Conversational AI in Healthcare: A Practical Guide for Health Systems

Conversational AI in healthcare refers to artificial intelligence systems that communicate with patients, providers, and staff through natural language, using text or voice, to handle tasks ranging from patient triage to appointment scheduling. These conversational AI tools combine natural language processing (NLP) and machine learning to understand what someone is asking, interpret the intent behind it, and respond in a way that feels like human conversation.

The technology is no longer theoretical. Healthcare organizations are deploying conversational AI systems to manage patient portal messages, which has become a problem that has grown too large for manual workflows. This “used to be a ratio of 1 message to 1 patient visit, but now some studies say it’s 6 to 1,” says Don Woodlock, President of InterSystems.

Whether you’re evaluating conversational AI solutions for your health system or trying to understand where this technology delivers measurable results, what follows is a breakdown of what’s working today which is grounded in real deployments and published research.

At its core, conversational artificial intelligence is the practice of using natural language processing to parse human language either written or spoken and machine learning to improve its responses over time. Unlike rule-based chatbots that follow rigid scripts, modern conversational AI systems understand context, recognize intent, and generate responses that adapt to the specifics of each interaction.

In healthcare, the stakes are higher than in other industries. Conversational AI tools must handle clinical terminology, detect urgency in patient messages, comply with privacy regulations, and integrate with electronic health records. Virtual assistants and AI chatbots in this space do not replace healthcare professionals. They handle the routine communication that consumes clinicians’ time, such as answering frequently asked questions, routing messages and drafting responses, so that clinical staff can focus on direct patient care.

The distinction matters: conversational AI is not synonymous with generative AI. Some systems use traditional NLP/classification, while newer systems may use generative AI for drafting, summarization, or dialogue

Patient portal messaging has surged. A 2025 MGMA poll found that 70% of medical groups reported an increase in patient portal message volume in 2024. Multiple studies confirm that the post-COVID spike in digital patient engagement has become permanent and patients now expect to communicate with their healthcare providers through portals as easily as texting.

The problem is not just volume. It is what is buried inside that volume.

“4 to 10% of messages are urgent... the patient really perhaps shouldn’t have put that in the portal, but did anyway, and it’s just sitting there,” Don explains. Most healthcare systems commit to a 24- to 48-hour turnaround time on portal messages. That means an urgent message sent at 11 p.m. on a Friday could sit unread until Monday morning.

Compounding the issue, patient self-categorization, where patients select the topic or department for their message, is only accurate about half the time. That is not reliable enough to drive clinical workflow. Messages land in the wrong queue, sit for hours, then get forwarded to the right team, adding delays to a system already under pressure.

This is the gap conversational AI was built to close. For health systems that are managing portal messages across multiple electronic health records, platforms like InterSystems HealthShare aggregate patient data and messages from disparate sources into a unified view which creates the foundation that conversational AI needs to triage, route, and respond effectively.

Three applications have emerged as the most proven uses of conversational AI in patient portal messaging, all with high accuracy and lower risk than other AI applications in healthcare.

The most straightforward application: training AI models to flag urgent patient messages automatically. The system scans incoming messages, identifies language patterns associated with high-acuity situations such as chest pain, difficulty breathing and medication reactions, and tags them for immediate review.

In practice, this means a contact center specialist arriving at 7 a.m. can sort their queue by urgency, work the critical cases first, and maintain pace through the rest of the day. Most urgent messages arrive overnight or on weekends, so this workflow catches the cases most likely to slip through manual review.

NYU Langone Health developed and deployed an urgency detection model using BERT (a natural language processing architecture) trained on over 40,000 patient messages. The system identifies language that warrants immediate action, such as calling the patient, and has been in production use for several years. In validation, the model achieved a 97% C-statistic and 72% average precision; in operational binary use, precision was 67% and sensitivity was 63%.

The second application: AI proposes a draft response to a patient message, and a healthcare professional reviews, edits, and sends it.

The surprising finding here is about empathy. A 2023 study published in JAMA Internal Medicine compared physician responses to patient questions with chatbot-generated responses. Evaluators rated the chatbot responses as higher quality and significantly more empathetic, they were 9.8 times more likely to be rated “empathetic” or “very empathetic” than physician responses alone.

“Not because computers are more empathetic than humans, but empathy takes time,” Don says. Writing a message that acknowledges the patient’s concern, expresses genuine care, and still delivers clinical information takes minutes a physician doesn’t have when managing dozens of messages. AI has those minutes.

Don describes the most effective format as an “empathy sandwich”:

• Opening: Acknowledge the patient, express care - “Dear Mrs. Jones, thank you for contacting us, your care is important to us…”
• Middle: The clinical substance - the human writes or heavily edits this part
• Closing: Reassurance and next steps - “We hope that answers your question, feel free to contact us again, here’s our phone number”

The AI drafts the empathetic bookends. The clinician focuses on what they do best: the clinical guidance in the middle. Patient satisfaction improves. Staff productivity improves. Both are measurable.

The third application eliminates one of the most wasteful bottlenecks in portal workflows: manual message forwarding.

In most health systems, incoming messages land in a first-line queue. A staff member opens each one, reads it, and forwards it to the appropriate pool - scheduling, refills, nursing, or something else. That message might sit in the first queue for a full day before someone routes it. The patient waits. The staff member’s time is consumed by sorting, not solving.

