Maria's client of fourteen months, the C-PTSD client whose thin 90837 notes once kept her up past ten, is ready to end treatment. PHQ-9 down from 19 to 6, the trauma work consolidated, both of them naming it in the same session: this is the ending. And here is the uncomfortable truth about behavioral health charts: most of them simply stop. No discharge summary, no continuity letter, just a last progress note followed by silence, which reads to a payer auditor like an unexplained cliff, to a board investigator like a question mark, and to the next treating clinician like a chart that withholds exactly what they need. The discharge cycle is the bridge between your episode of care and whatever comes next, and a bridge built halfway is worse than useless: in the clinician-initiated case, an incomplete ending is not just sloppy charting, it is the raw material of an abandonment complaint. By the end of this lesson you will run an AI-assisted discharge workflow that produces the two-document pair every clean ending requires: a discharge summary carrying diagnosis at admission and at discharge, interventions, outcome data, reason for termination, and recommendations; and a continuity-of-care letter to the PCP or next provider, scoped correctly, with 42 CFR Part 2 awareness where it applies, and with the abandonment-protection language the ACA and APA ethics codes require when the termination is yours rather than the client's.

The Bridge and the Cliff: Why Charts Are Not Allowed to Just Stop

Hold the image for the whole lesson: every episode of care ends at an edge, and the discharge documents are the bridge you build across it. On the far side stands someone who needs to cross: a future clinician who inherits the client after a relapse two years from now, a PCP managing the sertraline you watched work, a payer reviewer deciding whether the episode was medically necessary from intake to ending, a board investigator reconstructing whether the termination met the standard of care, or the client herself, requesting her records to start with someone new in another state. A chart that just stops gives each of them a cliff. The discharge summary is the bridge's span: it carries the whole episode in compressed, structured form. The continuity letter is the bridge's far ramp: it lands the relevant subset of that story exactly where the next provider stands, scoped to what that recipient may and should receive.

Notice that the two documents face different directions, which is why one cannot substitute for the other. The summary faces inward and backward: it is part of the medical record, written at full clinical register, complete enough that the entire fourteen months can be reconstructed from it alone. The letter faces outward and forward: it leaves the practice, enters someone else's record, and therefore answers to disclosure law, minimum-necessary scoping, the psychotherapy-notes carve-out, and, where SUD treatment is in the picture, 42 CFR Part 2 with its redisclosure prohibition. Practices that skip the pair tend to fail in one of two mirrored ways: they send the internal summary outward, over-disclosing, or they file the outbound letter as the only closure document, under-recording. The pair exists because the two audiences cannot be served by one artifact.

Why do so many charts stop at the cliff anyway? Because discharge documentation is the least urgent task in the building. The client is better, or gone; no session tomorrow depends on it; no one is calling for it this week. It is the perfect task to defer forever, and the perfect task for an AI-assisted workflow, because every fact the summary needs already exists in the chart you kept: the intake diagnosis, the dated scores, the interventions named session by session, the treatment plan reviews. The model's job is compression and structure across a long document set, which is the one thing models do tirelessly and clinicians at the end of an episode do not. Your job is everything the chart cannot say by itself: why it ended, what you recommend, and your signature on both documents, because the model assembles the bridge's materials and only the licensed engineer certifies it can bear weight.

Anatomy of the Discharge Summary: Five Load-Bearing Sections

The summary has five load-bearing sections, and each exists because a specific future reader needs it. First, diagnosis at admission and diagnosis at discharge, side by side with their ICD-10 codes. The pairing is the point: F43.10 at intake remaining F43.10 in partial remission at discharge tells the trajectory story in one line, and a payer reviewer reads that pairing as the skeleton of medical necessity for the entire episode. If the diagnosis changed mid-treatment, the summary says when and why, because an unexplained diagnostic shift is the kind of loose thread auditors pull. Second, the interventions: not "supportive therapy provided" but the actual modalities delivered and roughly when (trauma-focused CBT phases one through three, EMDR resourcing in months four through six, skills consolidation thereafter), because the next clinician's first question is always "what has been tried, and how did she respond?"

Third, outcome data, and this is the section where the program's payer rule applies with full force: the verifiable details come from the chart through you, never from the model's plausible imagination. Dated scores, named instruments, real deltas: PHQ-9 19 on the intake date, 11 at the 90-day review, 6 at the final administration, with dates attached to each. An AI asked to "summarize progress" without those anchors will write "client showed significant improvement," which is decoration, not data; an auditor cannot recoup decoration into evidence, and the next clinician cannot baseline against it. Fourth, reason for termination, stated plainly from a small honest menu: goals met, mutual agreement; client-initiated withdrawal; moved out of area; lost to contact after documented outreach; clinician-initiated for a stated clinical or practice reason; transfer at a level-of-care change. Fifth, recommendations and condition at discharge: current risk status as you assessed it (yours to assess, never the model's), medications observed, relapse indicators worth watching, and the concrete follow-up recommendation, since "return as needed" carries more legal weight when it is written than when it is remembered.

