Can AI Draft Mental Health Treatment Plans? The Review Steps

AI can draft the wording of a mental health treatment plan, but it should not choose the goals, diagnosis or interventions. The clinician sets those with the client, gives the AI only de-identified text, then reviews the draft, checks it with the client and signs it. Any identifiers stay out of tools without a signed data agreement.

That split is the whole approach. Academic reviews of AI in mental health tend to focus on what models can predict or generate. A practice owner needs a narrower answer: which part of a treatment plan is clerical, and which part is clinical judgement? Formatting a goal into measurable language is clerical. Deciding that this client's first goal is sleep rather than panic is judgement, and it belongs to the clinician and the client.

Follow me on Instagram@sagnikteaches

The four-part split: who owns what

A treatment plan has four parts that behave differently around AI. Keep them apart in your head and your workflow.

Connect on LinkedInSagnik Bhattacharya

If a tool starts proposing goals or diagnoses unprompted, switch that off or change the prompt. A model that suggests "treatment goal: reduce catastrophising" has made a clinical suggestion, and clinicians who see it first tend to anchor on it, even when the client's own priority was something else.

Subscribe on YouTube@sagnikteaches

Where identifiers stay out

A treatment plan is among the most sensitive documents a practice holds. The working rule: identifiers stay out of any tool that has not signed a business associate agreement or equivalent data contract with you. Identifiers means names, dates of birth, addresses, phone numbers, employers, schools, named family members, rare events that make the client recognisable, and anything else that could point to one person.

De-identification is harder than deleting a name. "A 34-year-old paramedic whose sibling died last March after a diving accident" is identifiable without a name. For drafting, reduce it to "an adult client with work-related stress and recent bereavement". If the plan needs the detail, write that line yourself after the AI has done the formatting.

Check vendor terms, not assumptions. One example of why: SimplePractice's Note Taker, a $35-a-month add-on per clinician, keeps a de-identified transcript by default for clinicians who switched it on from mid-June 2026, while earlier users stay opted out. You can opt out per clinician, client or session, but only if you know the setting exists. Settings and terms change, so read the current ones. For a fuller checklist, see AI note-taking tools for therapists: a confidentiality checklist, and for vendor questions about agreements, which vendors sign a business associate agreement.

The workflow in seven steps

1. Agree the goals in the session. Use the client's own words. Write them down by hand or in your own system.
2. Write a de-identified brief. Two or three lines on presentation and the agreed goals, with no identifiers.
3. Prompt the AI to format. Ask for objectives in measurable language, in your template, and nothing else.
4. Review against the checklist below.
5. Re-insert identifiers in your own system. The final document is assembled where the client record lives.
6. Review with the client. Read it together; change what does not sound like them.
7. Sign, date and file. Note in the record that AI assisted with wording.

A prompt that keeps AI in its lane

I am a licensed counsellor. Format the objectives below into measurable,
strengths-based language for a treatment plan. Use my headings: Goal, Objective,
Intervention, Review date. Keep the client's wording for goals where given.
Do not add goals, diagnoses, risk statements or interventions that I have not
listed. If something is unclear, ask me rather than guess. Plain English, no jargon.

Goals as agreed with the client (de-identified):
1. "I want to sleep through most nights."
2. "I want to be able to go to team meetings without leaving."
Interventions I have chosen: CBT for insomnia components (sleep diary, stimulus
control); graded exposure for meeting situations. Sessions weekly for 8 weeks,
review at session 8.

A realistic output, illustrative: the tool returns a tidy table, and in one place it adds "reduce anxiety symptoms by 50%". You did not give a number, and the target is an invented measure. Remove it, and decide yourself whether a number belongs there. Where you use a scale, say a sleep diary average or a standard anxiety questionnaire, name it and the baseline from your own record.

Some practice software now builds plan drafting in. SimplePractice's Care Aide add-on, for instance, includes a Treatment Planner for creating and updating plans, listed at $59 a month for the account owner and $49 for each additional clinician before introductory discounts. A built-in planner may be covered by the data agreement you already have with that vendor, so check before relying on it, but it does not remove the review: its drafts go through the same checklist as a general chatbot's, because the risk of invented goals or measures is the same.

The clinician review checklist

Review the draft in this order, and tick each item. Anything unticked goes back for rewriting before it reaches the client.

1. Goals match the session. Each goal is the client's, in words they would recognise. Nothing was added, nothing was lost.
2. No invented content. Every symptom, history item, risk statement, number or date appears in your own records.
3. Interventions are yours. The approach listed is what you will actually deliver, at the frequency you can offer.
4. Measures are real. Any scale, target or timeframe is one you chose. Numeric targets have a baseline.
5. Risk is yours. Any risk or safety content was written by you, not the tool. If a plan mentions risk, check it against your latest assessment.
6. Language is respectful. No deficit labels, blaming phrases ("refuses", "non-compliant") or stereotypes. The client could read it without feeling reduced to a diagnosis.
7. Cultural fit. The plan reflects the client's context and preferences, including those the model could not know.
8. No identifiers entered the tool. Confirm the prompt history holds none, and that you can delete it.
9. Consistency. Dates, session counts and review points agree throughout the document.
10. Sign-off is human. You read the whole plan, not just the changes.

