Maria's EHR sent her an email this morning: "Introducing AI-powered notes, built right into your workflow. Try it free." She already pays for a standalone scribe. Now her system of record, the place where her notes legally live, is offering to write them too. Should she run both? Cancel the scribe? Ignore the EHR's new button? This is the layer-versus-replace decision, and almost every clinician on SimplePractice, TherapyNotes, or a Therapy Brands product will face it this year, usually at 9:54 PM with seven notes left and no time to think it through. By the end of this lesson you will be able to explain how EHR-native AI in SimplePractice (with its Sidekick tooling and embedded Blueprint Health), TherapyNotes AI, and Therapy Brands differs structurally from a bolt-on scribe, why note version history becomes the quiet star witness in a board audit, and how to write a one-page layer-versus-replace decision memo for your own practice before the free trial decides for you.

The House and the Contractor: One Analogy for the Whole Lesson

Carry this analogy through the lesson: your EHR is the house your clinical records live in, and an AI scribe is a contractor you hire to work on that house. A bolt-on scribe is an outside contractor: skilled, specialized, but they work in their own truck (their own servers, their own retention rules, their own BAA) and hand you the finished cabinet to install yourself. EHR-native AI is the builder offering to do the work in-house: everything happens inside the structure you already own, under the roof of the agreement you already signed. The in-house option means fewer handoffs and one set of keys. It also means that if the builder cuts a corner, the defect is inside your house, in your walls, in your system of record, and there is no second party to point to.

That is the real shape of the layer-versus-replace question. Layering means keeping a standalone scribe (Mentalyc, Upheal, Twofold, Heidi, Eleos in an enterprise setting) alongside your EHR and moving drafts across the boundary between them. Replacing means letting the EHR's own AI do the drafting inside the chart. Neither is categorically right. The decision turns on three things this lesson walks through slowly: where the handoffs are (every handoff is an error opportunity and a compliance boundary), what the EHR's AI actually is under the hood (usually a third-party model wearing the EHR's interface), and what each architecture writes into note version history, the audit trail that decides arguments you have not had yet.

SimplePractice: The House With Built-In Helpers

SimplePractice is, by adoption, the default house for solo and small-practice clinicians, the converted-dining-room demographic Maria belongs to. Its AI story has two distinct parts that clinicians routinely conflate. The first is its note-drafting and practice-assistant tooling, the Sidekick layer, which lives inside the platform and touches scheduling, documentation, and administrative work. The second is embedded Blueprint Health, which is not a scribe at all: Blueprint is measurement-based care, the PHQ-9, GAD-7, and symptom-tracking layer that turns the questionnaire your client completes in the waiting room into a scored trajectory inside the chart. Conflating the two matters because they carry different risk profiles. A drafting assistant generates clinical language you must verify word by word before signing. A measurement layer moves validated instrument scores, where the risk is less hallucination and more workflow: an unimported PHQ-9, like the one sitting in Maria's queue, is a medical-necessity data point your 90837 note needed and did not get.

The credentialing questions from the last lesson do not relax because the AI lives inside the house. They relocate. With a bolt-on, you ask the scribe vendor for the BAA, the subprocessor list, and retention in days. With EHR-native AI, you ask your EHR the same questions, and the subprocessor question gets sharper, not softer: which model provider actually generates the text behind the Sidekick interface, is that provider under a BAA at the tier you pay for, and is your clients' data used to train or improve models? Jordan's discovery applies with full force here: the EHR vendor said yes to AI-assisted scoring, and the subprocessor list included a model provider that does not sign a BAA at Jordan's tier. The interface was native. The model was not. The house had quietly subcontracted the work to someone who never signed your agreement.

One more SimplePractice-specific note: because the platform hosts your record of care, anything its AI drafts is born inside the legal record's neighborhood. That is convenient and consequential at once. Convenient, because there is no paste-back step. Consequential, because a draft that was never reviewed now exists inside the system a board or payer will subpoena, and you will want the version history to show exactly what happened between draft and signature. Hold that thought; it gets its own section.

