Carmen is an AMFT in Fresno, twenty months from licensure, paying $59 a month out of a $32-an-hour wage for an AI scribe her agency never sanctioned, because the alternative is losing two unpaid hours every night to documentation. Last week, in the supervision meeting the BBS requires, her supervisor asked what tools she uses. Carmen told the truth. The supervisor said, "We need to talk about that," and both of them went home to reread a supervision agreement, signed eighteen months ago, that does not contain the word "AI" anywhere. Carmen does not know if her hours still count. The supervisor does not know if her signature on the next quarterly form is now exposed. Neither of them did anything wrong, and both of them are unprotected, because the document that defines their relationship is silent on the most consequential change in how associates work. This lesson rewrites that document. By the end you will have a Supervision Contract AI Addendum: the disclosure duty, the supervisor's review duty for AI-drafted notes, and the explicit line of responsibility when a pre-licensed AMFT, APCC, or ASW uses AI under supervision, written so a board auditor can follow it.

The Co-Signed Loan: What Supervision Legally Is

The controlling analogy for this lesson is the co-signed loan. When a parent co-signs a child's car loan, the parent is not making the payments, choosing the car, or driving it, and yet every missed payment lands on the parent's credit with the full force of the original obligation. That is supervision. The pre-licensed associate does the clinical work, but the supervisor's license is the collateral behind every hour, every note, and every clinical decision made under that supervision. Boards built the structure deliberately: the associate cannot practice independently precisely because the profession has not yet certified their judgment, so a fully licensed clinician lends their judgment, and their license, to cover the gap. The supervisor of record answers for the supervisee's practice, which is why the California BBS holds supervisors responsible for the supervision they provide, why New York's OPP, the Texas BHEC, Florida's DOH for Chapter 491 professions, and the Illinois DPR all build supervisor accountability into their pre-licensure frameworks, and why a board complaint about an associate's note routinely opens a file on the supervisor too.

Now run AI through the analogy. When Carmen pays for her own scribe, it is as if the borrower quietly refinanced the loan through a third party the co-signer has never heard of. The payments may keep arriving on time, the notes may even improve, but the co-signer's exposure changed without their knowledge or consent. The supervisor who signs Carmen's quarterly hour verification is attesting, to the BBS, that she has been supervising Carmen's work, and Carmen's work now includes a tool the supervisor did not know existed, processing client PHI under an account the agency never vetted, producing drafts the supervisor has never seen in raw form. Nothing about that is hypothetical malice; it is two good-faith people operating under a contract written for a world that no longer exists. The fix is not punishing Carmen. The fix is the addendum.

Why the Contract Must Say It Explicitly

Supervisors sometimes argue the agreement does not need an AI clause because "everything the associate does is already under my supervision." That sentence is exactly the problem: it concedes total exposure while securing zero visibility. If everything the associate does is on the supervisor's license, then the supervisor's only protection is knowing what the associate does, and the only enforceable mechanism for knowing is a written disclosure duty. An unwritten norm cannot be produced for a board auditor; a contract clause can. When the BBS audits a supervisor's file, when the BHEC reviews a complaint, when OPP examines whether supervision was real or rubber-stamped, the supervision agreement is the first exhibit. An agreement that anticipates AI, requires disclosure, defines the review duty, and assigns responsibility reads as competent supervision. Silence reads as a supervisor who never asked.

There is a second, quieter reason the contract must say it: the associate deserves to know the rules before breaking them. Carmen was not hiding anything; nobody told her there was anything to disclose. Pre-licensed clinicians are the most financially squeezed and most documentation-burdened people in the field, which makes them the earliest and most motivated adopters of AI scribes, and the least equipped to evaluate BAAs, subprocessor lists, or consent implications. A contract that is silent on AI does not prevent associate AI use; it guarantees that associate AI use happens invisibly, on consumer accounts, without client consent language, until a supervisor's casual question detonates it. The addendum converts a hidden practice into a governed one, and in doing so it protects the associate's hours, the supervisor's license, and the client's confidentiality in a single document.