Time yourself on the first five plans. Review with the checklist should take about 10 to 20 minutes; if it takes longer than drafting from your own template, the tool is not earning its place.

Three drafts and what the review caught

These are illustrative cases, not reports of real clients.

The invented number. A draft for a client with low mood said "increase behavioural activation to five activities a week". The clinician had agreed "more activity" and no count. Five a week was the model's plausible guess. The clinician replaced it with the client's own target of two walks a week, which was achievable and theirs.

The deficit word. A draft described a client as "avoidant of social situations and unmotivated". The client had said they found groups draining after work. The plan now says "wants to rebuild social contact at a pace that feels manageable". The change took twenty seconds and altered how the client felt reading it.

The added risk line. A draft for a client with panic symptoms included "no current suicidal ideation reported". The clinician had not asked that day. A risk statement the clinician did not make is worse than none, and it was deleted. The review habit that caught it: read the plan for anything you did not say.

Reviewing the plan with the client

The client review is not a courtesy. A plan the client has read and corrected is a better plan, and it protects you when memory differs. Offer it in the session, on screen or on paper, and ask three questions: is this what we agreed, is there anything here you would not want written, and what would you change?

Offer a plain-English version for the client, with the same goals and a short "what we will do" paragraph. AI is good at that rewrite when it is working from the signed clinical text with identifiers removed. A sample sentence: "We will meet weekly for eight weeks. We will start with keeping a sleep diary, then try changing a few evening habits. We will look at how it is going in week eight and decide together what comes next."

Tell clients that AI helped with the wording, in your consent form and in the session. Use a short script: "I use a tool to help tidy the wording of plans. I give it no names or details that could identify you, and I check everything myself. You can ask me not to use it." Record their answer. Disclosure expectations differ between licensing bodies and insurers, so check yours. Should therapists tell clients they use AI? covers the reasoning.

One plan end to end: a client with insomnia and meeting avoidance

This is an illustrative walk-through for a generic adult client, to show the hand-offs. Nothing here describes a real person.

In the session. The clinician and client agree two goals: sleeping through most nights, and attending team meetings without leaving. They agree on eight weekly sessions and a mix of sleep-focused CBT and graded exposure. The client says they would like the plan to mention that mornings are better than evenings for sessions. The clinician writes these notes in the client record.

The brief. The clinician writes the de-identified brief shown in the prompt above, with no names, no employer and no dates. It takes three minutes.

The draft. The tool returns two goals, each with two objectives. One objective reads "client will complete a sleep diary on at least five nights a week". The clinician had said a diary, but not five nights, and keeps it because it is realistic, adding "agreed with client" to mark it as theirs. Another reads "client will attend three consecutive team meetings without leaving by week eight". The clinician changes it to "stay for the first fifteen minutes of one meeting by week four, building from there", because three consecutive meetings is too steep a first step and was never discussed.

The review. The checklist catches two further problems: the draft says "treatment will begin with psychoeducation", which the clinician does not plan to do, and a review date is written as week eight when sessions run to week eight and the review should be in the final session. Both are fixed. Total review time: 14 minutes.

The client. At the next session the client reads the plan on paper and asks to change "avoid meetings" to "feel able to stay in meetings". The clinician makes the change by hand and re-saves the document. The client also asks whether the tool saw their name; the clinician explains it did not, and shows the de-identified brief.

The record. The signed plan, the date reviewed with the client and the note "wording assisted by AI tool, no identifiers entered" go into the file.

The saving here is modest: perhaps 15 minutes on a plan that would take 30 by hand. The larger gains come from consistency across a caseload and from the plain-English client copy. Do not expect the tool to remove the review, because the review is the product.

Variations that change the review

Children and adolescents. Goals involve the young person, parents or carers and sometimes school. The review adds a check that the language is understood by the young person, and that consent and confidentiality arrangements are described accurately. Never put school names or family details in a prompt.

Couples and family work. There is more than one client, and goals can conflict. The tool will tend to merge them into one smooth list. Check that each person's goals are recorded as theirs and that conflicts are named rather than smoothed away.

Substance use and trauma work. Language matters more here: "clean" and "dirty", "addict" or "abuser" labels, and clinical terms that a client may find shaming. Read for tone line by line, and use the person-first wording your service has agreed.

Plans submitted to funders or insurers. Some payers require specific fields or levels of detail. Confirm the payer's template and ask the AI to fit it, but check that you have not been led into stating medical-necessity claims the record does not support.

Testing a tool on made-up clients first

Before using any tool on real work, test it on three invented cases. Write short fictional briefs, one routine, one complicated and one with a trap such as a client whose goal is to stop a behaviour that is not clinically concerning but which a model might pathologise. Run the same prompt on each and score the outputs against the checklist. Count how often it invents content or adds risk language, and run the same case twice to see whether the output changes noticeably between runs.

