AI for Mental Health Providers: A Roll-Out Plan for Group Practices

Roll AI out to a group practice in this order: write the policy, pick one pilot clinician, agree consent wording, set an approved-tool list, run one training session, let the pilot run for six weeks, then hold a 90-day review before anyone else joins. Start with admin work, never session content.

Most guidance in this area is academic: reviews of how AI might affect therapy and position statements on ethics. Useful, but a practice owner still has to decide on Monday morning who gets a login. What follows is a sequence you can run with a team of five to twenty clinicians, with the documents each step produces.

Follow me on Instagram@sagnikteaches

Why a sequence beats a free-for-all

Clinicians are already trying AI on their own. Some are using a free chatbot to tidy a letter to a GP or to draft a handout, and a few may be pasting in anonymised case notes they believe are anonymous. A practice owner has a duty of care for client information whether or not the practice has a policy, so the choice is between knowing what is happening and finding out after a complaint.

Connect on LinkedInSagnik Bhattacharya

A staged roll-out protects three things at once: clients, whose information is the most sensitive a practice holds; clinicians, who carry professional responsibility for what they sign; and the practice, whose insurer, regulator or professional body may ask what it did. The sequence below keeps the early steps cheap, reversible and free of client content.

Subscribe on YouTube@sagnikteaches

The example practice used in this tutorial

All names and figures here are illustrative. Imagine a group practice with eight clinicians (psychologists, counsellors and a couple of associate therapists), an office manager and a part-time biller. It sees around 140 active clients a week, runs a waitlist of 30, and uses a practice-management system for scheduling and notes. The owner wants less time spent on admin, not a different way of doing therapy.

Step 1: a policy of two pages, written before any login exists

The policy answers six questions. Which uses are allowed? Which are banned outright? Which need extra approval and client agreement? Which tools may be used and on what accounts? Who is responsible for checking output? What happens if something goes wrong? Keep it plain enough for a new associate to read in ten minutes.

A filled-in version of the allowed and banned lines for the example practice:

Allowed (no client information): policy and handout drafts, newsletter,
website copy, job adverts, training quizzes, template letters with placeholders.
Allowed with approval: administrative summaries of non-clinical emails;
scheduling and reminder wording using first names only.
Banned: pasting session notes, assessments, risk information or any
identifiable client detail into any AI tool; using AI to assess risk; using
AI to produce a diagnosis or treatment recommendation.
Needs separate pilot and client consent: any recording, transcription or
note-drafting tool.

If you already have a policy and procedure manual, add the AI policy as a section rather than a stray document. Drafting a practice policy manual with AI shows how a manual can be built and reviewed without exposing client information.

Step 2: pick one pilot clinician and give them a narrow job

Choose the clinician who is curious, careful and respected by colleagues. Not the most enthusiastic; the one whose judgement others trust, because their verdict will persuade or put off the rest. Give them a one-page brief: the three admin jobs they may try, the log they must keep, and the person to tell about any slip.

In the example, the pilot clinician is a counsellor who spends about four hours a week on non-clinical writing. The three jobs are: drafting the monthly client newsletter, turning her own bullet points into a psychoeducation handout on sleep, and writing template reply wording for the common enquiries the office manager forwards to her (fees, cancellation policy, waiting times). None of it contains a client's information.

A sample task with a realistic output

Turn these notes into a one-page client handout on sleep and low mood.
Plain language, reading age around 12, no medical claims, no promises of
outcomes. Notes: regular wake time; limit screens in the hour before bed;
caffeine cut-off early afternoon; write worries down earlier in the evening;
see your GP if sleep problems last more than a few weeks.

The output is usually a clean handout with five headed tips. What the clinician fixes: AI tends to add confident claims such as "this will improve your mood within a week" and sometimes invents numbers like "adults need exactly eight hours". Both come out. She also removes a line suggesting a supplement, which is outside her notes. Prompting and editing take about 25 minutes against the hour she used to spend, and the handout carries her voice because she rewrote the opening.

