Jordan's compliance officer asks the question that stops the Tuesday leadership meeting cold: "If an AI-drafted note turns out to be wrong, at what exact moment did it become our problem instead of the vendor's?" Twelve of Jordan's twenty-five clinicians use an AI scribe daily, and not one of them can answer. They know the model drafts and they sign, but nobody has defined what the AI is allowed to produce at each step, what the clinician must verify before signing, or what happens when a draft mentions something the AI is not allowed to assess, like a passing reference to "wanting to disappear" buried in paragraph three. That undefined space is where board complaints, payer recoupments, and malpractice claims are born. By the end of this lesson you will design the human-AI handoff for every AI-touched step in your workflow: a written definition of exactly what the AI produces, exactly what the clinician verifies, the timestamped point where the clinician's signature transfers liability from draft to record, and an explicit escalation hand-back rule that yanks any output mentioning a risk indicator out of the automated lane and back into clinical judgment. The artifact is a Handoff Map you can show a supervisor, an auditor, or a malpractice carrier without flinching.

The Tower and the Cockpit: Why Handoffs Are Designed, Not Assumed

Aviation solved this problem decades ago, and the solution is the controlling idea of this lesson. When an aircraft moves from one controller's airspace to another's, nobody relies on vibes. There is a positive handoff: the transferring controller states what is being handed over, the receiving controller explicitly accepts it, and from a defined moment, recorded and timestamped, responsibility belongs to the receiver. Until acceptance, the aircraft is still the transferring controller's problem. After acceptance, no one gets to say "I assumed the other tower was watching." The entire architecture exists because the most dangerous place for an aircraft is the gap between two people who each believe the other one is in control.

Your AI workflow has exactly that gap, and most practices are flying through it daily without noticing. The scribe produces a draft note; the clinician skims it at 9:54 PM; the note gets signed; and if a hallucinated intervention or an unflagged risk mention surfaces later, the reconstruction begins: did the clinician verify that line? Was the tool supposed to catch it? Did anyone define whose job it was? In the cockpit-and-tower frame, the AI is an instrument, never a controller. It can compute, draft, format, and display, the way an autopilot holds a heading. But an instrument is never responsible for the flight. The clinician is the pilot in command at every moment, and the handoff you are designing is not between AI and human as peers; it is the defined moment when the pilot stops treating the output as instrument data and starts attesting to it as the official record.

This lesson sits on the two before it. You have already mapped your week as twelve to eighteen discrete tasks and culled the steps that fail the three-question ethics test (Does the client know? Does the BAA cover it? Does my ethics code permit it?). What remains is a list of AI-eligible steps: drafting progress notes from session content, summarizing prior notes for pre-session prep, structuring intake information, drafting coordination letters, formatting treatment plan reviews. For each surviving step, you now write the handoff in three parts: what the AI produces, what the clinician verifies, and where the signature lands. Steps without a written handoff go back in the cull pile until they get one. That is the whole discipline, and it is less work than it sounds: a practice with eight AI-touched steps needs eight rows on one page.

Column One: What the AI Produces, Stated Narrowly

The first column of every handoff defines the AI's output as narrowly as a work order. Not "the AI helps with notes," which is how vendors talk, but "the AI produces a draft DAP note from the session transcript, using only facts present in the transcript, in the practice's template, with all required fields present or flagged as missing." The narrowness is the safety feature. A broad mandate invites the model to fill gaps with plausible inference, and plausible inference is the polite name for hallucination. A narrow mandate gives the clinician a checkable contract: anything in the output that the mandate did not authorize is, by definition, a defect to be caught in column two.

Writing this column forces decisions practices usually dodge. Does the AI draft the assessment section, or only objective and plan? Does it propose CPT codes, or is coding a clinician-only field? Does it summarize measurement scores the clinician entered, or is it forbidden to mention scores it cannot see in the source? The answers differ legitimately across practices, but they cannot differ across clinicians within a practice without creating exactly the inconsistency an auditor loves. Jordan's twelve scribe users currently run twelve private versions of this column, which means the practice has no version of it. The Handoff Map collapses that into one.

Two boundaries are not practice choices; they are program-wide constants. First, the AI never supplies the verifiable details only the clinician possesses: time-in-session minutes for a 90837, the PHQ-9 delta from the chart, the modality actually used in the room. Those enter through the clinician, or they do not enter. Second, the AI never produces clinical determinations: no risk levels, no CSSRS scoring, no diagnosis assignments, no duty-to-protect calls, no mandated-report judgments. The output column for every step should be describable in one sentence that contains the words "draft" or "summary" or "format" and never the words "assess," "determine," or "decide." If you cannot write the sentence that way, the step does not belong in the AI lane.