Conversational AI models can classify messages and route them directly to the correct pool, bypassing the first-line queue entirely. Published routing models have shown meaningful accuracy and workflow improvements, but performance varies by health system, category design, training data, and implementation

The implementation insight is what makes this practical: “You don’t even have to label the data,” Don explains. “It’s auto-labeled by the past forwarding and pool assignment activity.” Healthcare organizations can train routing models on the last several months of their own forwarding behavior. The model learns where messages eventually end up and simulates that routing automatically. No manual annotation. Historical routing behavior can reduce manual labeling effort, but implementation still requires validation, governance, and monitoring.

Patient portal messaging is where the evidence is strongest, but conversational AI in healthcare extends across the entire patient journey:

• Appointment scheduling and reminders - Virtual assistants handle booking, rescheduling, and confirmations around the clock, reducing call center volume and missed appointments
• Symptom checking and clinical decision support - Structured triage flows help patients assess symptoms and find the right level of care, from self-care to urgent care to an in-person visit
• Medication management and chronic disease monitoring - AI-powered reminders improve adherence for patients managing conditions like hypertension or diabetes
• Mental health support - Conversational AI tools offer mental health check-ins, coping strategies, and guided self-care between therapy sessions, these are accessible and often anonymous, helping extend mental health support beyond clinical hours. These tools are not a substitute for clinical care because effective implementations include safeguards for crisis detection, escalation protocols, clinician oversight, and clearly communicated limits of use.
• Administrative tasks and billing inquiries - Answering frequently asked questions about insurance, billing, and patient records, reducing administrative tasks that consume staff time

The common thread: conversational AI handles routine, high-volume patient interactions so healthcare professionals can focus on complex clinical work.

The productivity gains are direct. Automated urgency tagging and intelligent routing reduce the time healthcare organizations spend sorting messages. AI-drafted responses cut composition time for contact center specialists. Published research on deployed routing models shows that routed message groups reduced initial response time by a median of one hour and total conversation time by over 22 hours compared to manually routed messages. Staff can serve more patients without increasing headcount.

Speed matters clinically. When urgent messages are flagged and addressed within minutes instead of hours, patient safety improves and patient outcomes follow. When empathetic, thorough responses reach patients faster, patient engagement and satisfaction increase. When medication reminders arrive on schedule, adherence improves, particularly for patients managing chronic diseases. Better patient care starts with faster, more accurate communication. Don’s assessment from working with InterSystems customers: “Very positive feedback from customers who have done it... improved turnaround time, contact center specialists have been very positive.”

Don makes a practical observation about risk: these AI in healthcare applications sit in a lower-risk category compared to diagnostic or clinical AI. “You’re reading all the messages anyway,” he says. For urgency tagging and routing, AI is sorting and filtering and the human is still reviewing every message. For proposed responses, the human edits and sends. The AI accelerates the workflow without removing clinical oversight.

The models themselves are small. “These are smaller language models which are fine-tuned versions of BERT from 2018. You can run them on a laptop or small server, no risk of sending data to cloud services,” Don notes. On-premises deployment eliminates the data privacy concerns that keep many healthcare organizations from adopting AI. Patient data never leaves the organization’s infrastructure.

Any conversational AI system handling patient data must comply with the Health Insurance Portability and Accountability Act (HIPAA). This means encryption, access controls, audit logging, and a Business Associate Agreement with any vendor involved.

Don’s point about on-premises models matters here. Urgency tagging and routing models are compact enough to run on local infrastructure with no cloud API calls, no external data transfer. Local deployment can reduce reliance on external AI APIs and limit data movement, but privacy depends on the full architecture, logging, vendor access, and governance controls.

For conversational AI platforms that require cloud connectivity, particularly those using larger language models for response generation, healthcare providers must verify HIPAA compliance, data residency policies, and whether patient data is used in model training. Electronic health records integration adds another layer: conversational AI tools need read access to patient records, which means governance around what data flows where.

Training data governance matters too. Past forwarding behavior can supply training labels; message content still requires secure handling, de-identification where appropriate, and governed access.

For healthcare organizations considering conversational AI, the practical path starts with lower-risk applications that deliver immediate value.

Start where accuracy is highest and risk is lowest. Urgency tagging and message routing both achieve 90%+ accuracy, operate with a human in the loop, and use compact models that can run on-premises. These are not moonshot projects. They are workflow improvements with published evidence behind them.

Use your own data. Don’s insight about auto-labeled training data removes a major barrier to adoption. If your contact center team has been forwarding messages for the past six months, you already have the training data for a routing model. No annotation project needed.

Keep clinicians in the loop. For AI-proposed responses, the model drafts; the human reviews and sends. This is the workflow that produced the empathy and quality gains in published studies.

Ensure your data foundation is ready. Conversational AI is only as effective as the data it can access. InterSystems IRIS for Health and Health Connect provide the interoperable data layer and integration engine that conversational AI platforms need to access patient records, route messages across systems, and connect with existing clinical workflows, particularly in environments with multiple EHRs.

The current generation of conversational AI in healthcare is largely reactive, it responds to messages, routes them, and drafts replies. The next generation of AI in healthcare will be proactive.

More sophisticated NLP models will detect not just urgency but sentiment gradients: a patient whose messages are growing shorter and more frustrated over time, a care plan that is generating more questions than usual. Conversational AI systems will integrate with wearable devices and remote monitoring data, identifying at-risk patients before they send a message at all.

Multilingual capabilities are expanding patient access, AI assistants that communicate fluently in a patient’s preferred language remove one of the most persistent barriers in care delivery. And as interoperability standards like FHIR mature, conversational AI tools will move more easily between healthcare systems, making deployment faster and less dependent on custom integration.

The health systems investing in conversational AI in healthcare today are building the communication infrastructure they will rely on for the next decade.