And say the central sentence in the contract itself, in plain language, because the document is also a teaching instrument: the supervisor's license is on the line for clinical work product the supervisee produces with or without AI assistance, and AI use by the supervisee is therefore a supervision matter, not a personal productivity choice. Once that sentence exists on paper, every other clause follows from it logically, and no one can later claim the stakes were unclear.

If everything the associate does rides on the supervisor's license, the supervisor's only real protection is a written right to know, and a written duty to look.

Clause One: The AI Disclosure Duty

The first clause of the addendum is the disclosure duty, and it has three parts. Part one, the inventory: the supervisee discloses, in writing, every AI tool used in connection with clinical work, including tools paid for personally, the account tier, and what client information the tool touches. The personal-payment point is the Carmen clause: $59 a month from her own pocket did not move the tool outside the supervision relationship, because the data it processes belongs to clients of the agency and the work product lands in charts the supervisor answers for. Part two, the standing duty: disclosure is not a one-time form but a continuing obligation, triggered before any new tool is used, not after. Part three, the approval gate: no AI tool touches client information until the supervisor, or the practice's governance process where one exists, has confirmed the basics, a BAA at the tier in use, retention settings, consent language in the affected clients' charts, and alignment with the practice's approved-tool list.

Write the gate for what it is: it is permission, not prohibition. A blanket AI ban in a supervision agreement is the contractual equivalent of abstinence-only education; it does not stop the 9:54 PM temptation, it just removes the disclosure incentive and drives use underground, which is the most dangerous configuration available. The agreement should name the sanctioned path, the approved tools, the account the agency provides or reimburses, the consent addendum clients sign, precisely so the associate never faces Carmen's choice between two unpaid hours and an unsanctioned tool. If the agency expects associates to produce timely, audit-quality notes at agency caseloads, the agency should provision the compliant tool. That sentence belongs in the supervision conversation even if it does not belong in the contract.

Clause Two: The Supervisor's Review Duty for AI-Drafted Notes

The second clause defines what the supervisor actually does about AI-drafted documentation, and this is where most draft addenda go vague. "Supervisor will review supervisee documentation" was already in the old agreement and already insufficient. The AI-era review duty has to answer three questions concretely. How much: a defined review sample, for example all notes for the first thirty days of any new tool, then a standing percentage per week, with every note from any session involving risk content, suicidal ideation, homicidal ideation, abuse, intimate partner violence, reviewed in full, always. Against what: the review checks the AI-specific failure modes by name, fabricated or softened risk content, instrument results for instruments never administered, wrong-client contamination, templated phrasing that will not survive a payer audit, and diagnostic framing the associate adopted from the draft rather than derived from the client. With what record: a supervision log entry for each review cycle, because the review that is not logged did not happen, in the precise sense that it cannot be produced when the board asks.

The review duty also carries the chapter's hard rule into the supervision relationship, stated in the contract so that neither party can miss it: AI never scores risk instruments, never assigns a risk level, never makes the duty-to-protect determination, never makes a mandated-report decision, for the supervisee's clients exactly as for anyone's. The supervisee makes the clinical determination first, under supervision; the supervisor reviews the determination, not just the prose; AI formats afterward. For a pre-licensed clinician this rule is doubly load-bearing, because the entire premise of supervised practice is that the associate's clinical judgment is still being formed and verified. An associate who lets a confident draft anchor their risk assessment has not just made the universal error; they have outsourced the exact faculty supervision exists to develop. The supervision hour is where that gets caught, named, and trained, which is why the addendum should require that AI use be a standing agenda item in supervision, not a confession extracted by surprise.

Clause Three: The Line of Responsibility

The third clause draws the line of responsibility, and it must be drawn in ink because everyone involved is otherwise tempted to assume it sits elsewhere. The structure that survives scrutiny has three tiers. The supervisee owns the clinical work and the signature: the associate signs their notes, the signature is a legal attestation that the content is accurate and the work was done, and reading every word of an AI draft before signing is part of the attestation, not an optional quality step. The supervisor owns the supervision: the adequacy of the review sample, the documented oversight of AI-assisted work, the training of the associate in verification discipline, and the escalation when something is wrong; the supervisor's countersignature and quarterly verifications attest to the board that this oversight was real. The practice owns the infrastructure: the approved-tool list, the BAAs, the client consent process, and the incident-response runbook from the previous lesson, because an associate cannot be made responsible for vendor diligence the practice never performed.