Step 3: consent wording clients can actually understand

Two documents do the work: a short notice about how the practice uses AI for administration, and a separate consent addendum that applies only if a client's sessions or records are ever touched by a tool. In the first phase of the roll-out, only the notice is needed, because no client information goes in. Write the addendum now anyway so the practice is ready, and state in the notice that no session content is used.

Client notice (for the website and intake pack)
We sometimes use AI tools to help with administrative writing, such as
newsletters, handouts and standard letters. We do not put your name, your
notes or anything you tell us in sessions into these tools. Your clinician
checks anything written with their help. If you have questions or would
like to know more, ask your clinician or the office manager at any time.

Consent addendum (only if a session-related tool is ever used)
I agree / do not agree that [tool name] may be used for [specific purpose,
for example drafting my clinician's notes from a recording of our session].
I understand that the recording is [stored for X days / deleted after the
note is written], who can access it, and that I can withdraw agreement at
any time without any effect on my care. Signed ... Date ...

Have an adviser who knows the rules for your clients and your insurer read both before use. Consent wording differs by profession and jurisdiction, and a practice should not rely on a template from the internet, including this one. The point of drafting early is that the conversation with the adviser happens before pressure to adopt a tool, not after.

Step 4: the approved-tool list and the settings behind it

The list is one table. For each tool: purpose, who uses it, which account, what may go in, and where the data goes. For the pilot, one general assistant on a business plan is enough. ChatGPT Business and Claude Team (each $20 a seat a month billed annually, $25 billed monthly) both state on their pricing pages, checked in November 2026, that they do not train on your content by default, and Microsoft 365 Copilot and Gemini in Workspace make similar commitments for business accounts. Personal accounts are different: consumer ChatGPT shares chats for training by default until the user turns the setting off, and Claude's consumer plans leave the choice to each user. Both business plans start at two seats, so a practice pays for two logins at the least.

Before any tool moves from "pilot" to "approved", the office manager saves the vendor's data-use and retention page, checks who at the practice can see chat history, and writes down how a departing clinician's access is removed. Resist adding tools because a vendor is persuasive; the most common reason for a messy register is a free trial nobody closed.

Step 5: a 60-minute training session with a live demonstration

Run one session, in person if possible, for every clinician and the office manager, led by the pilot clinician with the owner present. Showing a colleague do the work is more convincing than a policy. The structure that works:

1. Ten minutes: what AI can and cannot do, and the practice's rules, using the policy lines above.
2. Fifteen minutes: the pilot clinician's live demonstration of the sleep handout, including the corrections she made.
3. Fifteen minutes: a deliberately bad example. Show a prompt containing a client detail and ask the room to spot the problem before revealing it.
4. Ten minutes: how to report a slip and what happens next, with the assurance that reports are logged, not punished.
5. Ten minutes: questions, including from sceptics, with objections written up on a flipchart and answered in the review.

Give everyone a one-card summary of the rules. A card with five lines is read; a ten-page policy rarely is. Record attendance so that you can show the practice trained its team.

Step 6: the six-week pilot and its log

The pilot clinician keeps a simple log with date, task, tool, minutes spent, minutes the old way would have taken, and any correction needed. Weekly, the owner spends fifteen minutes reading it. The log shows what actually happened, not what people remember.

Illustrative log from weeks one to six in the example practice: two newsletters drafted in 35 minutes each instead of 90, three handouts in 25 minutes each instead of 60, and six template replies in 40 minutes instead of 2 hours. That is about 295 minutes saved, roughly five hours over six weeks for one clinician, which is the figure the review has to weigh. Two corrections of note: the assistant stated a fee the practice does not charge, and used spellings that do not match the practice's house style. Neither is dangerous; both show why a person reads everything.

Add one low-risk job from the credentialing side

Around week four, if things are steady, the office manager can try one more admin job that has no client data, such as tracking licence renewals and insurance dates for clinicians. A credentialing tracker built with AI is a good example, because the data is about staff documents, not clients, and mistakes are easy to spot.

