Jordan's practice bought the licenses, signed the BAA, and scheduled a single ninety-minute "AI tool training" delivered by the vendor's customer success rep over Zoom. Six weeks later the audit told the real story: four clinicians had never opened the tool, three were signing AI drafts without reading them, and one had a note in the chart describing an intervention that never happened in the session. The practice did not have a tool problem; it had a training problem, and a vendor demo is not training. Teaching therapists to use AI-assisted documentation is closer to teaching a new clinical skill than installing software: it requires sequenced learning, supervised practice, demonstrated competence, and a sign-off that means something. By the end of this lesson you will be able to design and run the four-session training series (concept, hands-on, verification protocol, ethics and consent) for your clinicians, and build the supervisor's evaluation rubric that makes competency sign-off a real clinical credential rather than an attendance certificate.

Train It Like a Clinical Skill, Not a Software Feature

Here is the controlling analogy for this entire lesson: training clinicians on AI-assisted documentation should be modeled on how the field teaches any clinical skill, the way a practicum teaches suicide risk assessment. Nobody hands a trainee the Columbia protocol PDF and considers risk assessment taught. The skill is built in layers: first the concept and the reasons behind it, then supervised hands-on practice with real-feeling material, then explicit training in the judgment layer (when does this answer change my decision?), then the ethics that wrap the whole thing, and finally a supervisor observes the trainee perform the skill and signs off. Competence is demonstrated, not assumed, and the sign-off has a name attached to it. AI-assisted documentation deserves exactly this structure, because the failure modes are clinical failure modes: an unverified hallucinated intervention in a signed note is a documentation integrity problem with license-level consequences, not a software bug.

The vendor demo fails because it teaches in the wrong direction. It starts from the tool's features and works outward, when clinicians need training that starts from their existing clinical workflow and shows where the tool enters it, where it must never enter it, and where their judgment remains the only authority. The vendor also trains to enthusiasm; you must train to competence. Those are different curricula. Enthusiasm training produces users who love the tool and skip the verification pass. Competence training produces clinicians who can tell you exactly which three error types they check for before signing and can demonstrate catching a seeded error under observation. The second group survives a payer audit; the first group generates the recoupment letter.

One design constraint from the previous lesson carries forward: the room contains technical anxiety, and training is where that anxiety either gets confirmed or dissolved. Every session in the series is built so that no clinician ever performs a new skill in front of the full group before practicing it in private or in pairs. Your nineteen-year LMFT learns alongside a peer, not on stage. Protecting dignity is not a soft consideration; it is the load-bearing wall of clinician AI adoption.

Session One: Concept, the Why Before the How

Session one contains no software. That is deliberate and it will feel strange to whoever schedules it, so defend it: the first ninety minutes establish the conceptual frame that every later behavior hangs on. Cover four things. First, what the tool actually does mechanically: it listens, transcribes, and drafts; it does not understand, diagnose, or decide. Demystifying the mechanism deflates both the hype and the dread. Second, the practice's non-negotiables, stated as clinical rules: the clinician signs the note, and the signature is a legal attestation, not a formatting step; every word gets read before signing; AI never scores a risk instrument, never assigns a risk level, never makes the duty-to-protect or mandated-report call; the clinician's determination comes first and the AI formats after. Third, the data path: where the audio goes, what the BAA covers, what zero-data-retention means, what the consent addendum tells clients, and what happens when a client opts out. Clinicians cannot ethically use a tool whose data path they cannot explain to a client who asks, and clients ask. Fourth, the honest expectations: the first two weeks are slower, the drafts will need editing, and the practice has scheduled for that ramp.

End session one with the artifact clinicians keep: a one-page Concept Card listing the non-negotiables, the data path in plain language, and the answer to the three questions clients most often ask. The card is the reference document for the hallway moment when a client says "wait, an AI listens to our sessions?" and the clinician has eleven seconds to answer well.

