply whether an AI agent can perform a healthcare administrative task. Current systems can already handle meaningful portions of scheduling, document processing, revenue-cycle work and other operational workflows. The more useful question is whether deploying the agent changes the economics and quality of the workflow as a whole.

How much human labor remains per completed case? How much additional capacity does the organization gain? Does the patient reach a resolution faster and with less friction? How much rework disappears? What happens when the agent cannot proceed? What does the full system cost to operate? Which outcomes can actually be demonstrated against a real baseline?

Those questions move the conversation away from flashy demos and toward operational value. Healthcare agents create meaningful ROI when they reduce the cost of completing work while increasing the amount and quality of service an organization can deliver. Organizations that measure that full picture will be in the best position not only to justify the agents they deploy today, but to take advantage of what those agents become capable of tomorrow.