A clinician three years out of licensure takes a contract with a telehealth platform: steady referrals, a built-in video room, no billing headaches. Eight months in, she notices a line in the session interface she had not seen before: "Session insights available." Somewhere between her contractor agreement and a product update, the platform began processing her sessions, and she cannot say when, under what consent, or where the transcripts live. If you work with or are considering Talkspace, BetterHelp, Tava Health, Lyra Health, Spring Health, NOCD, or Iris Telehealth, this is your lesson: these platforms have already made AI decisions for you, through contractor MSAs and platform-side BAAs you signed to get the caseload. By the end, you will be able to map how each category of telehealth platform deploys AI to contracted clinicians, identify where your contract may have consented on your behalf to ambient capture, transcript retention, and outcome data sharing, and produce a contract-clause checklist you can run before signing, or against the agreement you already signed.

The Furnished Office: One Analogy for the Whole Lesson

Extend the housing metaphor one final step. In lesson one you were a buyer credentialing a contractor. In lesson two you were a homeowner weighing the builder's in-house crew. In lesson three you were a tenant whose landlord renovates without asking. On a telehealth platform, you have moved into a fully furnished office inside someone else's clinic: the video room is theirs, the chat system is theirs, the scheduling, the intake, the outcome measures, the client relationship itself often originates with them, and increasingly, the microphone is theirs too. The furnished office is undeniably convenient, that is why clinicians take these contracts, but every piece of furniture can have instrumentation in it, and the inventory of what is instrumented lives in two documents most contractors skim: the Master Services Agreement (MSA) you signed as a contractor, and the platform-side BAA architecture that governs data flows you never see.

This changes the consent question in a way the previous three lessons did not. With a scribe, you decide to record and you obtain the client's consent. On a telehealth platform, the client often consented to the platform's terms before ever meeting you, the clinician consented through the MSA before ever meeting the client, and the AI capabilities can sit inside the infrastructure both of you already agreed to use. The phrase to hold onto: your contract may have already consented on your behalf. The work of this lesson is finding out whether it did, for which capabilities, and what your own professional duties still require on top of whatever the paperwork technically authorized.

Three Species of Platform, Three Different AI Postures

The seven names in this lesson's title are not one category; they are three, and the species determines the AI posture you should expect. Species one: direct-to-consumer therapy platforms, Talkspace and BetterHelp. The client subscribes to the platform, the platform matches them to you, and large parts of care may flow through asynchronous messaging, a modality that is natively text and therefore natively machine-readable. Where therapy itself is text in a platform's system, AI-adjacent processing (matching, message routing, quality monitoring, safety screening) is an architectural temptation the platform does not have to bolt on; the data is already sitting in its database. Species two: employer-market enterprise platforms, Lyra Health, Spring Health, and Tava Health. Here the paying customer is an employer, the product is measurable outcomes for a covered population, and the platform's AI energy concentrates on matching, outcome measurement, and population analytics, because demonstrating that therapy worked, at scale, in dashboards, is the product the employer is buying. Species three: specialty and infrastructure providers, NOCD (OCD-specialized care built around ERP) and Iris Telehealth (telepsychiatry staffing into health systems). Their AI posture follows their specialty: NOCD's value rides on structured, protocol-driven treatment and measured outcomes; Iris embeds clinicians inside health systems whose own AI and EHR rules then also apply, a contract inside a contract.

Why does the species matter to you? Because it predicts what the platform wants from your sessions. A consumer messaging platform wants engagement, retention, and safety coverage across millions of messages. An enterprise platform wants outcome deltas it can show an employer: PHQ-9 and GAD-7 trajectories, time-to-improvement, utilization patterns. A specialty platform wants protocol fidelity and condition-specific measures. Each desire shapes what gets captured, retained, and analyzed, and therefore which clauses in your contract deserve the closest read. The clinician who reads her Lyra or Spring contract looking only for "recording" language will miss the outcome-data clauses doing the heavier lifting; the clinician reading a BetterHelp or Talkspace agreement should be most alert to what happens to the message archive that is the therapy.

