AI Agents Are Coming to Healthcare. Most Medical Practices Are Not Ready.
Why the next competitive advantage won’t be another EHR, another chatbot, or another employee—it will be an intelligent workforce working quietly behind the scenes.
Monday morning begins before the doors open.
Voicemails have accumulated over the weekend. Portal messages are waiting. Patients want to schedule, reschedule, confirm instructions, request refills, locate referrals, ask about forms, and understand what happens next.
One patient says nobody returned the last call.
Another referral arrived without the required documentation.
A cancellation has opened valuable appointment capacity, but the patients on the waitlist have not yet been contacted.
Someone needs an insurance update. Someone else needs an interpreter. A staff member has called in sick.
Nothing catastrophic has happened.
No single employee caused the problem.
Yet the practice is already falling behind.
This is happening inside an organization that may own an EHR, a practice-management platform, a phone system, online scheduling, automated reminders, digital forms, a patient portal, a texting application, a payment system, and several additional technologies.
The information is digital.
The work is still waiting for a person to notice it.
That is the operational reality confronting many medical practices. They do not necessarily have a shortage of software.
They have a shortage of coordination.
Artificial intelligence is about to make that distinction impossible to ignore.
Healthcare Digitized the Record—Not the Work
Healthcare has completed an extraordinary digital transformation.
As of 2024, 91% of office-based physicians had adopted a certified electronic health record. More than 99% of nonfederal acute-care hospitals had done the same.
That achievement matters. EHRs have become essential systems of clinical record.
But digitizing medical information did not automatically connect all the work surrounding it.
The patient journey still crosses:
- Telephones and voicemail
- Patient portals
- Text messages
- Online scheduling
- Digital forms
- Referral systems
- Fax
- Payer portals
- CRM applications
- Payment systems
- Staff work queues
- Spreadsheets
- Human memory
A patient may enter through one system, be documented in another, receive messages from a third, submit information through a fourth, and require follow-up that depends on someone remembering to check a separate queue.
Each application may perform its assigned function.
The overall journey can still fail.
An EHR may document that a referral exists without ensuring the referral reaches completion. A phone platform may record an abandoned call without recovering the patient. A scheduling application may expose available appointments without identifying why other capacity remains unused. A portal may receive a message without ensuring that the request is routed, resolved, documented, and closed.
Healthcare digitized the record.
It did not fully digitize the work between the records.
That gap is where unanswered inquiries, incomplete referrals, delayed responses, unused capacity, employee frustration, and preventable patient attrition accumulate.
It is also where AI agents could eventually create their greatest value.
The World Is Moving From AI Assistance to AI Execution
The first major wave of generative AI helped people create, summarize, search, organize, and draft information.
The next wave is intended to help systems pursue defined objectives.
This is the difference between an AI assistant and an AI agent.
An automation follows a predetermined rule:
When this happens, do that.
An AI assistant helps a person produce or interpret information.
An AI agent works toward a defined goal. It can interpret context, use approved tools, take permitted actions, evaluate the result, and escalate the work when human involvement is required.
A multi-agent system coordinates several specialized agents across a larger process.
The distinction matters.
A chatbot may answer a question.
An assistant may draft the response.
An agent may identify what the patient needs, locate approved information, take the permitted next step, update the appropriate system, document the outcome, and alert a person when the request exceeds its authority.
A chatbot answers. An assistant helps. An agent acts. An intelligent practice coordinates many agents under human control.
Microsoft’s 2025 Work Trend Index describes the emergence of organizations built around hybrid teams of people and AI agents. The research incorporated a survey of 31,000 workers across 31 markets, LinkedIn labor-market information, Microsoft 365 usage signals, and interviews with experts and AI-native organizations.
Its 2026 research continued exploring how agents may reshape organizations while emphasizing that human anxiety, judgment, agency, and leadership remain central to the transition.
Healthcare will not adopt this model exactly as retail, banking, logistics, or travel will. The risks, regulations, workflows, and human consequences are different.
But the underlying shift is the same:
Software is beginning to move from displaying information to participating in work.