Session Two: Hands-On, in Pairs, with Role-Played Sessions

Session two is where the tool finally appears, and the design principle is reps without exposure. Clinicians work in pairs using role-played session material, never real client sessions, because training on live PHI before competence sign-off inverts the entire point. Each pair runs the full cycle three times: confirm consent language out loud, start the capture, conduct a ten-minute role-played session segment, generate the draft, and read it against what actually happened. The role-play scripts matter enormously and you should write them in advance: one routine anxiety presentation, one session containing material the AI typically mishandles (a client quoting another person's threatening statement, which drafts often misattribute), and one session where the role-played client mentions passive suicidal ideation, so clinicians see with their own eyes that the draft does not and must not contain a risk assessment the clinician has not made.

The teaching moment in session two is the gap between the transcriptlike draft and the clinical note. Have each pair mark up a draft with three colors of edit: factual corrections (the draft says "client's brother" when it was the client's father), clinical insertions (the medical-necessity language, the time-in-session minutes for a 90837, the specific PHQ-9 score and delta, the modality actually used, none of which AI can verifiably supply), and deletions (verbatim client material that belongs in the clinician's judgment about what enters the record, not in an automatic transcript dump). By the third rep, clinicians stop experiencing the edit pass as a chore and start experiencing it as the place their clinical authorship lives. That reframe is the session's real product.

Train to competence, not to enthusiasm. Enthusiasm produces users who love the tool and skip the verification pass; competence produces clinicians who can catch a seeded error under observation, and only the second group survives an audit.

Session Three: The Verification Protocol, the Heart of the Series

Session three teaches the single most important skill in the series: the verification protocol, the disciplined read a clinician performs on every AI draft before signing. Teach it as a named, ordered procedure, the way you would teach a mental status exam, because "read it carefully" is not a protocol and does not survive a Thursday with eight clients. The protocol has five passes. Pass one, identity and attribution: right client, right session date, statements attributed to the right speaker, especially in any session with collateral contacts or quoted third parties. Pass two, factual fidelity: every event, symptom, and quote in the draft actually occurred in the session; hallucinated interventions ("therapist conducted a grounding exercise" that never happened) are the signature AI failure and the most dangerous one in a chart. Pass three, omissions: what the draft left out that clinically matters, because absence of risk content in the draft is not evidence of absence in the session; the clinician checks the draft against memory, not memory against the draft. Pass four, clinical insertions: the verifiable details only the clinician can supply, the time-in-session minutes, the measured score and its delta, the named modality, the medical-necessity linkage to the treatment plan. Pass five, risk and judgment content: any risk-related material is rewritten in the clinician's own words reflecting the clinician's own determination, because the risk formulation belongs to the license, never to the model.

Then run the exercise that makes session three famous inside your practice: seeded-error drills. Before the session, prepare AI drafts into which you have deliberately planted errors of each type: a wrong attribution, a fabricated intervention, an omitted disclosure that appeared in the role-play script, a missing 90837 time statement, and a draft that quietly characterizes risk. Clinicians work through the drafts hunting errors, in pairs, scored as a game. The drill does what no lecture can: it calibrates the clinician's trust downward to the correct level and proves to each person that errors are findable in under five minutes when you know the five passes. The seeded-error drill is also the backbone of the competency evaluation in session four's sign-off, so the training and the assessment teach the same skill in the same shape.

Session Four: Ethics and Consent, Wrapping the Skill in the Profession

Session four returns the skill to the profession that owns it. Cover the consent conversation first, as a practiced skill, not a form: each clinician rehearses introducing the consent addendum to a new client and to an existing client, including the answers to the hard questions ("who hears the recording?", "what if I say no?", "can you delete it?"), and the practice's rule that opt-out has zero effect on access to care. Then cover the boundaries with cases: what the clinician does when a session unexpectedly contains a child-abuse disclosure (the mandated-report determination is the clinician's, made and documented in the clinician's own words; the AI never makes the call), what happens when a client in couples work asks for the recording, and how 42 CFR Part 2 material changes the calculus for SUD program records. Close with the documentation of the documentation: how the practice logs consent status, opt-outs, and the clinician's attestation, because in a payer audit or board inquiry the practice must be able to show not just that notes were signed but that the consent and verification infrastructure existed and was followed.