Ambient Capture: When the Room Itself Listens

Ambient capture means the session medium itself records or processes the encounter: the video room transcribes, the audio is analyzed, the chat is processed, without the clinician pressing a record button. On a telehealth platform, ambient capability is architecturally trivial, the platform already owns the pipes, which is exactly why your scrutiny has to be contractual rather than technical. Three questions define the ambient terrain. Is capture happening (transcription, audio analysis, message processing), and is it disclosed to you as the clinician, not buried in a product-update email? Who consented, and when: did the client's platform terms cover it, did your MSA cover it, and does either consent meet the standard your license holds you to? And what is retained: a transcript that exists for the seconds it takes to draft a note is a different object from a transcript archived for "quality and product improvement."

Hold two professional anchors here. First, your state's recording-consent and your board's informed-consent expectations apply to you personally; "the platform's terms covered it" has never satisfied a licensing board examining whether your client understood who or what was listening. If ambient processing touches your sessions, your own consent conversation and paperwork need to describe it in plain language, whatever the platform's terms say. Second, the regulatory bright line from this chapter's first lesson holds with special force here: under the Illinois WOPR Act and Nevada AB 406, AI cannot provide therapy, and New York's AI companion safeguards law (2025) polices the adjacent consumer-chatbot territory. A platform whose marketing drifts toward AI that "supports clients between sessions" is operating near a line that statutes now patrol; your name and license are attached to the care on that platform, so you need to know exactly which side of the line every feature sits on. AI that drafts your note after the session is documentation support. Anything that talks to your client therapeutically is a different animal, and in several states now, a regulated or prohibited one.

On a telehealth platform, the microphone may belong to the house, but the license in the room is yours: no MSA clause, platform consent, or product update transfers your duty to know what is captured, tell your client the truth about it, and keep every clinical determination human.

Transcript Retention and the Message Archive That Is the Therapy

Retention questions from lesson one return here with a twist worth slowing down for. On a messaging-first platform, the transcript is not a byproduct of therapy; in asynchronous care, the transcript is the therapy. Every reframe you offered, every disclosure the client made at 1 AM, every rupture and repair exists as durable text in the platform's systems. That archive raises the same questions as scribe retention (how long, for what purposes, deleted or soft-flagged, used for training or not) at a categorically larger scale, because there is no version of messaging therapy where the content was never captured. For video-first platforms, the questions attach to transcripts and any session-derived artifacts: summaries, "insights," quality-review samples. In both cases, ask where your own clinical notes live, who can access them platform-side, and what follows you if you leave: the contractor who departs a platform typically leaves the records behind, which means your documentation of your own clinical decisions, the record that defends you in a board complaint years later, lives in a system you no longer have access to. Negotiate, or at least document, your continuity-of-access expectations before you need them.

And carry forward the psychotherapy-notes discipline from lesson one: the HIPAA carve-out (45 CFR 164.501 definition; 45 CFR 164.508(a)(2) protections) covers your private process notes only if they are kept separate. A platform's chat thread, session transcript, or note field is not that separate place. Whatever reflective, hypothesis-level writing you do about clients on a platform caseload, do it outside the platform's systems, in a location you control, under the heightened protection the carve-out was written to provide.

Outcome Data Sharing: The Dashboard Your Client Becomes

Enterprise behavioral health runs on measurement. Lyra, Spring Health, and Tava sell employers a population product: access, engagement, and improvement, demonstrated with instruments like the PHQ-9 and GAD-7 administered through the platform at intervals. Measurement-based care is clinically legitimate, often clinically excellent; the questions are about the data's second life. Your client's score trajectory feeds, in aggregated or de-identified form, the dashboards that justify the employer contract. Three things to understand calmly. First, aggregation and de-identification are real protections when done properly, and the platforms have strong incentives to do them properly. Second, your client may not realize that their employer's purchase of the benefit is justified by analytics their scores participate in; nothing improper has to be occurring for the client to deserve a plain-language understanding of it, and your consent conversation is where that understanding can live. Third, the clinical instrument itself stays clinical: the PHQ-9 is administered and interpreted in your care of this client, and a platform analytics layer never replaces your reading of item nine, the suicidal ideation item. Which is the place to restate this chapter's non-negotiable: AI never scores the CSSRS, never assigns a risk level, never makes the Tarasoff or duty-to-protect determination, never makes a mandated-report call. Platforms run safety-screening systems over messages and check-ins, and those systems can surface content to you faster than you would have seen it; treat every such flag as routing, not assessment. The flag tells you where to look. You decide what it means, you conduct the risk assessment, and you document a determination that is yours.