Healthcare Is Adopting AI—but Not Yet as an Intelligent Workforce
Healthcare is not ignoring artificial intelligence.
The American Medical Association’s 2026 survey of 1,692 physicians found that more than 80% were using AI professionally—approximately twice the level reported in 2023. The average physician reported 2.3 AI use cases, compared with 1.1 in 2023. More than three-quarters believed AI could improve their ability to care for patients.
The most frequently reported applications included:
- Summarizing medical research and standards of care
- Creating discharge instructions, care plans, or progress notes
- Supporting billing-code and clinical documentation
- Generating chart summaries
- Drafting patient-portal responses
- Translation
- Diagnostic assistance
Medical groups are also expanding AI operationally. In a September 2025 MGMA poll, 68% of responding groups reported adding or expanding AI during the year. Common targets included documentation, scheduling, patient communication, revenue-cycle functions, coding, prior authorization, and analytics.
AI has even overtaken EHR usability as a leading technology priority among medical-practice leaders surveyed by MGMA.
This is meaningful progress.
It is not yet an intelligent workforce.
Most present-day healthcare AI helps individuals or improves isolated tasks. A physician receives a draft note. A patient receives a reminder. A coder receives a suggested code. A staff member receives a summary. A caller interacts with a virtual assistant.
These applications can be valuable.
But they do not necessarily create coordinated operational capability.
Healthcare is not behind in trying AI.
It is behind in connecting isolated AI tools to accountable, end-to-end work.
The Hidden Patient-Engagement Gap
Patients do not experience a medical practice as a collection of departments and technologies.
They experience one organization.
They expect the practice to know that they called, submitted a form, requested an appointment, received a referral, canceled a visit, asked a question, or needed help.
Internally, however, those interactions may arrive through unrelated channels.
That creates what we call the Patient Engagement Gap.
What Patients Expect
- A timely response
- Clear instructions
- Convenient communication
- Continuity between interactions
- Language and accessibility support
- A human being when the situation requires one
What Many Practices Can Consistently Provide
- Limited telephone capacity
- Fragmented communication channels
- Delayed callbacks
- Generic automation
- Inconsistent follow-up
- Separate departmental queues
- Limited visibility into unfinished work
What Accumulates in the Gap
- Patient frustration
- Staff pressure
- Incomplete referrals
- Unfilled cancellations
- Weak reactivation
- Poor schedule utilization
- Repeated work
- Reputation risk
- Lost revenue opportunities
- Operational blindness
Not every failure is visible immediately.
A patient may simply call another practice.
A referral may remain unfinished.
A cancellation may remain open.
An overdue patient may never be contacted.
A complaint may appear later as a negative review.
A staff member may spend ten minutes reconstructing a journey that should have been visible in seconds.
Every unanswered request creates operational debt. Someone must eventually resolve it—or the patient, appointment, referral, or revenue opportunity disappears.
AI agents could help close this gap, but only if practices first pass six critical readiness gates.
The Six Gates Blocking Agentic Healthcare
Gate 1: Integration
An agent cannot complete work reliably if it cannot interact appropriately with the systems where the work lives.
MGMA has identified EHR and practice-management integration as central to obtaining value from healthcare chatbots and virtual assistants.
Consider an appointment inquiry.
A conversational system may understand that the patient wants to schedule. But can it identify the correct appointment type? Apply provider and location rules? Recognize referral requirements? Access approved availability? Account for visit duration? Update the scheduling system? Send the correct instructions? Document the interaction? Escalate an exception?
Without integration, the system may conduct a pleasant conversation and still leave the work for a person.
Conversation without integration is not transformation. It is another inbox.
Integration must also extend beyond the EHR.
Patient engagement crosses telephony, scheduling, CRM, forms, messaging, payments, referrals, analytics, and staff workflows. Connecting only one system may automate one step while leaving the rest of the journey fragmented.
Gate 2: Governance
Healthcare AI must operate within explicit boundaries.
