Your AI Scribe Only Heard Half the Telehealth Visit. The Other Half Is What Auditors Read.

Medicare telehealth is no longer a cliffhanger. In early February 2026, the Consolidated Appropriations Act, 2026 (H.R. 7148) extended the flexibilities through December 31, 2027 in Section 6209 — ending a 43-day lapse during the 2025 shutdown and giving practices the longest planning runway they have had since 2020. Video visits are a permanent line item in the schedule now, not an emergency measure.

Which means the documentation problem is permanent too.

Most practices running virtual visits have also adopted an AI medical scribe. The pairing looks obvious: the encounter is already audio, the scribe is already listening. But ambient scribes were designed for a room — one microphone, two people, shared air. A telehealth visit is a stream, and two things go wrong that nobody warns you about. The scribe often hears only your half of it. And the parts of a telehealth note that auditors actually read are the parts no ambient scribe will ever write.

What is locked in, and the one date that is not

Before the workflow, the policy floor, because it determines what you have to document:

• Coverage through December 31, 2027. Geographic and originating-site waivers, the patient's home as an eligible site, audio-only services, the expanded practitioner list (including PTs, OTs, SLPs and audiologists), and FQHC/RHC distant-site billing all run to the end of 2027.
• Behavioral telehealth is permanent. Geographic and originating-site restrictions were permanently removed for mental health services back in the Consolidated Appropriations Act, 2021. The in-person visit requirement that would otherwise attach to tele-mental health remains delayed — currently due to take effect in January 2028.
• The CY2026 Physician Fee Schedule cleaned up the list. Effective January 1, 2026, CMS eliminated the "provisional" versus "permanent" distinction on the Medicare Telehealth Services List (everything on it is now permanent), removed frequency limits on subsequent inpatient visits, subsequent nursing facility visits and critical care consultations, and permanently adopted a definition of direct supervision that can be satisfied through real-time audio and video — audio-only does not count, and higher-risk surgical services with 010 and 090 global indicators still require physical presence.
• The one real cliff is the DEA's. The Fourth Temporary Extension of COVID-19 Telemedicine Flexibilities for Prescription of Controlled Medications, published in the Federal Register on December 31, 2025, allows Schedule II–V prescribing via telemedicine without a prior in-person exam only through December 31, 2026. It imposes no new documentation, registration or technology requirements. The special-registration framework DEA proposed in January 2025 has still not been finalized. If controlled substances are part of your virtual practice, that is four months of runway, not two years.

Everything above tells you what you may bill. None of it tells you what your note has to say — and that gap is where AI scribes quietly fail.

Problem one: your scribe is probably half-deaf

In a physical room, the microphone hears everyone. On a video call, it does not, and the reasons are technical rather than clinical.

Browsers capture a tab, not a system. Heidi's own documentation is unusually candid about this: browser-based capture is "limited to tabs in the same browser," so audio playing through any other application can be missed entirely. Its desktop app exists precisely to capture system-wide audio instead. Heidi requires you to select Telehealth mode — not Transcription or Dictation — to enable tab audio sharing at all, and its mobile app with headphones cannot capture the patient's side; the guidance is to use a separate device.

Echo suppression eats the patient's half. This is the failure that catches people. Browsers apply echo cancellation to stop feedback loops, and that processing "can unintentionally block parts of the patient's audio." Heidi's documented workaround is to toggle echo suppression off in Settings when running the scribe and the telehealth platform in the same browser. Safari and Firefox users are told to abandon headphones and run loudspeaker mode instead.

Microphone contention is real. Two applications competing for the same input device produces a transcript of you talking to nobody. Both Heidi and other vendors advise confirming permissions and closing any other app holding the microphone.

Vendors solve it differently, and the difference matters. Freed uses a Chrome or Edge extension with a dedicated "Capture telehealth" mode to pick up both the provider's microphone and the patient's audio, supporting browser-based Zoom, Google Meet, Microsoft Teams and browser-based EHR video modules — plus a Zoom Marketplace app that joins the meeting as a silent participant and needs no extension at all, with capture up to two hours. Nabla and Notable also sit in the telehealth platform category. The right question for any vendor is not "do you support telehealth" — everyone says yes — but how they capture the far-side audio, and on which browser, app and operating system.