Assemble it in that order and read it as the future clinician will: can someone who never met the client safely pick up her care from this one document? That question, not word count, is the completeness test. Most summaries built from a well-kept chart run one to two pages. The C-PTSD client's summary will note the co-occurring F43.12 history, the trajectory of the trauma work, the score arc, the mutual ending, and the door left open, and it will take Maria perhaps twenty minutes with AI assembly and a real verification pass, against the zero minutes the task got in the years when charts at her practice simply stopped.

Reason for Termination: The Section That Touches the Abandonment Line

Of the five sections, reason for termination is the one with ethics-code teeth, because it is where the chart proves the ending was a termination and not an abandonment. When the ending is mutual or client-initiated, the documentation burden is light: state it, date it, note the client's voice in it. When the termination is clinician-initiated, the burden inverts, and both major ethics codes say so directly. The ACA Code of Ethics requires counselors to avoid abandoning clients and, when terminating, to provide pre-termination counseling and recommend other service providers. The APA Ethics Code runs parallel for psychologists: terminate when the client no longer needs or is not benefiting from the service, and except where precluded by circumstances, provide pre-termination counseling and suggest alternative providers. NASW's standards put social workers under the same expectations. The codes differ in wording; the architecture is identical: notice, a clinical rationale, a transition path, and a bridge that does not drop the client mid-span.

Translate that architecture into chart language, because an ethics requirement that is not documented is an ethics requirement you cannot prove you met. The clinician-initiated discharge summary states: the clinical rationale for ending (no longer benefiting, needs exceed scope of practice, level-of-care change, practice closure, an irreparable frame rupture); the date pre-termination counseling occurred and what was discussed; the referrals actually given, specific and plural where possible, with names or referral pathways rather than "client advised to seek services"; the reasonable notice period offered; and the interim coverage arrangement, including crisis resources provided in writing. That cluster is the abandonment-protection language, and it has one purpose: a board investigator reading this chart two years later should find, inside the record itself, every element the ethics codes require, dated and concrete.

A termination is an ending the chart can defend. An abandonment is the same ending with the defense missing. The difference, two years later, is entirely in what you wrote down.

Now the AI boundary, stated bluntly: the model does not decide to terminate, does not pick the reason from the menu, and does not generate the abandonment-protection cluster from thin air, because every element of that cluster is a fact about what you actually did. If the pre-termination conversation has not happened, no drafting tool can make it have happened; a model asked to "write the termination section" without the facts will hallucinate a tidy compliance story, which is worse than a blank, because it is a falsifiable fiction in a legal record. The workflow is: you do the clinical work of ending well, you log the facts (dates, referrals, notice, coverage), and the AI formats those facts into the section. Drafting follows the ending. It never substitutes for it.

The Continuity Letter: Scoped, Outbound, and Aware of Part 2

The second document of the pair travels. The continuity-of-care letter to the PCP, psychiatrist, or next therapist is short by design: identification of the shared client and the treatment dates; diagnosis at discharge; a two-or-three sentence course-of-treatment summary; the outcome anchors (the same dated scores, not the narrative); medications you observed and the prescriber who manages them; current risk status with its assessment date, as you determined it; and the specific follow-up recommendation with the door-open line ("client was advised she may return; please contact me for coordination with a signed release"). Everything else in the fourteen months, the trauma narrative, the family material, the content of sessions, stays home. The letter is built from the medical record, never from psychotherapy notes, which sit behind their own authorization wall under the HIPAA carve-out, and it is scoped to the minimum necessary for the recipient's purpose: a PCP managing an SSRI does not need the chapter, only the bridge ramp.

Then the federal layer that turns a routine letter into a violation when missed: 42 CFR Part 2. If the client received substance use disorder treatment from a Part 2 program, records of that treatment carry protections beyond HIPAA even at discharge, under the 2024 final rule that aligned Part 2 more closely with HIPAA while preserving its core: the disclosure must be covered by compliant consent, and the letter must carry the redisclosure prohibition notice telling the recipient the information cannot be further disclosed without consent except as the regulation permits. The discharge moment is precisely where this gets missed, because the writer is summarizing "the whole picture" and the SUD treatment is part of the picture. Keep the notice as a saved verbatim block and insert it whenever Part 2 information rides in the letter; a paraphrased legal notice is a defective legal notice, and a model will paraphrase anything it is allowed to touch. If you are unsure whether Part 2 attaches to your setting, that is a question to resolve with compliance counsel before the letter exists, not a judgment to delegate to a drafting tool.