Then the clause says the uncomfortable sentence directly: as between supervisee and supervisor, the supervisor of record bears professional responsibility to the licensing board for the supervisee's clinical work product, including AI-assisted work product, and nothing in the supervisee's use of any tool transfers, dilutes, or delegates that responsibility. The supervisor's license is on the line; the contract must say so explicitly, in those terms, because the sentence does two jobs at once. It tells the supervisor the true price of lax review, and it tells the associate why the disclosure duty and the approval gate are not bureaucratic hazing but the mechanism by which someone else's license can safely cover their work. Boards differ in machinery, the BBS with its quarterly hour verifications and supervisor responsibility statements, OPP, BHEC, DOH 491, and DPR with their own supervision plans and attestation forms, but every one of them converges on the same premise: the supervisor answers for the supervision.

Hours, Incidents, and the Board: The Operational Clauses

Three operational clauses finish the addendum. First, hours protection: the agreement should state that supervised experience hours are earned by clinical work performed by the supervisee, that AI-assisted documentation does not convert clinical work into non-qualifying work, and that the disclosure and review provisions exist partly to keep the hour record unimpeachable. This is the clause that answers Carmen's 2 AM fear. Her hours are not voided because a scribe formatted her notes; her hours become questionable only if the supervision around them looks fictional. An agreement that shows active oversight of AI-assisted work is precisely what makes the quarterly verification signature safe to give, for the supervisor, and safe to rely on, for Carmen, when the board reviews her application for licensure.

Second, the incident clause: the supervisee reports any suspected AI incident, a wrong-client draft, fabricated content discovered after signing, a possible exposure of client information, to the supervisor the same day, and the practice's incident-response runbook governs from there, with the no-blame reporting principle restated in the supervision agreement itself. Associates are the likeliest first detectors of AI failures and the most discipline-fearful people in the building; the supervision contract is where their reporting safety has to be guaranteed in writing, because an associate who fears for their hours will bury an incident, and a buried incident ages into exactly the late-discovered breach the previous lesson priced out. Third, the modification clause: AI terms are reviewed at a set cadence, at least annually and whenever the practice's approved-tool list or the regulatory landscape changes, so the addendum does not fossilize the way the original agreement did. Carmen's agreement was eighteen months old and already obsolete; the cure includes a review date, not just new language.

The Conversation Before the Signature

The addendum will fail if it arrives as a disciplinary artifact, so sequence the human side deliberately. The supervisor who discovered Carmen's scribe has one productive opening: "You did what the workload made rational, the agreement failed us both by being silent, and we are going to fix the document, not punish the disclosure." Then walk the addendum clause by clause in a supervision session, because every clause is a teaching opportunity: the disclosure duty teaches vendor diligence, the review duty teaches the verification pass and the risk-content rule, the responsibility clause teaches what a license actually is, the incident clause teaches the detection culture, and the hours clause converts the associate's deepest anxiety into the reason the whole structure exists. A pre-licensed clinician who internalizes this addendum has learned more about professional risk management than most licensees learn in a decade, which is the real argument for doing it well: supervision contracts shape the licensees the profession gets.

For the supervisor, one more discipline: do not sign the next quarterly verification until the addendum is executed and the first review cycle is logged. The verification is the supervisor's attestation to the board, and attesting to oversight of a tool-assisted practice you have not yet documented overseeing is precisely the exposure this lesson exists to close. The gap between "we talked about it" and "the file shows it" is the entire distance between a defensible supervisor and an exposed one, and the addendum plus the first logged review closes it in under two weeks.