The HIPAA Security Rule requires covered entities and business associates to protect electronic protected health information using appropriate administrative, physical, and technical safeguards. HHS describes risk analysis as foundational to determining which safeguards are necessary, and it provides a risk-assessment tool designed specifically to assist small and midsize healthcare organizations.
AI adds another layer of questions:
- What information may the agent access?
- What actions may it take?
- What decisions must remain human?
- How is identity verified?
- When must the interaction be escalated?
- How are actions logged?
- How are errors detected and investigated?
- Who reviews performance?
- Who remains accountable?
NIST’s AI Risk Management Framework is designed to help organizations incorporate trustworthiness into the design, deployment, evaluation, and use of AI systems. Its generative-AI profile identifies additional risks that organizations should consider when adopting these technologies.
MGMA’s guidance for medical groups similarly emphasizes policies, accountability, oversight, and clearly defined human responsibility.
A signed business associate agreement may be an important contractual component when protected health information is involved.
It is not, by itself, an AI-governance strategy.
A responsible deployment requires permissions, controls, testing, monitoring, documentation, escalation, risk management, and named human accountability.
Gate 3: Implementation Capacity
Buying technology and successfully changing a medical practice are not the same activity.
Practices have already lived through difficult EHR conversions, imperfect interfaces, portal launches, telephone migrations, and applications that promised efficiency but created additional work.
That history matters.
Employees do not evaluate a new AI system in a vacuum. They evaluate it against every previous implementation that disrupted their work, required duplicate entry, failed to integrate, confused patients, or left them responsible for cleaning up the result.
Smaller practices may be able to make decisions quickly, but they often have limited internal resources for:
- Information technology
- Compliance
- Data governance
- Workflow design
- Training
- Analytics
- Change management
- Continuous monitoring
This creates a paradox:
The practices that may benefit most from automation often have the least disruption capacity available to implement it.
The federal Health IT Playbook treats implementation and optimization as distinct disciplines requiring planning, stakeholder involvement, workflow evaluation, training, and ongoing improvement.
AI-agent adoption must be approached with the same seriousness.
A technology demonstration may take minutes.
Operational trust takes longer.
Gate 4: Economics
An AI system is not valuable merely because it can speak, write, or complete a demonstration.
Practice leaders need evidence that it improves the operation.
MGMA has found that practices reporting workload reductions often connect those gains to specific applications, while organizations seeing fewer benefits cite early-stage adoption or additional complexity.
The wrong measures include:
- Messages generated
- Calls answered
- Conversations conducted
- Features activated
- Agents deployed
Those may be useful activity measures, but they do not establish value.
Better measures include:
- Response time
- Resolution rate
- Appointment conversion
- Referral completion
- Recovered cancellations
- Reactivated patients
- Schedule utilization
- Staff handling time
- Escalation rate
- Error rate
- Patient satisfaction
- Capacity recovered
- Revenue protected or recovered
The value of an AI agent is not how much it talks. It is how much accountable work it completes.
A system that handles thousands of interactions while creating duplicate work, inappropriate responses, unresolved exceptions, or poor patient experiences may produce activity without producing value.
The practice must measure completion—not conversation.
Gate 5: Workforce Trust
No responsible discussion of AI can ignore employee anxiety.
People hear the language of automation and reasonably wonder whether leadership is trying to eliminate their jobs.
Practice administrators may carry an additional concern: even when they did not select the technology, they may become responsible for its failures.
Administrators are already accountable for:
- Staffing
- Patient access
- Vendor performance
- Complaints
- Compliance
- Training
- Workflow reliability
- Financial performance
- Operational breakdowns
An AI deployment can therefore look less like assistance and more like one more system to supervise.
This is why implementation language matters.
The objective should not be to portray employees as inefficient obstacles standing in the way of automation.
The objective should be to remove repetitive, low-value work that prevents capable people from performing at their highest level.
Staff should spend less time:
- Copying information between systems
- Repeating routine instructions
- Conducting basic status checks
- Searching for missing context
- Chasing predictable documentation
- Manually detecting forgotten follow-ups
They should have more time for:
- Exceptions
- Judgment
- Empathy
- Complex coordination
- Patient reassurance
- Service recovery
- Relationship building
The administrator of the future does not become less important.