Column Two: What the Clinician Verifies, Stated as Actions

The second column converts "review the draft" into named, executable checks, because "review" is the word practices hide behind and auditors see through. A real verification spec for a draft progress note reads like a preflight checklist: confirm every factual claim against memory of the session and the source material; confirm no invented quotes, interventions, or client statements; enter or confirm the verifiable details (time in session, scores, modality named in session); confirm the note supports the CPT code billed; scan the full text for any risk-relevant content (more on that in a moment); and only then sign. Each item is binary, performable in seconds once habitual, and, critically, describable after the fact: a clinician who follows this column can testify to what verification meant, instead of gesturing at "I read it over."

Calibrate the column to the stakes of the step. A pre-session summary of your own prior notes carries low external risk; its verification might be two checks (no content from a different client, no invented history) because the document never leaves your prep folder and you will be in the room correcting it in real time. A discharge summary leaving the practice for a PCP carries high stakes; its column is long. The handoff map is allowed to be unequal across rows, and should be. What it is not allowed to do is leave a row blank, because a blank verification column is a written confession that for this step, nobody checks.

An unsigned AI draft is a suggestion. A signed note is testimony. The handoff map exists to make sure no one in your practice confuses the moment one becomes the other.

Column two is also where supervision plugs in. For a pre-licensed clinician, the verification spec includes the supervisor's layer: which AI-touched documents the supervisee may sign after self-verification, which require supervisor co-signature, and how AI use appears in the supervision log. Carmen's situation in Fresno, paying for her own scribe with a supervision agreement that never mentions AI, is precisely a missing column two: nobody defined what she verifies, what her supervisor verifies, and whose signature carries which weight. A supervisor who co-signs an AI-drafted note has accepted the handoff too, and the map should say so in writing before the quarterly BBS form makes it awkward.

The Signature Line: The Timestamped Point Where Liability Transfers

The third element is the one Jordan's compliance officer was really asking about: the defined, timestamped moment when responsibility for the content stops being ambiguous and becomes wholly the clinician's. That moment is the signature. Before it, the document is a draft: an instrument reading, vendor work product, an unaccepted handoff that no payer should be billed from and no court should treat as the record. After it, the document is the legal medical record, and the signature is an attestation, the same legal act it has always been, that a licensed professional reviewed the content and adopts it as accurate. The EHR timestamp on that signature is your positive-handoff record: it marks, to the minute, when the pilot accepted the aircraft.

Three operational rules follow, and they belong on the map verbatim. Rule one: nothing downstream happens from an unsigned draft. No billing submission, no claim, no coordination letter sent, no record release. If a biller can pull from drafts, the handoff has a hole in it that a recoupment letter will eventually find. Rule two: the verification in column two happens before the signature, every time, with no batch-signing of unread drafts. Batch-signing is the practice-level equivalent of accepting twenty aircraft at once without looking at the radar, and "I sign them all on Friday" is a sentence that should make a supervisor's hands cold. Rule three: the signature timestamp must sit inside your documentation window, signed before midnight or within the payer-dependent window (some payers run 72 hours), because a handoff accepted late is a handoff a concurrent reviewer can question.

Understand what the signature does to the liability question. It does not create your liability; you were always responsible for your records. What it does is collapse the ambiguity. Before a defined handoff, a bad note spawns a three-way argument among clinician, practice, and vendor about who should have caught it, and the clinician usually loses that argument anyway because the license is the deepest anchor in the room. After a defined handoff, the answer is documented: the vendor's product responsibility runs up to the draft (and lives in the BAA and the service agreement); the clinician's professional responsibility begins at the verification pass and is sealed at the timestamped signature. You are not taking on more risk by writing this down. You are converting an ambient, unbounded risk into a bounded, dated, defensible one.

The Escalation Hand-Back: When the Output Mentions What the AI Cannot Assess

Now the rule that earns this lesson its place in the program. Sometimes the AI's output will mention a risk indicator: the draft note includes the client's passing "I just want to disappear sometimes," the intake summary surfaces a reference to an old attempt, the transcript-derived draft contains a sentence about the client's partner "going through my phone and showing up at my work." The AI is not allowed to assess any of it. It cannot score a CSSRS, assign a risk level, judge IPV danger, make the Tarasoff-type duty-to-protect determination (in California, duty to protect under Civ Code §43.92), or decide whether WIC §11166 mandated reporting is triggered. Those are clinical and legal judgments that belong to a licensed human, full stop. But the AI's output can contain the raw material of those judgments, and the question your workflow must answer in advance is: what happens at that moment?