The Applied Problem: Draft Your Supervision Contract AI Addendum

Your artifact is the Supervision Contract AI Addendum, one to two pages, attached to the existing supervision agreement by reference, and you can draft it in an afternoon. Structure it in six sections. One, recitals: name the parties, the existing agreement, and the governing board (BBS, OPP, BHEC, DOH 491, or DPR) so the document anchors to the right regulatory frame. Two, the disclosure duty: written inventory of all AI tools including personally paid tools, continuing disclosure before any new tool, and the approval gate with its four checks (BAA at the tier in use, retention settings, client consent on file, approved-tool alignment). Three, the review duty: the defined sample, the full review of all risk-content sessions, the named AI failure modes checked, and the supervision-log entry per cycle. Four, the responsibility clause, including the explicit sentence that the supervisor's license bears professional responsibility to the board for the supervisee's AI-assisted work product, and the supervisee's signature attestation duty to read every word before signing. Five, the risk-determination rule verbatim: AI never scores risk instruments, never assigns risk levels, never makes duty-to-protect or mandated-report determinations; the supervisee determines, the supervisor reviews, AI formats afterward. Six, the operational clauses: hours protection, same-day incident reporting with the no-blame principle, and the annual review date.

You may use AI to assemble the draft from these specifications; the prompt: "Draft a supervision agreement addendum from the following six numbered specifications. Use plain contractual language a state board auditor can follow. Do not add obligations, board names, or legal standards I have not provided. Flag any clause that lacks a defined trigger, owner, or record." Then run the verification pass with both parties reading: Does the disclosure clause cover personally paid tools by name? Does the review clause define how much, against what, and with what record? Does the responsibility clause contain the license-on-the-line sentence in plain terms? Does the risk-determination rule appear verbatim? Is there a review date? "Done" looks like this: the addendum signed by supervisor and supervisee, walked through clause by clause in a logged supervision session, the first review cycle completed and logged, and the document filed where the quarterly verification, the practice's approved-tool list, and the incident runbook can all point to it. Carmen keeps her hours, the supervisor keeps her signature safe, and the next associate never has to guess what the right thing is.

Key Takeaways

• Supervision is a co-signed loan: the associate does the clinical work, but the supervisor's license is the collateral behind every note and decision, and AI use by a supervisee changes the co-signer's exposure whether or not anyone intended it. The California BBS, New York OPP, Texas BHEC, Florida DOH 491, and Illinois DPR all converge on the same premise: the supervisor answers for the supervision.
• "Everything the associate does is under my supervision" concedes total exposure while securing zero visibility. The only enforceable protection is a written disclosure duty plus a written review duty, because an unwritten norm cannot be produced for a board auditor and silence reads as a supervisor who never asked.
• The disclosure clause must cover personally paid tools explicitly, impose a continuing duty triggered before any new tool is used, and run through an approval gate checking the BAA at the tier in use, retention settings, client consent, and the approved-tool list. A blanket ban drives use underground, the most dangerous configuration available; the sanctioned path is the real control.
• The review duty is defined by how much (a set sample, plus every risk-content session in full), against what (the named AI failure modes: fabricated or softened risk content, phantom instrument results, wrong-client contamination, templated phrasing, adopted diagnostic framing), and with what record (a supervision-log entry per cycle, because an unlogged review cannot be produced).
• The contract states the hard rule verbatim: AI never scores risk instruments, never assigns a risk level, never makes duty-to-protect or mandated-report determinations. The supervisee determines under supervision, the supervisor reviews the determination itself, and AI formats afterward; for a pre-licensed clinician, outsourcing judgment to a draft defeats the very faculty supervision exists to develop.
• The responsibility clause says the uncomfortable sentence in ink: the supervisor of record bears professional responsibility to the board for the supervisee's AI-assisted work product, and no tool transfers or dilutes it. The supervisee signs every note as a legal attestation; the practice owns the infrastructure of BAAs, consent, and the approved-tool list.
• The operational clauses do the daily work: hours are protected by documented oversight, not threatened by disclosed tools; incidents are reported same-day under a written no-blame principle; and the addendum carries an annual review date so it never fossilizes the way the original agreement did. The supervisor signs no quarterly verification until the addendum is executed and the first review cycle is logged.