One more scoping discipline: the continuity letter inherits the release. Confirm the signed ROI or the applicable treatment-disclosure basis before drafting, and write the letter inside its boundaries: named recipient matches, scope covers what the letter contains, expiration valid. The AI never decides what may be disclosed, to whom, or under which consent; it drafts inside a scope you set, and your verification pass checks every sentence against that scope the same way the handoff lesson taught: line by line, before the timestamped signature, before anything transmits.

The AI Assembly Workflow: From Chart to Draft Pair

Here is the workflow, end to end. Step one, gather the source set: intake assessment, treatment plan and its reviews, the dated measurement scores, the last several progress notes, and your termination facts log (reason, pre-termination counseling date, referrals, notice, coverage). De-identify per your standard practice if your tool sits outside a BAA-covered EHR integration; better, run this inside the covered tool you vetted in earlier chapters. Step two, the summary prompt, and the structure matters more than the eloquence: "From the attached chart materials only, draft a discharge summary with these sections: (1) diagnosis at admission and at discharge with ICD-10 codes as documented; (2) interventions delivered with approximate timeframes as documented; (3) outcome data using only the dated scores provided, verbatim, with dates; (4) reason for termination using only the facts in the termination log; (5) recommendations and condition at discharge, leaving risk status as a bracketed field for clinician completion. Use only facts present in the materials. Do not infer, embellish, or characterize progress beyond the data. Flag any required field for which the materials contain no information."

Step three, verify the summary the way the Handoff Map prescribes: every diagnosis and code against the chart, every score and date against the actual instruments, every intervention against the notes, the termination section against your log, and then complete the bracketed risk field yourself, because risk status at discharge is a clinical determination and the model leaves a bracket precisely so there is nothing to "verify," only something for you to author. Step four, the letter prompt, fed by the signed summary, not the raw chart: "From the verified discharge summary, draft a continuity-of-care letter to [recipient, role] containing only: client identification and treatment dates, discharge diagnosis, a three-sentence course summary, the dated outcome scores, observed medications and prescriber, the clinician-completed risk status with date, and follow-up recommendations with a coordination invitation. Under 300 words. No session content, no history beyond the listed elements, no psychotherapy-note material. If the materials indicate SUD treatment records, stop and flag rather than including them." Step five, the Part 2 decision and the verbatim notice block if applicable, made by you. Step six, sign both documents (timestamped, liability transfers, exactly as the handoff lesson defined), transmit the letter through the secure channel, log the send against the release, and file the pair as the episode's closing entries.

Two failure modes to watch in verification, both fluent. Trajectory inflation: the model smooths the score arc into a triumph narrative ("steady and remarkable improvement") that the actual data, with its month-nine spike during the custody hearing, does not support; auditors compare adjectives to numbers, so keep the adjectives out. And menu substitution: asked to state the reason for termination, a model with thin facts will pick the most flattering plausible option (goals met) when the truth is lost-to-contact; that substitution converts an honest difficult ending into a misrepresentation in a legal record. Both are caught the same way everything in this program is caught: line-by-line reading against the source, before the signature, every time.

The Hard Endings: Lost to Contact, Level-of-Care, and Risk at Discharge

Clean mutual endings are the easy case. The discharge cycle proves its worth on the hard ones. Lost to contact: the client stops responding after session nineteen. The summary documents the outreach sequence (the calls, the letters, the dates), states the true termination reason, and records that a closing letter was sent to the client with referrals and crisis resources, with the door explicitly open for return; that documented outreach is the abandonment protection when the client never formally said goodbye. Level-of-care transfer: the client steps up to IOP or PHP; the summary states the clinical rationale for the change, and the continuity letter to the receiving program is the most consequential letter in this lesson, because the receiving team treats your outcome data and intervention history as the baseline for a higher-acuity admission. Practice closure or clinician departure: the notice period, the transfer options offered, and the records-access instructions all go in writing, because closures are where abandonment complaints cluster.

And the case that overrides everything: elevated risk at the moment of discharge. If your assessment at the final session finds active risk, the discharge does not proceed on autopilot, and neither does the documentation. The clinician, never the model, makes and documents the risk determination, the safety planning, and the disposition (continued care, higher level of care, crisis linkage), and the discharge paperwork follows the clinical reality rather than the calendar. This is the discharge cycle's version of the escalation hand-back from the handoff lesson: any risk-relevant content in the chart materials or the drafts pulls the process back to clinical judgment before any document advances. A discharge summary that papers over an unresolved risk picture is the single worst document this workflow can produce, because it is the one most likely to be read aloud, slowly, to a jury.