Session four is also where Carmen's situation gets named for any associates in the room: pre-licensed clinicians under supervision must have their AI use known to and approved by their supervisor, written into the supervision agreement, because the supervisor's signature on quarterly hour forms carries the exposure. An associate paying out of pocket for an unsanctioned scribe is not a discipline problem; it is a visibility problem the training series exists to solve, and the practice's sanctioned, BAA-covered tool is the path out of it.

The Supervisor's Competency Rubric: Making Sign-Off Mean Something

Now the assessment layer. After the four sessions, each clinician sits a sign-off observation with a supervisor or designated champion, and the supervisor scores a written rubric. Build the rubric with six competency domains, each rated on a three-level scale: Not Yet Demonstrated, Demonstrated with Prompting, Demonstrated Independently. Domain one, mechanism and data path: the clinician can explain in plain language what the tool does, where the audio goes, what the BAA and retention settings provide, and can answer a simulated client question about it. Domain two, consent practice: the clinician conducts a role-played consent conversation that covers the addendum's content, the opt-out right, and the no-penalty rule, without reading from the form. Domain three, capture workflow: the clinician runs the start-to-draft cycle correctly, including confirming consent status before recording. Domain four, the verification protocol: the clinician performs the five passes aloud on a seeded-error draft and catches at least the planted attribution error, the fabricated intervention, and the missing clinical insertions; this domain is the heart of the rubric and a miss here is an automatic Not Yet. Domain five, risk boundaries: presented with a draft containing risk content, the clinician correctly rewrites it in their own words and articulates that AI never scores instruments, assigns risk levels, or makes duty-to-protect or mandated-report determinations. Domain six, escalation: the clinician knows what to do when something goes wrong, a draft for the wrong client, a client revoking consent mid-treatment, a suspected transcription of a third party who never consented.

The sign-off rule: all six domains at Demonstrated Independently before the clinician uses the tool on real client sessions. Demonstrated with Prompting in any domain triggers a scheduled re-observation within two weeks, with targeted practice in between, exactly the way a supervisor handles a trainee who is almost there on a clinical skill. The rubric is signed and dated by the supervisor and filed, because that signature is the document that answers the malpractice carrier's questionnaire, the payer auditor's process question, and the board's inquiry about whether the practice trained its clinicians before deploying AI in clinical documentation. An attendance sheet proves people sat in chairs; a scored rubric proves the practice taught a skill and verified it landed.

Scheduling, Make-Ups, and Keeping the Skill Alive

Practical architecture, because training that cannot be scheduled does not exist. Run the four sessions weekly over four weeks, ninety minutes each, during paid time; asking clinicians to train on unpaid evenings tells them the practice values the tool more than the people, and the resistance lesson explains where that leads. Cap cohorts at eight to ten so the pairs structure works and nobody hides. Offer each session twice in the week so a clinical emergency does not break a clinician's sequence, and record session one (the concept lecture) for make-up viewing, but never allow video make-up for sessions two and three; hands-on and verification are performance skills and watching is not doing. New hires enter the next available cohort, which is why this series becomes the spine of the onboarding flow two lessons from now.

Sustainment is the part most practices skip and regret. Schedule a refresher seeded-error drill quarterly, ten minutes inside an existing staff meeting, with new planted errors each time; verification skill decays toward rubber-stamping exactly as fast as trust in the tool grows, and the drill is the antidote. Track two signals between refreshers: edit distance (a clinician whose signed notes are byte-identical to drafts has stopped verifying, and the conversation is supportive, not punitive) and time-to-sign (a clinician signing thirty seconds after draft generation is not performing five passes). Both signals route to the clinician's supervisor as a coaching flag, never to performance evaluation, consistent with the surveillance commitments the practice made during the resistance work. The training series is a beginning, not an event; the rubric tells you the skill arrived, and the sustainment loop tells you it stayed.

