Patient portal messaging is now an operating capacity problem

Patient portal messaging has crossed an important threshold. It is no longer a convenience layered onto the electronic health record. It is a care-delivery channel with its own demand, queues, staffing requirements, response risks, and patient-experience consequences.

A 2026 JAMA study analyzed deidentified data from 2,067 hospitals and 47,100 clinics. Patient-authored portal messages increased from 0.99 per active patient per year in 2020 to 2.50 in 2025, a 153% increase. Office visits also rose 17% during the same period, from 2.37 to 2.77 per patient per year.

The operational conclusion is direct. Portal messaging did not simply replace visits. It added another stream of work.

For health-system executives, the question is no longer whether patients will use secure messaging. The question is whether the organization has designed the capacity and controls required to manage it.

More messages do not automatically mean better access

Digital communication can remove friction for patients who need a refill, an appointment change, a result explanation, or guidance about a new symptom. It can also create a hidden queue when every message arrives in a clinician inbox without clear routing or ownership.

That distinction matters because a patient sees one entry point, while the health system receives many different kinds of work. A scheduling request, medication question, billing issue, referral problem, and high-acuity symptom report should not follow the same path or wait under the same service standard.

When those needs remain mixed together, health systems often ask highly trained clinicians to sort administrative work before they can address clinical work. The patient experiences delay. The clinician experiences inbox burden. The organization loses a clear view of demand.

A better model treats patient portal messaging as an operating system with defined categories, owners, escalation rules, and measurable response times.

Routing can improve speed, but it must match real workflows

A 2026 JAMA Network Open quality-improvement study offers a useful example. Southern California Permanente Medical Group deployed a natural-language processing tool that classified incoming portal messages into labels aligned with clinical care pathways.

The study included 3,030,247 messages from 1,042,418 unique patients and reached inboxes associated with more than 1,000 primary care clinicians. The new tool achieved 81% classification accuracy, compared with 44% for the legacy system that relied on patients to select a subject. Median time to first read for messages classified as high acuity fell from 22.03 hours before deployment to 5.02 hours after deployment.

These results do not prove that classification alone caused the improvement. The work took place in one integrated, value-based system and used a quality-improvement design rather than a randomized trial. Still, the operating lesson is strong.

Technology worked because the labels connected to real care pathways. The organization also monitored changes in message patterns and updated its categories over time. Classification was not a stand-alone model. It was part of a maintained workflow.

Health systems considering automated routing should start with the work itself:

1. Define the message types that require different actions.
2. Assign each type to a team with the authority to resolve it.
3. Set service levels based on urgency and patient need.
4. Establish escalation paths when classification confidence is low.
5. Monitor drift as patient behavior, services, and risks change.

The goal is not to automate every response. It is to place each need in the right queue quickly enough for a qualified person to act.

Response performance is also an equity measure

A fast average response time can conceal meaningful differences among patient groups.

A 2025 JAMA Network Open study examined 3.5 million messages sent by 341,836 primary-care patients at Mass General Brigham. After adjustment, the rate of receiving a care-team response within one business day was 1.1 percentage points lower for Black patients and 1.1 points lower for Hispanic patients than for White patients. It was 4.9 points lower for patients with dual Medicare and Medicaid eligibility than for commercially insured patients, and 4.1 points lower for patients who preferred Spanish than for those who preferred English.

The study was observational and limited to one health system in 2021. It also found that clinic and time differences accounted for much of the observed disparity. That finding is operationally important. It suggests that response gaps can reflect where capacity, language support, and workflow resources are available.

Executives should therefore review portal performance by language, payer category, clinic, location, and other locally appropriate access measures. The purpose is not to assign intent to a difference. It is to identify where the operating model produces unequal response reliability.

Generative AI needs narrower jobs and stronger controls

The pressure to use generative AI for patient messages is understandable. Drafting support may reduce cognitive burden, but a fluent response can still contain a harmful error.

In a 2025 npj Digital Medicine simulation, 20 practicing primary-care physicians reviewed 18 patient messages with AI-generated draft responses. Four drafts contained errors. Each error was insufficiently corrected by 13 to 15 participants, and 35% to 45% of the erroneous drafts were submitted without any edits.

At the same time, 80% of participants said the drafts reduced cognitive workload, 90% trusted the tool, and 75% considered the drafts safe. The study was small and simulated, and participants did not have full patient records. It should not be treated as a failure rate for deployed health-system tools. It does show why positive user perception cannot substitute for safety measurement.

A second 2025 study found that patients reported high satisfaction across AI and human response scenarios. Disclosure that AI was involved produced only a small satisfaction reduction. The authors still recommended disclosure to preserve patient autonomy.

The practical implication is to give AI bounded roles. Classification, summarization, translation support, and draft preparation may each have value, but health systems should define where human review is mandatory, what evidence the reviewer sees, when automation must stop, and how errors are detected after launch.

A five-part operating model for patient portal messaging

Health systems can manage patient portal messaging through five connected controls:

Demand: Measure message volume, topic, source, timing, repeat contacts, and growth by clinic or service line.

Routing: Connect message types to teams that can resolve them, not merely forward them.

Service levels: Define response targets by urgency and need, then measure both first response and resolution.

Equity: Monitor whether language, clinic resources, insurance status, digital access, or workflow design create uneven reliability.

Accountability: Require human judgment for clinical decisions, maintain escalation rules, test AI-supported workflows, and disclose automation where appropriate.

This is the same operating principle PEC360 applies across patient-experience work. Technology should reduce repetitive coordination while preserving human judgment for clinical context, empathy, exceptions, and complex problem-solving. The value comes from a designed workflow, not the presence of AI alone.

The executive takeaway

Patient portal messaging has become infrastructure. If it remains an unmanaged inbox, more digital access can produce more hidden work, slower responses, and uneven service.

The stronger approach is to treat secure messaging as a measurable operating channel. Define the demand. Route work to resolution. Set service levels. Watch for access gaps. Use AI for bounded tasks with visible human accountability.

Health systems do not need fewer ways for patients to reach them. They need an operating model that can reliably respond when patients do.

Authoritative sources

1. Long JJ, McAdams-DeMarco MA, Schwartz MD, et al. “Trends in Patient Portal Messages, Office Visits, and Telephone Encounters.” JAMA. 2026. https://jamanetwork.com/journals/jama/fullarticle/2850607
2. Nguyen D, Lee S, La K, et al. “Performance of an Intelligent Messaging Tool for Clinical Communications.” JAMA Network Open. 2026. https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2843966
3. Rotenstein LS, Hardy B, Tang M, et al. “Patient-Physician Messaging by Race, Ethnicity, Insurance Type, and Preferred Language.” JAMA Network Open. 2025. https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2839740
4. Biro JM, Handley JL, McCurry JM, et al. “Opportunities and risks of artificial intelligence in patient portal messaging in primary care.” npj Digital Medicine. 2025. https://www.nature.com/articles/s41746-025-01586-2
5. Cavalier JS, Goldstein BA, Ravitsky V, et al. “Ethics in Patient Preferences for Artificial Intelligence-Drafted Responses to Electronic Messages.” JAMA Network Open. 2025. https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2831219