The escalation hand-back rule is the answer, and in the aviation frame it is the instrument's stall warning: the system does not fly the recovery, it surrenders control loudly. The rule, written for the map: "Any AI output that mentions a risk indicator (suicidal or self-harm ideation, harm to others, abuse or neglect of a child, elder, or dependent adult, IPV indicators, acute substance-related danger, psychosis-driven safety concerns) exits the automated lane immediately. The draft is not signed, not billed from, and not transmitted. The clinician reviews the source material directly, performs whatever clinical assessment the content warrants using clinical judgment and the practice's risk protocols, documents the assessment and disposition personally, and only afterward may AI assistance resume, limited to formatting the clinician's completed determination." Hand the controls back to the pilot, and do not let the autopilot re-engage until the pilot has flown the maneuver.

Notice the rule's three design choices. It triggers on mention, not on the model's opinion of severity, because asking the AI to decide which mentions are serious is asking it to assess risk through the back door. It interrupts the entire downstream chain, not just the note, because a coordination letter or billing claim built on an unassessed risk mention is the same defect wearing different clothes. And it defines re-entry: the workflow resumes only after the clinician's determination exists, at which point the AI may format and structure that determination, never alter or grade it. If your scribe vendor offers a feature that flags risk language, treat the flag as a smoke detector, useful, never sufficient: the absence of a flag clears nothing, and column two's full-text risk scan remains the clinician's job on every draft, every time.

Building the Map, Row by Row: A Worked Example

Here is one complete row, worked the way you will write all of yours. Step: post-session progress note for an individual 90834 session. AI produces: "Draft note in practice DAP template from the clinician's session shorthand, using only facts present in the shorthand; missing required fields flagged, not filled; no codes, no scores, no risk language generated; assessment section left as bullet points for clinician authorship." Clinician verifies: "Read full text; confirm every fact against the session; confirm no invented quotes or interventions; enter time-in-session minutes and confirm code fit; enter the GAD-7 score from the chart; full-text scan for risk-relevant content; if any risk mention, invoke hand-back rule before anything else." Signature point: "Clinician e-signature in the EHR, same day, before midnight; timestamp is the liability transfer; no billing from unsigned drafts." Hand-back trigger: "Any mention of SI, HI, abuse, IPV, or acute danger in draft or shorthand." That is one row: four cells, maybe a hundred and forty words, and it answers the compliance officer's question for this step completely.

Now run the variation exercise, because the map earns its keep on the steps that differ. The pre-session prep summary row has a thin verification column but the same hand-back trigger, because a summary of last week's note can surface last week's risk content and the clinician walking into the room needs it handled as clinical signal, not as recycled text. The intake-packet structuring row adds a verification line about collateral and third-party information. The coordination-letter row adds scope verification against the signed release and the minimum-necessary standard, plus a harder signature rule: nothing transmits until signed, and the send itself is logged. The treatment plan review row adds the requirement that trajectory claims trace to actual dated scores. Different cargo, same tower discipline: every row ends in a named verification, a timestamped signature, and a hand-back trigger.

Resist two temptations as you draft. The first is writing aspirational columns nobody will execute at 9:54 PM; a verification spec with nineteen items per note is a spec that will be skipped in whole, which is worse than six items performed always. Write the column you will actually run on your worst night, then run it. The second temptation is exempting "low-risk" steps from the hand-back rule. There is no AI-touched step in a behavioral health practice whose source material cannot contain a risk mention; the rule costs nothing on the days it never fires, and it is the only part of the map that matters on the day it does.

The Map as Governance: Supervisors, Carriers, and Auditors Read It Too

A finished Handoff Map does quiet work far beyond the clinical lane. For Jordan, it is the spine of the written AI policy the malpractice carrier's renewal questionnaire is asking about: a carrier that asks "do you use AI?" is really asking "do you control it?", and a one-page map with defined outputs, verification specs, signature rules, and a hand-back protocol is the most persuasive possible answer. For a payer audit or a concurrent review, the map plus signature timestamps demonstrates that notes were verified and adopted by a licensed clinician inside the documentation window, which defuses the insinuation that the practice bills from machine output. For a board inquiry, it shows the standard of care was designed, not improvised. None of these audiences requires the map to be elaborate. All of them notice when it does not exist.

For supervisors, the map is the missing annex to the supervision agreement. Carmen's supervisor can attach a supervisee-specific version: which steps Carmen may use AI on at all, which verification items are hers and which are reviewed in supervision, which documents require co-signature before they count as accepted, and the explicit statement that the hand-back rule applies to supervisees with one addition, that any triggered hand-back is brought to supervision. That single page converts "we need to talk about that" into a working agreement, protects the supervisor's signature on the quarterly form, and gives Carmen what she has wanted all along: someone finally telling her what the right thing is, in writing.