TherapyNotes AI: The Builder Who Drafts in Your Templates

TherapyNotes ships its own AI inside an EHR that was already opinionated about behavioral health documentation. TherapyNotes' templates have long enforced structure: psychotherapy notes with defined fields, treatment plans with goals and objectives, termination summaries with required elements. TherapyNotes AI drafts into that structure. For a clinician, this is the strongest practical argument for the native option: the AI is filling in the exact fields your note already requires, in the format your payer reviewer and your board already expect, rather than producing freestanding prose you reformat by hand. Structured drafting reduces a specific, mundane error class: the field left blank, the plan section missing, the time-in-session line that a 90837 audit lives or dies on never making it from the scribe's output into the EHR's field.

But structure is not verification. A template with every field filled looks finished, and "looks finished" is precisely the state in which a tired clinician signs without reading. The cardinal rule does not bend for native tooling: the clinician signs the note, and the signature is a legal attestation, not a formatting step. Read every word. The native AI can fill the Mental Status Exam field with plausible boilerplate; it cannot have observed your client's affect. It can suggest a CPT code; it cannot know the session ran 56 minutes unless you told it. It can populate a risk field with language; it must never be the source of the risk determination. AI never scores the CSSRS, never assigns a risk level, never makes the Tarasoff or mandated-report call. The clinician decides; the AI formats after the decision. If a native template auto-fills a risk field, your job is to treat that field as empty until you have written it yourself.

Ask TherapyNotes the same relocated credentialing questions: which model generates the drafts, what is retained beyond the signed note, what happens to the audio or dictation source, and whether drafts that were never signed persist somewhere discoverable. An EHR that answers in writing earns the native option a fair hearing in your memo. An EHR that answers with a compliance badge earns the same score a bolt-on vendor would: marketing language, score 1.

Therapy Brands: Native AI for the Medicaid-Heavy Clinic

Therapy Brands is not one product but a family (NextStep, ProcedureFlow, TheraNest among them) serving a different segment: Medicaid-heavy clinics, community behavioral health, SUD programs, and group practices whose documentation lives under state Medicaid manuals rather than commercial payer rules. Therapy Brands has rolled native AI into this EHR layer, and the segment changes what the stakes are. Medicaid documentation is unforgiving in specific ways: service start and stop times, location, modality, credential-aligned signatures, and in many programs, daily notes. A native AI that drafts into those required structures can be a real relief for clinicians writing multiple notes a day under productivity requirements. A native AI that drafts plausible but unverified content into those structures is manufacturing audit findings at scale, because a Medicaid Recovery Audit Contractor does not read notes for tone; it reads them for required elements that are present, consistent, and true.

The Medicaid-heavy segment also makes the 42 CFR Part 2 question unavoidable rather than diagnostic. Many Therapy Brands clinics hold themselves out as providing SUD treatment, which puts their records squarely under Part 2 as revised by the 2024 final rule. The relocated credentialing question becomes concrete: when the EHR's native AI processes a Part 2 client's session content, where does that content go, which subprocessors touch it, and how does the platform honor redisclosure prohibitions when the AI-generated note is later released under a consent? If your clinic's compliance officer cannot answer those questions about the EHR's own AI, the native feature is not safer than a bolt-on; it is an unexamined bolt-on wearing the house's paint.

EHR-native AI does not remove the vendor questions; it relocates them. The model behind the native interface is usually a subcontractor your agreement never named, and the version history is the witness that will testify about what you actually reviewed.

Note Version History: The Quiet Star Witness

Here is the part of this lesson that will matter most in three years, and it is the part no vendor demo covers. Every EHR keeps some form of note version history: when the note was created, what was changed, when it was signed, and sometimes who or what produced each revision. In ordinary practice nobody looks at it. In a board audit, a payer audit, or a malpractice deposition, it becomes the star witness, because it answers the question your attestation depends on: did the clinician actually review this note before signing it?