The Applied Problem: The Four-Session Training Plan and Competency Rubric

Your artifact is a two-part document: the Four-Session Training Plan and the Supervisor's Competency Rubric, ready to hand to whoever will run cohort one. Build the training plan first as a one-page-per-session design: for each session, write the learning objectives (three to five, stated as observable behaviors: "clinician can name the five verification passes in order"), the materials list (Concept Card, three role-play scripts, seeded-error draft set, consent addendum copies), the minute-by-minute run sheet, and the exit check (what each clinician must do before leaving the room, such as completing one full capture-to-edit cycle in session two). Draft it with AI using this prompt: "Design a 90-minute training session run sheet for behavioral health clinicians on [session topic]. Audience: licensed therapists and pre-licensed associates with mixed technical comfort, trained in pairs, never performing new skills in front of the full group. Include learning objectives as observable behaviors, a materials list, a minute-by-minute agenda, and an exit check. Constraints: the clinician signs every note after reading every word; AI never scores risk instruments or makes risk determinations; consent addendum governs all capture." Run it four times, once per session topic.

Then build the rubric as a single scoring sheet: six domains down the left (mechanism and data path, consent practice, capture workflow, verification protocol, risk boundaries, escalation), the three-level scale across the top, a behavioral anchor sentence in each cell describing exactly what Demonstrated Independently looks like for that domain, and a signature and date block for the supervisor at the bottom with the sign-off rule printed above it: all six domains independent before real client use; any Demonstrated with Prompting triggers re-observation within two weeks.

Now the verification pass on your own artifact. Walk the rubric against the training plan and confirm every rubric domain is actually taught and practiced in a specific session; a rubric that assesses what the training never taught is an ambush, and clinicians will read it as one. Walk the seeded-error draft set and confirm it contains at least one planted error per verification pass, including one fabricated intervention and one risk-characterization error. Check every claim in the Concept Card against the practice's real BAA, retention configuration, and consent addendum. Strip any objective that is not observable.

"Done" looks like this: four run sheets a substitute facilitator could deliver without you in the room, a rubric a supervisor can score in a twenty-minute observation, a seeded-error set with an answer key, and a pilot read-through with one champion and one skeptic who both confirm the plan teaches the skill they would want their own supervisee to have before touching a real chart.

Key Takeaways

• Train AI-assisted documentation the way the field trains any clinical skill, like suicide risk assessment in practicum: concept, supervised practice, judgment, ethics, then observed sign-off. A vendor demo trains to enthusiasm; you must train to competence, and only competence survives an audit.
• The four-session series runs concept (no software, non-negotiables and data path), hands-on (pairs, role-played sessions, three-color edit markup), verification protocol (the five named passes), and ethics and consent (practiced conversations, mandated-report and Part 2 boundaries, supervision visibility for associates).
• The five verification passes are identity and attribution, factual fidelity, omissions, clinical insertions, and risk and judgment content. The clinician checks the draft against memory, never memory against the draft, and AI cannot supply the verifiable details: time-in-session minutes, the measured score and delta, the modality used.
• Seeded-error drills are the engine of both training and assessment: deliberately planted attribution errors, fabricated interventions, omissions, and risk characterizations calibrate trust downward to the correct level and prove errors are findable in minutes.
• The supervisor's rubric scores six domains (mechanism and data path, consent practice, capture workflow, verification protocol, risk boundaries, escalation) on a three-level scale; all six must be Demonstrated Independently before real client use, and a verification-domain miss is an automatic Not Yet.
• The signed, dated rubric is a legal artifact: it answers the malpractice carrier's questionnaire, the payer auditor, and the board. Attendance proves chairs were filled; a scored rubric proves a skill was taught and verified.
• Sustainment is mandatory: quarterly seeded-error refreshers, plus edit-distance and time-to-sign monitoring routed to supervisors as coaching flags only, never to performance evaluation, honoring the surveillance commitments made during the resistance work.