NOCD illustrates the specialty version of the same structure: condition-specific measures (for OCD care, symptom severity tracking through validated instruments) woven through a protocol-driven model. Specialty measurement is often the best version of measurement-based care, and the same two duties ride along: know where the data goes, and keep the interpretation human. Iris Telehealth illustrates the layered version: an Iris clinician working inside a hospital system sits under the Iris MSA and the health system's own EHR, AI, and consent policies simultaneously, and must know which rules govern which artifact, the kind of layered-obligation reading this chapter has been training you for.

Reading the MSA: The Five Clauses That Decide Everything

Apply the targeted-pass skill from the aggregator lesson to the contractor MSA, where five clause families decide everything this lesson cares about. One: data and recording. Find every grant of rights over session content, audio, video, chat, transcripts, and the purposes attached (care delivery, quality, safety, research, product improvement, model training). "Product improvement" and "research" are the phrases to circle; they are where a session can become training data. Two: technology and modification. Platforms reserve the right to update their technology; find whether new AI capabilities require notice to you, and what form notice takes, because "session insights available" appearing in the interface is what no-notice looks like. Three: documentation and records. Where do clinical records live, who owns them, what access do you retain during and after the contract, and what is the records-retention commitment if the platform is acquired or shuts down? Four: your professional independence. Strong agreements state plainly that clinical judgment belongs to the clinician; read how yours handles protocol requirements, safety-system responses, and anything that could read as the platform directing care. Five: indemnification and insurance. Who bears the cost if the platform's AI mishandles data from your sessions, and does your own malpractice coverage (the carriers' renewal questionnaires now ask about AI) align with what the platform's systems are doing around your care?

Read with the same auditor's eye you trained on vendor marketing. "We use advanced AI to improve care quality" means session-derived data feeds analytic systems; ask which data and which systems. "Clinicians remain responsible for all clinical decisions" is true and is also the liability architecture: the platform's systems inform, your license decides, and the MSA is built to keep it that way. That is not a reason to refuse the contract; it is the reason to know precisely what the systems do, because responsibility without visibility is the worst position a licensed professional can occupy.

Employed, Contracted, and Still a Clinician: Your Moves Inside the Machine

The realistic posture is not refusal; tens of thousands of clinicians earn their living on these platforms, and the platforms have built access infrastructure the field needed. The posture is informed residence in the furnished office. Concretely: get your own copy of every agreement you signed, including amendments and updated terms, and store them outside the platform. Send the dated, written question set (the same discipline as the aggregator lesson): What ambient capture or processing applies to my sessions? What is retained, for how long, for what purposes, and is any of it used for model training? What safety-screening systems run over my clients' content, and how are flags routed to me? What outcome data leaves the clinical context, in what form? What notice will I receive before new AI features touch my sessions? Keep the answers with the agreement. Update your own consent conversation: even on a platform with its own client terms, your intake conversation can say, in your voice, what the technology around the care does, and a client who heard it from you is a client whose trust survives the next product announcement. And keep your zone-one duties bright: every note you sign is your attestation, read every word; every risk determination is yours, with platform safety flags treated as routing; and the verifiable details that defend your care, the minutes, the instrument deltas, the modality you actually delivered, go into the record in your words.