Why this matters more than it sounds: the dominant error mode of ambient scribes is already omission, not fabrication. In a controlled evaluation of five ambient scribe platforms across 14 simulated ambulatory encounters (Mayo Clinic Proceedings: Digital Health, 2025;3(4):100292), omissions accounted for 76.3% of all errors. Platforms averaged 13.9 transcript errors per case, roughly 19.5% of which carried through into the clinical note, producing a mean note error rate of 26.3% and an average of 3.0 errors per case with potential for moderate-to-severe harm. Only 35.8% of clinical elements were consistently correct across all five platforms.

And here is the part that should worry any telehealth clinician: those encounters were run under optimized conditions — prerecorded audio, no background noise, clean capture. A scribe that omits three-quarters of its errors by leaving things out, fed a stream where the patient's voice is being suppressed by the browser, does not produce an obviously broken note. It produces a fluent, confident note about a conversation it only partly heard. Our guide to what the research says about AI scribe accuracy covers the review workflow that catches this.

Test before a real patient. Run a five-minute rehearsal call with a colleague on the far end, using the exact browser, headset and platform you will use in clinic. Then read the transcript and count how many of their sentences appear. If the answer is "most but not all," you have found your problem before a patient did.

Problem two: the five lines the AI will not write

Your scribe produces a clinical narrative: HPI, assessment, plan. Telehealth claims are not usually denied over the narrative. They are denied over the metadata — and the metadata is exactly what an ambient model has no way to know.

The clearest illustration is still an OIG audit from April 1, 2020 (A-04-18-00122). Reviewing a stratified random sample of 100 South Carolina Medicaid fee-for-service telemedicine payments, OIG found 3 allowable and 97 unallowable. In 95 of them, providers had documented neither the start and stop times nor the consulting site location of the service. Extrapolated, an estimated 96% of the state's Medicaid fee-for-service telemedicine payments were unallowable. Not one of those failures was a clinical judgment error. They were missing lines.

Telehealth remains an active OIG audit target — the Work Plan item covering audits of Medicare Part B telehealth services runs in phases, with the first covering evaluation and management, opioid use disorder, end-stage renal disease and psychotherapy services.

So add these to your telehealth note template, and enter them yourself:

1. Modality, stated explicitly. Real-time audio and video, or audio-only. This single line drives the modifier, and getting it wrong is the most common denial driver in telehealth billing today.
2. The patient's physical location during the encounter. Not their mailing address — where they actually were. This governs licensure, recording law and place of service.
3. Your location as the distant-site practitioner.
4. Consent, and which kind. Most states require telehealth-specific informed consent documented in the medical record; the Center for Connected Health Policy tracks the state-by-state variations, which range from verbal-if-documented to written-only.
5. Who else was on the call, and start/stop times. Interpreters, family members, supervising physicians, residents. Under the CY2026 rules, if a teaching physician participated virtually, the record must state whether they were physically present or joined virtually.

If a scribe drafts any of these, treat it as a hallucination and delete it. The model is inferring, not observing.

The coding traps a scribe will walk you into

Ambient scribes default to producing an office-visit note. Left alone, that default is wrong in three ways:

• Modifier 93 versus 95. Modifier 93 identifies a synchronous audio-only encounter; modifier 95 identifies synchronous audio-video. Billing 95 on a visit that never had working video is overcoding, and it is recoupable. Many commercial payers still do not recognize 93 at all, so confirm per payer before you send audio-only claims. If video was attempted and failed, or the patient declined it, say so in the note — that sentence is what supports the audio-only bill.
• POS 02 versus POS 10. POS 10 is the patient at home; POS 02 is the patient anywhere else. Since 2022 the overwhelming majority of visits are POS 10, which is exactly why POS 02 gets missed when a patient calls in from work or a car.
• The 98000 series is a trap on Medicare. CPT's telemedicine E/M codes 98000–98015 carry status indicator "I" on the Medicare PFS — they will deny. Medicare wants standard office-visit codes 99202–99215 with the telehealth POS and modifier. The one exception is 98016, the brief virtual check-in that replaced HCPCS G2012 for 5–10 minute technology-based communication with an established patient, not originating from an E/M in the prior 7 days and not leading to one within 24 hours. Some commercial and Medicaid plans have adopted the full 98000 series; many have not.