Walk the two architectures through that lens. With a bolt-on scribe and copy-paste, the drafting history lives at the scribe vendor; your EHR's history shows a finished block of text arriving at 10:02 PM and a signature at 10:03 PM. That one-minute gap, multiplied across a caseload, is a pattern an auditor can read as rubber-stamping, fairly or not, and the exculpatory drafting history sits on a vendor's servers under a retention policy you scored in the last lesson. With native AI, the draft, the edits, and the signature can all live in one timeline inside the EHR. If you actually review your notes, native version history is your best defense: it shows the AI draft at 9:58, your three edits at 10:04, the corrected time-in-session at 10:06, the signature at 10:07. If you do not review your notes, the same timeline is the prosecution's exhibit. Version history does not make you compliant; it makes you legible. It records whichever clinician you actually are.

Three practical moves follow. First, learn what your EHR's version history actually captures before you adopt its AI; ask the vendor to show you the audit trail for an AI-drafted, clinician-edited, signed note. Second, make your review visible in the record: real edits, corrections, the verifiable details only you can supply (the 56 minutes in session, the PHQ-9 delta from 18 to 11, the modality you actually used, CBT with a thought record, not generic "processed feelings"). An AI cannot supply those because it does not know them; their presence in your final note is the fingerprint of a real review. Third, if you layer a bolt-on, keep your own lightweight evidence of review, because the EHR timeline alone will not show it.

Layer or Replace: The Actual Decision Logic

Now the decision itself, taken slowly. Layering (bolt-on scribe plus EHR) wins when the standalone tool is meaningfully better at the clinical drafting your caseload needs: a therapy-native scribe tuned for 53-minute sessions, your modality, your note format, with a retention policy and BAA you have already verified and filed. Layering costs you a handoff: every copy-paste is an error opportunity (Client A's note in Client B's chart), a workflow tax, and a split audit trail across two vendors. Layering also means two BAAs, two subprocessor lists, two retention policies to track, double the surface your compliance file has to cover.

Replacing (EHR-native AI only) wins when the native tool drafts into your required structures well enough, when the unified version history matters to you (it should), and when reducing vendor count is itself a compliance goal, as it is for Jordan, who currently has twelve clinicians on an unauthorized free tool and needs to consolidate, not multiply. Replacing costs you choice: you inherit whatever model the EHR subcontracts to, on the EHR's timeline, and if the native drafting quality is mediocre for your modality, you will quietly drift back to workarounds, which is how shadow tools are born. The worst outcome is not layering or replacing; it is the undecided middle, where clinicians use both inconsistently, paste between systems ad hoc, and nobody can reconstruct which tool drafted which note. That middle is where Jordan's practice lives today, and it is the thing your memo exists to end.

The honest tiebreakers, in order: documentation quality for your actual caseload (test both on your hardest note type, not the vendor's anxiety client); the relocated credentialing answers (BAA at your tier, model subprocessor coverage, retention in days, Part 2 if it applies); version history legibility; and only then price. Price last, because Carmen's math from the last lesson cuts both ways: $59 a month is cheap against two unpaid hours a night, and any tool that fails the credentialing floor is expensive at free.

Running the Test Without Betting a Client on It

Whichever way you lean, do not run the comparison on live client data first. Build one synthetic session: a fabricated but clinically realistic 53-minute case in your modality, with the elements vendor demos avoid, a passing reference to a custody stressor, a substance-use mention, a mood instrument score, and a session time of 56 minutes. Run it through the EHR-native AI and through your bolt-on candidate. Then grade both drafts against the same five-line rubric: Did it preserve the time-in-session detail? Did it keep the instrument score exact or drift it? Did it stay out of risk determination, or did it generate risk language you never dictated? Did it invent anything (a quote, a homework assignment, an intervention)? Does the output land in the fields your payer and board require?

This synthetic test is also where you discover the difference between a drafting model and a template engine. Some native AI is excellent at structure and bland at clinical specificity; some bolt-ons are vivid at narrative and careless with required fields. Your caseload decides which failure is cheaper to correct. A psychiatrist on Therapy Brands' stack doing med management wants dense structure; a trauma therapist on SimplePractice doing EMDR wants narrative fidelity; a Medicaid clinic needs the start-stop times and credential lines present every single time. There is no best answer in the abstract, which is why the artifact for this lesson is a memo, not a recommendation.