The Applied Problem: Your Telehealth Contract-Clause Checklist

Your artifact is the Telehealth Contract-Clause Checklist, one page, twelve items, usable in two modes: before signing a platform contract, or retrospectively against the agreement you already work under. Items one through five, the clause hunt, one per clause family: (1) quote every grant of rights over session content and list the attached purposes, circling "product improvement," "research," and any training language; (2) quote the technology-modification clause and record what notice, if any, precedes new AI capabilities; (3) quote the records clauses: where clinical documentation lives, who owns it, and your access during and after the contract; (4) quote the professional-independence language and note anything that could read as the platform directing care; (5) record the indemnification and insurance allocation for data and AI failures. Items six through nine, the capability inventory, answered from the agreement plus your dated written questions: (6) ambient capture and processing applicable to your sessions, disclosed in writing; (7) transcript and session-artifact retention, in days, with training use and opt-outs; (8) safety-screening systems running over client content and how flags route to you; (9) outcome data flows: which instruments, what leaves the clinical context, in what form (aggregate, de-identified, identified).

Items ten through twelve, your duties, which no platform answer can satisfy for you: (10) your own consent conversation and paperwork describe the platform's technology in plain language, in your voice; (11) your process notes live outside the platform, preserving the 45 CFR 164.501 / 164.508(a)(2) protection; (12) your standing rule, written at the bottom of the page: platform safety flags are routing, not assessment; every risk determination, CSSRS administration, Tarasoff evaluation, and mandated-report decision is made by you and documented in your words, and every note you sign has been read in full.

Run it in one sitting per platform: agreement open, checklist beside it, every item resolved as a quoted clause, a recorded answer, or "NOT FOUND, asked in writing on [date]," exactly the no-blank-cells standard from the aggregator lesson. The verification pass: items one through nine must trace to the agreement text or a dated platform reply in your file; items ten through twelve must trace to your own artifacts (your consent paperwork, your notes location, your written rule). Done looks like one completed page per platform, filed in the decision log beside your vendor matrix, your layer-vs-replace memo, and your panel-agreement checklist. With those four pages, you have done something most of the field has not: you know, in writing, every AI system positioned around your clinical work, what it is authorized to do, and where your judgment remains, by your own documented insistence, the only instrument in the room.

Key Takeaways

• Telehealth platforms are the furnished office: Talkspace, BetterHelp, Tava, Lyra, Spring Health, NOCD, and Iris Telehealth own the video room, the chat, the measures, and often the client relationship, and their AI decisions reach you through contractor MSAs and platform-side BAAs signed before any feature appeared, sometimes consenting on your behalf to ambient capture, transcript retention, and outcome data sharing.
• The three platform species predict the AI posture: consumer messaging platforms (Talkspace, BetterHelp) hold archives where the transcript is the therapy; enterprise platforms (Lyra, Spring Health, Tava) concentrate AI on matching, outcomes, and the population dashboards employers buy; specialty and infrastructure players (NOCD, Iris Telehealth) follow their protocol-driven or embedded models, with Iris clinicians living under two rulebooks at once.
• Ambient capture is architecturally trivial for a platform that owns the pipes, so scrutiny must be contractual: what is captured, who consented and when, and what is retained. Your state recording-consent and board informed-consent duties are personal, and platform terms have never satisfied a board examining what your client understood.
• The statutory bright line patrols this territory directly: under the Illinois WOPR Act and Nevada AB 406, AI cannot provide therapy, and New York's AI companion safeguards law (2025) polices adjacent consumer chatbots; documentation support after the session is one side of the line, anything therapeutically conversing with your client is the other.
• Outcome measurement is clinically legitimate and deserves plain-language client understanding: PHQ-9 and GAD-7 trajectories can feed aggregated employer dashboards, while the clinical reading, especially item nine, remains yours. Platform safety-screening flags are routing, not assessment: AI never scores the CSSRS, never assigns risk, never makes the Tarasoff or mandated-report determination.
• The MSA's five deciding clause families are data and recording rights (circle "product improvement" and "research"), technology modification and notice, documentation ownership and post-contract access, professional independence, and indemnification and insurance, and "clinicians remain responsible for all clinical decisions" is both true and the liability architecture, which is why responsibility demands visibility.
• Your artifact is the twelve-item Telehealth Contract-Clause Checklist (five clause quotes, four capability answers, three personal duties), completed to the no-blank-cells standard and filed as the fourth page of your AI decision log, beside the vendor matrix, the layer-vs-replace memo, and the panel-agreement checklist.