And for you, solo or not, the map is how the workflow survives your own fatigue. The whole reason handoff design exists in aviation is that competent professionals under load make assumption errors; the procedure carries them when attention cannot. Maria at 9:54 PM with seven notes left does not need to remember her philosophy of AI oversight. She needs four cells per row: what the draft is allowed to be, what she checks, where she signs, and what stops everything. Design it once, at a desk, in daylight. Execute it nightly, tired, safely.

The Applied Problem: Build Your Handoff Map

The deliverable is a one-page Handoff Map covering every AI-touched step that survived your ethics cull, built as a five-column table: Step, AI Produces, Clinician Verifies, Signature Point (timestamped liability transfer), and Hand-Back Trigger. Start by listing your surviving steps down the left side; most practices land between five and nine rows (post-session note, pre-session prep summary, intake structuring, treatment plan review, discharge summary, coordination letter, psychoeducation materials, between-session message drafts). If a step you use daily is not on the list, it either failed the ethics test and should not be AI-touched, or you forgot it, and the map just did its first job.

Fill column two using the one-sentence discipline: each AI output described with "draft," "summary," or "format," never "assess," "determine," or "decide," and always including "using only facts present in the source" and "missing fields flagged, not filled." Fill column three as named binary checks calibrated to the stakes of each row, always ending with the full-text risk scan, and always including the verifiable details only you can supply: time-in-session minutes, the actual score deltas from the chart, the modality named in session. Fill column four with the signature rule for that document type: who signs, by when (same day before midnight, or your payer's window), and the standing line "no downstream action from unsigned drafts." Then write the hand-back rule once, in full, across the bottom of the page: the trigger list (SI, HI, abuse or neglect, IPV indicators, acute danger), the immediate exit from the automated lane, the clinician's direct review and personal documentation of the assessment and disposition, and AI re-entry limited to formatting the completed determination.

Now pressure-test the map with two drills. Drill one: take a real recent draft from your scribe (or generate a realistic test case) and execute its row exactly as written, timing yourself; if the verification column takes longer than the note used to take, the column is aspirational and needs cutting to what you will run on your worst night. Drill two: plant a risk mention, a single sentence like "client mentioned sometimes wishing she would not wake up" inside a test draft, and walk the hand-back: stop, no signature, review the source, document the clinical assessment yourself as the clinician, and only then let AI format. If anything in your tooling makes the stop awkward, a biller who can see drafts, an auto-sign setting, a vendor flag you have been treating as the assessment, fix that before the map goes live.

Done looks like one page: every AI-touched step in a row, every row with four completed cells, the hand-back rule written in full at the bottom, a version date, and your signature, because the map itself deserves one. File copies with your AI policy and, if you supervise or are supervised, attach the supervisee version to the supervision agreement. From now on, when anyone asks Jordan's compliance officer's question, the exact moment a draft becomes your problem, you point to column four: at the timestamped signature, after column three, never before, and never at all if column five fired first.

Key Takeaways

• Handoffs are designed, not assumed. The most dangerous place in an AI workflow is the gap between a model that drafted and a clinician who believes someone else checked; aviation closes that gap with positive handoff, and your practice closes it with a written Handoff Map.
• The AI is an instrument, never a controller. Define each step's output narrowly, in one sentence using "draft," "summary," or "format," never "assess," "determine," or "decide," with only source facts allowed and missing fields flagged rather than filled.
• Verification is a named checklist, not a vibe. Each row lists binary checks calibrated to the document's stakes, always including the verifiable details only the clinician can supply (time-in-session minutes, real score deltas, modality used) and always ending with a full-text risk scan.
• The timestamped signature is the liability transfer point. Before it, the document is an unaccepted draft from which nothing may be billed, sent, or released; after it, it is the legal record the clinician has adopted. Verification precedes signature every time, with no batch-signing, inside the documentation window.
• The escalation hand-back rule is non-negotiable: any output mentioning a risk indicator (SI, HI, abuse or neglect, IPV, acute danger) exits the automated lane on mention, not on the model's severity opinion. The clinician assesses and documents personally; AI never scores risk, and it re-enters only to format the clinician's completed determination.
• The map is also governance: it answers the malpractice carrier's questionnaire, shows auditors that signed notes were verified inside the window, gives supervisors a written annex defining supervisee AI use and co-signature rules, and carries the tired clinician at 9:54 PM when attention cannot.
• Build the artifact as a five-column, one-page table (Step, AI Produces, Clinician Verifies, Signature Point, Hand-Back Trigger), pressure-test it with a timed real draft and a planted risk mention, then date it, sign it, and attach it to your AI policy and supervision agreement.