The role becomes more strategic.
Practice administrators will increasingly conduct a hybrid workforce composed of people, systems, automations, agents, analytics, and external partners.
AI should not remove people from patient care. It should remove the friction that prevents people from caring well.
Gate 6: Patient Trust
Patients do not need to become enthusiastic about artificial intelligence for administrative AI to be useful.
They do need to trust the experience.
A national Pew Research Center survey found that 60% of U.S. adults would be uncomfortable if their healthcare provider relied on AI for functions such as diagnosis and treatment recommendations.
That finding should be taken seriously.
It should not be misrepresented as proof that patients oppose every use of AI in healthcare.
There is an important difference between:
- An AI system making a treatment decision
- An agent helping a patient locate an appointment
- An agent providing approved logistical instructions
- An agent routing a refill request
- An agent identifying an unfinished referral
- An agent offering a human callback
- An agent helping a patient communicate in a preferred language
Trust depends heavily on the nature and risk of the task.
Patient portals offer a useful lesson. In 2024, 65% of individuals nationally were offered and accessed an online medical record or patient portal. Yet 59% had multiple portals, and only 7% reported using an application that consolidated information from them. Patients encouraged by their healthcare provider to use portals accessed them at higher rates than those who were not encouraged.
Technology adoption is not produced merely by making a tool available.
It depends on design, guidance, convenience, relevance, and trust.
A patient-facing agent should therefore be:
- Easy to understand
- Transparent where appropriate
- Multichannel
- Accessible
- Sensitive to language and health literacy
- Conservative about clinical boundaries
- Able to reach a human
- Designed without dead ends
Patients do not need to love AI. They need to trust the experience it helps the practice deliver.
Why Another Chatbot Will Not Solve the Problem
The patient journey is larger than a conversation.
A patient may:
- Discover the practice.
- Submit an inquiry.
- Call for more information.
- Need a referral.
- Select a provider.
- Schedule an appointment.
- Complete forms.
- Receive instructions.
- Reschedule.
- Attend the visit.
- Need follow-up.
- Ask a question.
- Become overdue for care.
- Leave feedback.
- Refer another patient.
A standalone chatbot may participate in one or two of these moments.
The practice remains responsible for the entire journey.
This is why adding isolated AI features to disconnected applications may actually worsen fragmentation. Each vendor may provide an assistant, bot, copilot, or agent that operates inside its own product.
The practice can end up with more intelligence inside individual silos while the silos themselves remain disconnected.
One agent knows about the phone call.
Another knows about the appointment.
Another knows about the marketing inquiry.
Another knows about the referral.
Another knows about the patient’s message.
No system understands the whole operational journey.
The future of patient engagement will not be one bot talking to every patient. It will be a coordinated system ensuring that every appropriate next step happens.
Medical practices do not need dozens of unrelated AI features.
They need an intelligent operating layer connecting:
- People
- EHR and practice-management systems
- Telephony
- CRM
- Messaging
- Scheduling
- Forms
- Referrals
- Payments
- Analytics
- Patient feedback
- Operational workflows
This is the difference between purchasing more software and building an intelligent medical practice.
The Invisible Workforce
The medical practice of the future may be supported by dozens of narrowly defined capabilities.
Patients should not have to see or understand them.
We call this the Invisible Workforce: a governed network of specialized agents and automations working quietly behind the people delivering care.
It may include five broad families.
1. Access Agents
These agents support approved administrative inquiries, appointment requests, basic navigation, after-hours interactions, routing, and human handoffs.
Their objective is not to prevent patients from speaking with people.
It is to prevent patients from being abandoned while waiting for people.
2. Continuity Agents
These agents support the work that happens between visits:
- Referral follow-up
- Form completion
- Pre-visit preparation
- Cancellation recovery
- Waitlist management
- Follow-up scheduling
- Patient reactivation
- Overdue-care outreach
Their purpose is to reduce the number of patients who disappear between one appropriate next step and the next.