Any of these can turn a clinically excellent AI-drafted note into a denied claim. The related risk runs the other direction too: a scribe that generates a richly detailed exam for an encounter conducted over video invites the pattern analysis described in our piece on AI scribes, coding intensity and billing audits. Document the exam you actually performed through a screen — nothing more.

Consent stacks up faster on telehealth than in person

In an exam room, recording consent is one conversation. On a video visit it can be four at once:

• State telehealth consent, documented in the chart before or at the start of the encounter.
• Recording consent under the wiretap statute of the patient's state — a live two-way electronic communication is squarely within these laws, and all-party-consent states key off where the patient is sitting, not where you are. Cross-state telehealth means more than one rulebook applies to the same visit.
• A HIPAA authorization if the audio will be used for anything beyond treatment, payment or operations — model training in particular.
• 42 CFR Part 2 consent if substance use disorder treatment is involved.

For behavioral health, where telehealth flexibilities are permanent and virtual delivery is the norm, this stack is the everyday case rather than the exception. Practices using AI therapy notes should read our deeper treatment of AI scribe consent and recording laws, alongside the state AI disclosure rules that increasingly require telling patients an AI was involved at all.

A workflow that holds up

1. Pick your capture path and freeze it. One browser, one platform, one headset, documented for the whole practice. Prefer a desktop app with system-audio capture over a browser extension where the vendor offers both.
2. Rehearse it, then re-rehearse after every vendor or browser update. Echo suppression settings reset. Extensions update. Verify the far side is still in the transcript.
3. Take consent before you start capture, and log it in that order. Telehealth consent, then recording consent, then record.
4. Confirm the patient's physical location out loud at the top of every visit. It is one sentence, it lands in the transcript, and it answers three compliance questions at once.
5. Use a telehealth note template with the five metadata fields pre-stubbed so the AI draft drops into a scaffold that already has blanks you must fill.
6. Read the patient-attributed content hardest. That is the half most likely to have been suppressed, and omissions do not announce themselves.
7. Set modality, POS and modifier before the claim leaves. Never let them inherit from an in-person template.

Done properly, the payoff is real. A 2025 preprint from Included Health describing a custom ambient scribe built for a telehealth-first practice — Whisper transcription with GPT-4o note generation — reported that 94% of surveyed clinicians (n=63) experienced reduced cognitive load during visits and 97% (n=66) reported decreased documentation burden, across more than 540 clinicians who used the tool at least once. It is a vendor-authored preprint on a purpose-built internal system rather than an independent trial of a commercial product, so read the numbers as directional. But the direction is consistent with what multisite ambient scribe research has been showing in person.

The bottom line

Telehealth stopped being temporary, and your documentation process should stop being improvised. The two failure modes are specific and both are fixable in an afternoon: an ambient scribe that captures only your side of the stream, and a note that reads beautifully while omitting the modality, the two locations, the consent and the times that determine whether the claim survives.

Fix the audio path first, because everything downstream is built on the transcript. Then stop asking the AI to supply facts it cannot observe, and put those five lines in a template where a human fills them in every single time.

If you are choosing a tool, start with the telehealth-capable platforms in our directory, compare scribes by specialty, or read what Medicare actually pays for when AI is involved.

---

This article is for general information only and is not legal, medical, or billing advice. Telehealth coverage, coding, prescribing and consent rules change frequently and vary by state and payer; verify current requirements with CMS, your state board, your payers, and qualified counsel before relying on any specific rule. Always confirm a vendor's current telehealth capture capabilities and compliance posture directly with the vendor.