Across all the hard endings, the rule that organizes them is the same bridge rule: the documentation answers the question "what would the next responsible person need, and what did the client receive on the way out?" If both answers are in the chart, dated and specific, the ending is defensible whatever shape it took. If either is missing, no elegance in the prose will supply it.

The Applied Problem: Produce the Discharge Pair

Your artifact is the Discharge Pair: a completed discharge summary and matching continuity-of-care letter for one real, de-identified, recently ended (or nearly ended) case, plus the reusable prompt blocks that generated them. Start with the source set: pull the intake, the treatment plan and reviews, every dated score, the final notes, and write the termination facts log by hand first, one short paragraph per element (reason from the honest menu; if clinician-initiated: pre-termination counseling date and content, the specific referrals given, the notice period, the coverage and crisis resources provided). Writing the log first enforces the lesson's deepest rule: the ending happens in clinical reality before anything is drafted, and the model formats facts, it does not manufacture them.

Run the summary prompt from the workflow section against your source set. Then verify in this order: codes against the chart; the admission and discharge diagnoses paired correctly; every score, instrument name, and date checked against the actual administrations; interventions against the notes; the termination section against your handwritten log word for word; and the risk-status bracket completed in your own words, with the date of your final assessment. Hunt specifically for the two named failure modes: strike any progress adjective the numbers do not support, and confirm the stated termination reason is the true one, not the flattering one. Sign it, timestamped.

Now generate the letter from the signed summary using the letter prompt, addressed to the actual or hypothetical next recipient (PCP for a medication-sharing case; receiving program for a level-of-care case). Verify against three rulers: the release (recipient, scope, expiration), the minimum-necessary standard (could every sentence be read aloud to the client without a flinch, and does the recipient need it), and the Part 2 question (if any SUD treatment records ride in this letter, the verbatim redisclosure prohibition block goes in, inserted by you, never paraphrased by the model). Confirm it contains the dated outcome anchors and your clinician-authored risk status, runs under 300 words, and ends with the coordination invitation. Sign, transmit securely, log the send.

Done looks like four files saved together: the signed summary, the signed letter, your termination facts log, and the two prompt blocks with a dated note of every correction your verification pass caught. That correction list is not bookkeeping; it is your evidence of supervision over the tool, and it is what makes the next discharge faster, because the prompts get amended wherever the model drifted. From this case forward, no chart in your practice ends at a cliff: every episode closes with a bridge, built from chart facts, certified by your signature, and carrying, when the ending was yours to initiate, the abandonment-protection language that turns an ending into a defensible termination.

Key Takeaways

• Charts are not allowed to just stop. The discharge cycle builds the bridge between your episode of care and whatever comes next, and it always produces a pair: an inward-facing discharge summary for the record and an outward-facing continuity letter scoped to its recipient. One document cannot serve both audiences.
• The summary has five load-bearing sections: diagnosis at admission and at discharge with ICD-10 codes, interventions with timeframes, outcome data as dated scores, reason for termination from an honest menu, and recommendations with the clinician-authored condition and risk status at discharge. The completeness test is whether a clinician who never met the client could safely pick up her care from this one document.
• Outcome data is where the payer rule bites: real, dated, named-instrument scores (PHQ-9 19 to 11 to 6, with dates), never the model's "significant improvement." Auditors compare adjectives to numbers, and trajectory inflation is a fluent failure mode you catch only by reading against the chart.
• Clinician-initiated terminations carry the abandonment line: the ACA and APA codes (and NASW's parallel standards) require pre-termination counseling, referrals, notice, and a transition path, and the chart must show each element dated and concrete. The AI never decides to terminate, never picks the reason, and never fabricates the protection cluster; you do the ending, log the facts, and the model formats them.
• The continuity letter is short, built from the verified summary rather than the raw chart, scoped to minimum necessary, free of psychotherapy-note material, and Part 2-aware: when SUD treatment records ride in it, the 42 CFR Part 2 redisclosure prohibition goes in as a saved verbatim block under the 2024 final rule, inserted by you, because a paraphrased legal notice is a defective one.
• The hard endings are the point: lost-to-contact requires documented outreach and a closing letter with referrals and crisis resources; level-of-care transfers make the continuity letter the receiving program's baseline; and elevated risk at discharge halts the autopilot entirely, because the clinician, never the model, makes and documents the risk determination before any paperwork advances.
• The artifact is the Discharge Pair plus its scaffolding: signed summary, signed letter, handwritten termination facts log, and the reusable prompt blocks with your correction list. Both signatures are timestamped liability transfers, nothing transmits unsigned, and the correction list is your documented supervision of the tool.