3. Communication Agents
These agents help coordinate voice, text, portal messages, email, multilingual navigation, and escalation.
The patient should not need to understand which department owns a request.
The system should help route the request to the right workflow.
4. Operational Agents
These agents can assist with rules-based administrative work such as eligibility workflows, authorization preparation, status checks, task routing, revenue-cycle queues, documentation support, and inventory processes.
Human review remains essential wherever judgment, ambiguity, regulation, or clinical significance requires it.
5. Intelligence Agents
These agents analyze the operation itself.
They help leaders understand:
- Where inquiries are being lost
- Why patients cancel
- Which referrals remain incomplete
- Which channels perform best
- Which patients are becoming disengaged
- Where work is accumulating
- Which workflows consume excessive staff time
- Whether automation is producing measurable improvement
The patient does not need to see this infrastructure.
The patient should experience:
- A faster answer
- Less repetition
- Clearer communication
- Fewer forgotten next steps
- Better continuity
- A human being when one is needed
The best AI may be the AI patients barely notice. They will not remember the agent. They will remember that the practice answered, followed through, and cared.
What Changes for the People Running the Practice?
For Physician Owners
AI agents represent more than potential labor efficiency.
They represent capacity, consistency, visibility, risk, and enterprise value.
Owners should be able to ask:
- Where are new-patient inquiries being lost?
- Which referrals reach completion?
- Why do patients cancel?
- Which locations respond most effectively?
- How much appointment capacity is recovered?
- Which patients are becoming disengaged?
- Are automated workflows reducing work or merely relocating it?
- Where does human intervention create the greatest value?
A practice cannot improve what it cannot see.
For Chief Operating Officers
The opportunity is to establish common operating standards without erasing the differences among specialties, locations, physicians, and patient populations.
The objective is not identical care everywhere.
It is dependable execution everywhere.
For Practice Administrators
The role evolves from manually detecting failures to managing performance, exceptions, capacity, escalation, service recovery, governance, and continuous improvement.
The administrator becomes the conductor of the intelligent practice.
For Private-Equity Operating Partners
Agentic systems may eventually help standardize operations, improve multi-location visibility, recover capacity, and support acquired practices operating across different EHRs, phone platforms, and local workflows.
But value will depend on repeatable deployment, defensible governance, trustworthy reporting, and measurable outcomes—not isolated demonstrations.
For Employees
The promise should not be fewer people.
It should be less wasted human potential.
For Patients
The benefit is not access to artificial intelligence.
It is better access to the practice.
The Safest Way to Begin
A practice does not need to automate everything.
It should begin where work is already being delayed, repeated, forgotten, or lost.
A disciplined starting model is:
One Workflow. Ninety Days. Five Measures.
Choose One Contained Workflow
Strong starting points may include:
- Unanswered new-patient inquiries
- Referral follow-up
- Cancellation recovery
- Waitlist outreach
- Patient reactivation
- Pre-visit form completion
- Appointment-request routing
These workflows are generally easier to define and measure than broad, autonomous patient engagement.
Establish Clear Boundaries
Before deployment, specify:
- What the agent may do
- What the agent may not do
- What data it may access
- Which language and content are approved
- When a human must intervene
- How identity will be verified
- Who will monitor the system
- What happens during downtime
- How incidents will be documented
Measure the Baseline
The practice must understand current performance before claiming improvement.
Track Five Outcomes
- Response time
- Completion or conversion rate
- Human handling time
- Escalation and error rate
- Capacity, access, or revenue recovered
Expand Only After Evidence
Scale should follow proof of:
- Reliability
- Safety
- Integration
- Staff acceptance
- Patient acceptance
- Operational value
- Financial value
Do not begin by automating the entire practice. Begin where patients are already waiting and work is already being lost.
The Practice That Always Follows Through
The competitive medical practice of the future will not necessarily employ fewer people.
It will waste less of their time.
Staff members will not need to remember every follow-up manually.
Patients will not need to chase the office repeatedly.
Administrators will not discover every operational failure through a complaint.
Owners will not have to guess where demand, capacity, and revenue are disappearing.