A tool that adds unrequested clinical content in two of three cases needs a tighter prompt, or is the wrong tool. Keep the three test cases and rerun them whenever the vendor updates the model, since behaviour can shift without notice.

Checking after a month that it is working

Use a few measures that need no special software. Count plans drafted with AI and the minutes spent on each from brief to signature. Count the edits you make per plan; if you are changing more than a handful of items, the prompt needs work. Note any client who asked not to have AI used, and whether you honoured that. Ask two or three clients, in your own words, whether the plan felt like theirs. Finally, have a colleague read a plan with no idea which parts were drafted and tell you if anything reads as generic.

If clients start describing their plans in words that sound like the tool's phrasing rather than theirs, you are accepting too much of the draft. Go back to the client's own language in step one.

One last guard concerns reuse. It is tempting to keep a finished plan as a model and ask the tool to adapt it for the next client. Do not paste a real plan back in, because it carries the previous client's details into a new prompt. Keep a separate, fictional template with placeholder goals, and adapt that. Over time this library becomes the practice's own house style, which makes drafts more consistent and reviews faster.

Sign-off and the audit trail

Sign-off means the clinician has read the complete plan, accepts it as their own work, and is responsible for it. A signature on a plan you skimmed is exactly what the review checklist is there to prevent.

Keep three things in the record: the signed plan, the date reviewed with the client, and a note that AI assisted with wording. You do not need to store the prompt text if it contained no identifiers, but keep your template and the brief rules so a colleague can see how plans are made. If your tool has chat history, check whether it keeps identifiers and delete history on a schedule.

When the plan is reviewed or updated, repeat the process with the new session notes. Never ask an AI to "update the plan" from the old document alone, because it will carry forward goals the client has since dropped.

When to skip AI for a plan entirely

Some plans are better written by hand. Skip the tool when the client has asked you not to, when the plan involves high-risk content such as safety planning, when the case is unusual enough that a template would distort it, and when you cannot de-identify without losing the point. A hand-written plan is still fine, and nothing here suggests every plan should use AI.

Group and agency practices should add one more rule: the practice, not the individual clinician, decides which tools are approved. A clinician using a personal chatbot account for client work is a data incident waiting to happen. Put the approved list in your policy, and revisit it quarterly. A therapist's AI draft of a practice policy and procedure manual shows how to write the policy itself.

For the surrounding note-taking decisions, best AI note taker for therapists: recording or dictation style compares the two approaches, and can therapists use ChatGPT for session notes goes through the consumer-tool risks. If you want help setting up the routine for your own practice, my AI implementation consultation can cover the data rules and checklist together.

Questions from therapists about AI-drafted plans

Can I paste a client's presenting problem into ChatGPT?

Not with identifiers, and not on a consumer plan unless you have checked the data terms. Strip names, dates, places, employers and distinctive details first, or use a tool that has signed a business associate agreement. When unsure, write the clinical content yourself and use AI only on wording.

Do clients need to know AI helped draft their plan?

Tell them. Say what the tool did (formatting and wording), what it did not do (choose goals or diagnose), and let them decline. Add a line to your consent form and record that you discussed it. Rules on disclosure vary, so check your licensing body.

How long should reviewing an AI-drafted plan take?

Plan on 10 to 20 minutes for a first plan, mostly on checking that goals match what was agreed and that language is measurable and respectful. If review takes longer than writing from your own template, the tool is not helping.

Should AI suggest the diagnosis or the interventions?

No. Diagnosis and the choice of treatment approach are clinical judgements, and a draft that includes them invites anchoring. Give the AI your decisions and ask it to phrase them, not to originate them.

Further reads

• Safe AI Uses for Counsellors That Never Touch Client Notes — AI uses for counsellors that never touch client notes.
• Therapy Practice Management Software: What the AI Add-Ons Cost — What AI add-ons in practice software cost per clinician.
• Therapist Termination Letter to Client: Wording and an AI Draft — Another wording task where AI helps and the clinician decides.
• Blueprint AI for Therapists: Notes, Cost and Alternatives to Check — A tool designed around notes and plans, with alternatives.
• Best AI Software for Therapists in a Group Practice: Seat Costs — Seat costs when several clinicians share a tool.
• How to Use AI on SimplePractice: Note Taker or Care Aide? — What a practice-software note taker does and does not do.
• AI for Mental Health Providers: A Roll-Out Plan for Group Practices — A seven-step roll-out sequence for a group therapy practice, with consent wording, an approved-tool list and a 90-day review.
• Mental Health Services AI: Intake, Waitlist and Referral Admin Jobs — A checklist of the intake, waitlist and referral admin jobs AI can take on in a mental health service, and the lines it must never cross.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: SimplePractice pricing page, Note Taker FAQs and transcript retention FAQ, checked 4 November 2026; the review checklist is practice guidance, not a regulatory standard. Academic reviews of AI in mental health care exist but none is quoted here.