The demos ended. The deployments are reporting.
Two years of agentic AI keynotes have finally produced something more useful than keynotes: production evidence. Commercial triage systems with built-in risk detection launched this summer. The major EHR vendors are shipping agent orchestration, one now reports the overwhelming majority of its customers using AI features. And across DTC telehealth, operators have moved past the pilot phase into the unglamorous part: agents handling real patient conversations at volume, with metrics.
The production evidence tells a more specific story than the keynotes did. Some agent patterns work brilliantly and patients prefer them. Some work only with tight human-in-the-loop design. And a few failure modes recur so predictably that avoiding them is now just professionalism. This post is the field report, organized for operators deciding what to turn on next.
The platform-level architecture, data, events, actions, runtime, is covered in The Agentic Telehealth Platform. The earlier support-layer map is in AI Agents for Telehealth Support. This is the patient-facing production layer.
The four patterns that have proven out
1. Status and logistics resolution
The workhorse. Where is my medication, when is my appointment, how do I reschedule, what did my payment cover. In production, agents grounded in live fulfillment, scheduling, and billing data resolve the large majority of these conversations end to end, faster than any human queue, at any hour. Patients do not merely tolerate this; they prefer it, because the agent answers in seconds with the actual tracking state.
The dependency is the grounding: an agent reading real pharmacy status events, per the pipeline in Pharmacy Status Visibility in Telehealth, is a service; an agent guessing is a liability.
2. Intake assistance and completion recovery
Agents that help patients through intake, clarifying questions, resolving upload problems, resuming abandoned sessions, and increasingly parsing free-text history into structured data, have earned their place in production. The measured effect shows up directly in completion rates, with the strongest gains on mobile and among the AI-informed arrivals described in When Your Next Patient Arrives via ChatGPT.
3. Structured triage with risk detection
The newest proven pattern, validated by this summer's commercial launches: agents that conduct structured symptom and side-effect check-ins, score responses against protocol, and route accordingly, routine reassurance delivered directly, concerning patterns escalated with context attached. The design insight from the systems working in production: the agent's job is detection and routing, not judgment. It widens the funnel of patients checked on, and narrows the stream reaching clinicians to the ones who need them.
4. Between-visit adherence and milestone touches
Daily-medication check-ins, titration-window education, taper monitoring per the cadences in The Maintenance Evidence Just Arrived, delivered conversationally and logged to the chart. Production experience says patients engage with these at rates static reminders never achieved, provided the agent remembers context: the check-in that knows last week's answer feels like care; the one that does not feels like a form with a personality.
The escalation architecture
Every production failure story traces to the same root: escalation treated as an exception rather than an architecture. The deployments that work design the handoff first:
Hard boundaries, encoded. Clinical judgment, dosing decisions, distress signals, and anything ambiguous route to humans, not because agents always fail there but because the cost asymmetry demands it. The boundary list is written by clinical leadership, versioned, and enforced in the runtime, which is governance made operational per Clinical Governance Is a Growth Asset.
Context-carrying handoffs. The measure of a good escalation is that the patient never repeats themselves. The agent's summary, the relevant history, and the trigger reason arrive with the conversation. Support teams report this single property, warm handoff versus cold transfer, dominates patient satisfaction with the entire AI layer.
Escalation as signal. The pattern of what escalates is a product roadmap: recurring escalation clusters mark either an agent capability to add or, more often, an upstream program defect to fix, the same upstream logic as ticket clustering in Your First 100 Telehealth Patients.
Disclosure, now with statutes
Honest disclosure was always the right design; as of this summer it is increasingly the law, with state AI-in-care statutes taking effect in July and more scheduled, covering disclosure, consent, and boundaries on AI in therapeutic contexts.
Production experience has converged on the pattern that satisfies both patients and regulators: disclose clearly at first contact, in plain language, with an always-available path to a human, and never design the agent to pass as one. The operators who feared disclosure would hurt engagement report the opposite: patients use disclosed agents more readily for logistics and check-ins, precisely because the boundaries are legible.
Platforms that build disclosure, consent capture, and human-path affordances into the agent runtime turn a compliance requirement into configuration, which is the recurring theme of every governance-meets-AI question this year.
Measuring agents like employees
The deployments that improve month over month share a measurement habit: agents get scorecards, reviewed in the same weekly ops rhythm as The Weekly Telehealth Ops Dashboard:
| Metric | What it tells you |
|---|---|
| Containment with satisfaction | Conversations resolved end to end where the patient rated the outcome positively; containment alone is a vanity number |
| Escalation precision | Of escalated conversations, how many genuinely needed a human; both false positives and false negatives cost |
| Time to resolution | Against the human-queue baseline, per conversation type |
| Handoff quality | Repeated-information complaints as the inverse indicator |
| Trust trajectory | Patient willingness to use the agent again, trending by cohort |
| Chart completeness | Whether agent interactions land in the record as reliably as human ones |
The last row is the sleeper. Agent conversations are care interactions; production programs route them to the chart with the same discipline as visits, which is what makes the whole layer auditable when partners and regulators ask, and they now ask.
FAQ
What AI agent use cases actually work in telehealth today? Four have proven out in production: status and logistics resolution grounded in live data, intake assistance and completion recovery, structured triage with risk detection and routing, and between-visit adherence check-ins with context memory. Clinical judgment remains with humans by design.
How should telehealth AI agents hand off to humans? Through designed escalation: clinically-authored hard boundaries encoded in the runtime, context-carrying handoffs so patients never repeat themselves, and monitoring of escalation patterns as product signal. Handoff quality dominates patient satisfaction with the entire AI layer.
Do patients accept AI agents in healthcare? For logistics, intake help, and structured check-ins, patients often prefer them, faster answers, any-hour availability, provided the agent is clearly disclosed, remembers context, and offers an always-available human path. Undisclosed or context-free agents damage trust quickly.
Are there laws about AI agents in patient care? Yes, and growing: state statutes effective mid-2026 impose disclosure, consent, and boundary requirements for AI in care contexts, particularly therapeutic ones. Programs should treat clear disclosure and human-path affordances as both compliance and good design.
How do you measure an AI support agent? Like an employee: containment with satisfaction, escalation precision, time to resolution versus baseline, handoff quality, repeat-use trust, and chart completeness of agent interactions, reviewed on the same cadence as other operational metrics.
The boring maturity is the news
The agentic era's real milestone is not a demo; it is the week an operator stops thinking about the agent because it simply runs, resolving the logistics, catching the risks, escalating with context, and logging to the chart. Hundreds of programs quietly reached that week this year.
The patterns are proven, the guardrails are knowable, and the infrastructure is a configuration screen away. What remains is the operator's move: turn on the first pattern, measure it like a hire, and let the scorecard tell you when to turn on the second.