Technology will become less visible even as the practice becomes more intelligent.
The winners will not be the practices with the greatest number of AI products.
They will be the practices that combine:
- Exceptional people
- Connected systems
- Clear governance
- Measurable workflows
- Responsible agents
- Human accountability
AI should not become the center of the patient experience.
People should.
The purpose of the technology is to quietly remove the fragmentation, repetition, delay, and uncertainty that prevent people from delivering exceptional care.
The future of patient engagement is not a patient trapped in a conversation with a robot. It is a patient receiving the right response, through the right channel, in the right language, at the right time—while the practice quietly ensures that the work behind the interaction is completed.
Frequently Asked Questions
What are AI agents for medical practices?
AI agents are software systems designed to pursue defined goals, use approved information and tools, take permitted actions, and escalate work when human involvement is required. In medical practices, they may support scheduling, referrals, patient communication, reactivation, administrative workflows, and operational analysis.
How are AI agents different from healthcare chatbots?
A chatbot generally answers questions within a conversation. An AI agent may take additional permitted actions, such as locating approved appointment availability, routing a request, updating a workflow, documenting an outcome, or creating a human task. The difference is not how naturally it talks, but whether it can complete accountable work.
Can AI agents integrate with an EHR or practice-management system?
Some can, but the depth and safety of integration vary significantly. Practices should examine what information the agent can access, which actions it can perform, whether the integration is real-time, how permissions are controlled, how activity is logged, and what happens when the connection fails.
Are AI agents HIPAA compliant?
HIPAA compliance cannot be determined from an “AI” or “HIPAA-compliant” label alone. Compliance depends on the complete deployment, including the parties involved, permitted uses, contracts, safeguards, access controls, risk analysis, data handling, monitoring, and organizational policies. A business associate agreement may be necessary, but it is not sufficient by itself.
Will AI agents replace medical receptionists or practice staff?
Some repetitive tasks may become automated, but medical practices will continue to require people for empathy, judgment, complex coordination, exception handling, reassurance, service recovery, and accountability. The stronger model is a hybrid workforce in which AI handles appropriate repetitive work and people focus on work requiring human ability.
What patient-engagement tasks can AI agents handle?
Potential administrative uses include appointment-request routing, approved frequently asked questions, referral follow-up, cancellation recovery, waitlist outreach, forms, pre-visit preparation, follow-up scheduling, patient reactivation, multilingual navigation, and service-recovery escalation. Clinical boundaries and human-review requirements should be explicitly defined.
Can AI agents communicate with patients in different languages?
AI can assist with multilingual administrative communication, but practices must evaluate accuracy, context, health literacy, accessibility, and the risk of misunderstanding. High-risk, ambiguous, or clinical communication may require qualified human interpretation or translation rather than unrestricted automated output.
How can a medical practice measure the ROI of AI agents?
Practices should measure operational outcomes rather than conversation volume. Useful measures include response time, appointment conversion, referral completion, recovered cancellations, patient reactivation, staff handling time, escalation rate, error rate, schedule utilization, patient satisfaction, and capacity or revenue recovered.
What is the safest workflow to automate first?
The safest starting point is usually a contained, high-volume administrative workflow with clear rules, measurable outcomes, limited clinical risk, and reliable human escalation. Examples include unanswered appointment inquiries, waitlist outreach, referral-status follow-up, form completion, or cancellation recovery.
How should practice administrators prepare for agentic AI?
Administrators should inventory workflows, identify fragmentation, establish baseline performance, define permissions and escalation rules, involve employees, evaluate integrations, create governance policies, and begin with a limited pilot. Their role will increasingly include supervising a hybrid workforce of people, systems, automations, and AI agents.
Conclusion
The next competitive advantage in healthcare will not be another isolated application.
It will be the ability to coordinate people, systems, automations, and AI agents around accountable work.
Medical practices that begin with integration, governance, measurable workflows, and human trust will be better positioned to create an intelligent workforce without losing the compassion, judgment, and accountability that healthcare requires.
Not more AI. A more intelligent way to operate medicine.