AI Agents for Telehealth and Virtual First Groups

Because in virtual care, intake is not a form. It is the substrate of the visit.

A well run intake agent collects history, symptoms, medications, allergies and consents before the encounter, asks the follow up questions a static form cannot, and hands the clinician a structured summary rather than a wall of free text. Three things follow. Visit length drops because the clinician is confirming rather than discovering. Documentation quality rises because structured data is already in the record. And some encounters are correctly routed to asynchronous care or to a different channel entirely, which is capacity you did not have to hire for.

This is also where the difference between an intake form and an intake agent actually shows. A conditional form branches on rules somebody wrote. An agent can ask the obvious follow up question that nobody anticipated, which in practice is most of the useful questions.

Two design constraints keep it safe. The agent gathers and summarises; it does not conclude. And the summary presented to the clinician must show what the patient actually said, not only the agent's paraphrase, because a clinician reviewing a paraphrase is reviewing the agent rather than the patient. Our comparison of intake products covers which vendors expose the underlying responses and which do not.

• Asynchronous intake. Highest leverage, for the reasons above, and it touches the metric that matters most: completed visits per clinician hour.
• Documentation inside the video platform. Easier here than in any physical setting, because the audio stream is already in software rather than in a room. No microphones, no hardware rollout, no exam room acoustics.
• Message triage. Virtual first models generate far more asynchronous messaging per patient than clinic models do, and the volume grows faster than the panel. Drafting replies for clinician approval is the defensible pattern.
• Scheduling with licensure constraints. Matching a patient to a clinician who is licensed in that patient's state, available, and appropriate for the complaint is a genuine constraint problem and a real source of abandoned bookings.
• Voice support. Lower priority for most virtual first groups, since patients who chose an app rarely want a phone call, but it matters if you serve older or Medicare populations.

Effort assumes an in house engineering team and API access to your own platform, which is the normal case in this segment and materially lower than elsewhere.

The pattern here is the inverse of the hospital pattern. Technical effort is low and clinical governance is the gate. That is a better problem to have, and it is a different one from the one most published guidance addresses.

This is the compliance question that is genuinely specific to your setting, and it is usually underestimated.

Start with the licensure layer. The Interstate Medical Licensure Compact provides an expedited pathway for eligible physicians across a large majority of states plus the District of Columbia and Guam. It is worth being precise about what it is: it speeds up obtaining separate licences from each participating state. It does not create a single national licence, and it does not harmonise the practice rules those states apply. Comparable compacts exist for nurses and for psychologists, with their own membership and their own limits.

Now add the AI layer on top. Because the governing rules generally follow the patient's location, a group operating in thirty states inherits every applicable state AI provision at once. California requires disclosure when generative AI is used in certain patient communications and constrains AI in utilisation review. Utah imposes disclosure duties and has legislated specifically about mental health chatbots. Colorado and Texas have each taken a different approach again.

Maintaining thirty configurations is not a viable operating model, and every group that tries it discovers the same thing: the configurations drift and the audit finds the drift. The workable answer is to build to the strictest applicable requirement and apply it everywhere. Disclose that an agent is software at the start of every patient facing interaction, keep a human review step wherever a state requires one anywhere, and log both. It costs you a little conversion and removes an entire category of legal exposure.

Ambient documentation is easier to deploy in telehealth and harder to consent for, because the encounter spans two jurisdictions by construction.

Several states require the consent of all parties to record a conversation. In a video visit, the parties are in different states, and the conservative and workable approach is to obtain explicit consent from the patient every time rather than to maintain a state by state matrix that will be wrong within a year. Make it a spoken question at the start of the visit, log the answer, and make it easy for the clinician to proceed without recording if the patient declines.

Then ask your vendor the questions that matter more than the consent language. Is video or audio retained, and for how long. Is a transcript stored separately from the note, and can retention be set to zero. Is any content used to train models, and can that be declined in the business associate agreement rather than in a policy page that can change. Our HIPAA and AI page works through what those contract terms need to say.

One further point specific to virtual care: the patient is at home, and other people are in the room. A consent that covers the patient does not cover a family member who wandered past the camera. Vendors rarely raise this. It is worth a line in your consent script.

Sometimes, and less often than the engineering team will argue.

Build where the workflow is your differentiator. If your intake logic encodes a clinical model that is the reason your outcomes are good, that belongs in house, and you have the platform access to do it properly. Build also where you need the agent embedded so deeply in your own product that a vendor's interface would be visible to patients.

Buy where the capability is commodity and the maintenance is continuous. Documentation is the clearest example. Note formats, specialty templates, model updates and payer expectations all change constantly, and a team that builds a scribe signs up to maintain it forever against vendors whose entire company does nothing else.

The failure pattern we see most often in this segment is a six month internal build of something available off the shelf, justified by an integration concern that a two week evaluation would have resolved. Before committing engineering quarters, run the shortest possible bake off against one bought option. If the bought option is within reach, the build is a decision to spend your scarcest resource on a solved problem.

Whichever way you go, price it per encounter. A tool at a fixed monthly price per clinician looks expensive at low volume and very cheap at high volume, and virtual first groups change volume faster than any other provider type.

The absence of a physical room removes several safety nets that clinic based providers rely on without noticing, so the line sits in a different place here.

• Prescribing decisions, and controlled substances above all. The federal rules for remote prescribing of controlled substances have been extended repeatedly and remain unsettled, so verify current status before designing anything near them. Independent of that, the prescribing decision is the clinician's, and an agent that pre populates one is exerting influence you cannot audit.
• Emergency recognition and routing as a fully automated path. In a clinic, someone in acute distress is visible to staff. In a queue, they are a message. An agent should escalate aggressively on any signal and should never be the last reviewer of one.
• Licensure and state matching without a hard verified check. Getting this wrong is practising without a licence. Treat the credentialing database as the source of truth with a blocking check, not as an input the agent weighs.
• Identity verification by conversation alone. Virtual care has a genuine identity problem, and a conversational agent is the wrong tool for solving it.
• Diagnosis or triage conclusions in asynchronous care. Gathering asynchronously is fine and valuable. Concluding asynchronously without a clinician is a different product with a different regulatory status, which our page on FDA regulation of AI medical devices covers.

Instrument before you build. The four numbers worth having are median visit length, completion rate from booking to visit, message response time, and note completion lag. You almost certainly already log all four, which puts you ahead of most providers before you start.

Then take intake first. It is the highest leverage change available to you, it improves every downstream metric at once, and the clinical governance work it forces is work you would have to do eventually anyway. Documentation is a good second project precisely because it is easy: use it to prove your review, logging and consent patterns work before you apply them somewhere harder.

Where an outside view helps most in this segment is the multi state compliance layer, because it is the part your engineering team cannot solve by building well. Our AI readiness audit covers exactly that ground: which of your workflows are ready, what your state footprint actually obliges you to do, and where a build is genuinely warranted. The free AI readiness assessment gives you a first read in a few minutes. Groups delivering therapy or psychiatry by video should also read our behavioral health page, where the confidentiality rules are stricter again, and those serving walk in style same day demand may find the front door reasoning on our urgent care page transfers.