The Applied Problem: Your Layer-vs-Replace Decision Memo

Your artifact is a one-page Layer-vs-Replace Decision Memo, written for your own practice (or, if you are an associate like Carmen, written as a recommendation to your supervisor, which is itself good supervision hygiene). Structure it in five short sections. Section one, Current state: name your EHR (SimplePractice, TherapyNotes, a Therapy Brands product, or other), any AI tools currently in use, authorized or not, and who is using them. Honesty here is the point; Jordan's memo would begin "twelve clinicians currently use an unauthorized free scribe." Section two, The native option: what AI your EHR offers today, and the relocated credentialing answers in writing: which model subprocessor drafts the text, BAA coverage at your tier, retention in days for drafts and source audio or dictation, training opt-out, and, if you serve SUD clients, the Part 2 handling under the 2024 final rule.

Section three, The synthetic test: run the one fabricated session through both candidates and record the five-line rubric results (time-in-session preserved, instrument score exact, no risk language generated, nothing invented, required fields populated). Paste the actual rubric scores into the memo. Section four, Version history: one paragraph stating what your EHR's audit trail shows for an AI-drafted, edited, signed note, and how you will make your review visible in the record (the verifiable details only you can supply: minutes in session, the specific PHQ-9 delta, the named modality). Section five, Decision and rule: one sentence choosing layer, replace, or staged pilot, and one bright-line rule for the practice, for example: "All AI drafting occurs in [named tool] only; no other AI touches client information; every signed note must contain at least two clinician-supplied verifiable details."

The verification pass: every factual claim in sections two and four must trace to a document or a screenshot in your file, not a sales call. The memo is done when a colleague who has never met you could read it and know your EHR, your tool decision, why you made it, and how an auditor would see your review in the record. Done also means dated and signed, because this memo is the first entry in your practice's AI decision log, the paper trail that turns "we sort of adopted some tools" into "we made a documented decision on June 3, 2026, and here is the reasoning." That sentence is worth more to your malpractice carrier than any vendor badge.

Key Takeaways

• The layer-versus-replace decision is the EHR-era version of vendor selection: layering keeps a standalone scribe beside SimplePractice, TherapyNotes, or Therapy Brands and accepts handoffs; replacing uses the EHR's native AI and accepts its subcontracted model. The worst position is the undecided middle, where nobody can reconstruct which tool drafted which note.
• EHR-native AI relocates the credentialing questions instead of retiring them: the native interface usually fronts a third-party model, so you still demand the BAA at your tier, the subprocessor list, retention in days, and training opt-out, this time from your EHR. Jordan's EHR said yes to AI while its subprocessor list included a model provider with no BAA at Jordan's tier.
• SimplePractice's AI story has two parts with different risk profiles: drafting and assistant tooling (Sidekick), which generates clinical language you must verify word by word, and embedded Blueprint Health, a measurement-based care layer whose risk is workflow fidelity, like the unimported PHQ-9 your 90837 medical-necessity argument needed.
• TherapyNotes AI drafts into enforced templates, which reduces blank-field errors but creates the "looks finished" trap; Therapy Brands brings native AI to Medicaid-heavy and SUD settings where start-stop times, credential-aligned signatures, and 42 CFR Part 2 (2024 final rule) handling are non-negotiable required elements, not style preferences.
• Note version history is the quiet star witness in any board audit or deposition: native AI can put draft, edits, and signature on one legible timeline, which defends a clinician who really reviews and convicts one who rubber-stamps. Make your review visible with the verifiable details AI cannot supply: minutes in session, the exact PHQ-9 delta, the modality actually used.
• The hard lines hold inside the EHR exactly as outside it: AI never scores the CSSRS, never assigns a risk level, never makes the Tarasoff or mandated-report determination, and a template that auto-fills a risk field should be treated as empty until the clinician writes it. The signature remains a legal attestation: read every word before signing.
• Your artifact is the one-page Layer-vs-Replace Decision Memo: current state stated without spin, the native option's credentialing answers in writing, a synthetic-session test scored on a five-line rubric, a version-history paragraph, and one dated, signed decision with a bright-line rule. It is the first entry in your practice's AI decision log.