Four admin jobs for the second quarter, once the review is clean

If the review at day 90 is clean, widen the admin list before touching anything clinical. Each job below uses practice information, not client information, and each has a quick way to check the output.

Waitlist wording that sets honest expectations

The office manager writes the facts: current wait of about nine weeks, two clinicians with openings on weekday evenings, a cancellation-list option. The assistant turns them into a friendly 90-word holding email with a placeholder for the first name. Before: "Thanks for getting in touch, we will be in touch when we can." After: a message that says what the wait is, what the client can do meanwhile and how to reach the office if their situation changes. The fix is always the same: delete any reassurance the practice cannot back, such as "we will find you a slot soon".

Cancellation and late-arrival policy in plain language

Paste the existing policy text, which contains no client details, and ask for a version at a reading age of around 12 with the fee rule in a single sentence. Illustrative result: a 420-word policy becomes a 150-word notice. Check that the notice keeps the notice period, the fee and the exceptions exactly as the policy has them, because shortened text tends to drop exceptions first.

A clinician job advert and interview questions

Ask for an advert from real facts: modality mix, supervision hours, caseload range, days offered. Then ask for ten interview questions about working in a group practice and handling a difficult supervision conversation. Practice owners often find the questions generic, so ask for scenarios based on the practice's own policies. No client information is involved, and the owner remains the judge of candidates.

Meeting agendas and action lists from the owner's own notes

For the monthly clinical meeting, the clinical lead can turn bullet points about practice matters into an agenda and, afterwards, a list of actions with owners. Keep case discussion out of the notes entirely; it stays in supervision. An example: eight bullet points about room bookings, a new referral route and the cancellation policy become a one-page agenda in about five minutes, with the actions copied into the shared task list.

What to decide about notes, scribes and treatment plans

The jobs people most want from AI in this field are session notes and treatment plans, and they are the ones that need the most care. A tool that records or transcribes a session captures the most sensitive material a client will ever share. Before any practice considers one, it needs client consent in writing, a data agreement with the vendor, a clear answer on where recordings and transcripts are stored and for how long, and a clinician who reads and signs every note as their own.

Drafting treatment plans from a clinician's own bullet points is lower risk than recording but still shapes a client's care. The review steps for AI-drafted treatment plans set out what a clinician should check line by line. In this roll-out, both uses wait for the review at day 90, and each gets its own mini-pilot with its own measures. If a clinician in the team wants it sooner, the answer is "not yet, and here is the date we decide".

Step 7: the 90-day review and what it should decide

Hold the review with the owner, clinical lead, office manager and pilot clinician. Use four questions and write each answer in a one-page memo.

The memo also names the next pilot, if any, with an owner and a date. Waiting is a legitimate outcome. A practice that decides "not notes, not yet" and records why has done its job; so has one that stops after finding the time saved was too small to bother with.

Three ways a practice roll-out quietly drifts

Three failures recur. The first is a policy nobody reads, so the card and the demonstration carry more weight than the document. The second is the quiet expansion, where the pilot clinician starts using the tool for a case summary because it is quicker, and nobody notices. The log, reviewed weekly, catches it. The third is a vendor with a polished demo for a recording tool that arrives before consent wording exists; the rule that no session tool is trialled before the addendum is reviewed protects against it.

An example of the second failure: after a busy week, the pilot clinician pastes a paragraph describing a client's presenting problem, with the name removed, into the assistant to tidy up a referral letter. She realises it is still identifiable by context and tells the owner that afternoon. The owner thanks her, logs it, deletes the chat, adds a line to the rules card ("Referral letters are written by hand or from templates") and shares the lesson at the next team meeting without naming anyone. The practice's culture, not the policy, is what made the report happen.

Small additions that make the roll-out stick

• Add AI use to the induction checklist for new associates so it is not rediscovered every hire.
• Put the client notice in the intake pack and on the website the same week, so no client first hears of it from a stranger.
• Check your professional indemnity or liability cover and ask whether AI use changes any condition; a short email to the insurer is enough, and the answer goes in the file.
• Repeat the review every six months, because tools and vendor terms change; a tool approved in November may behave differently by spring.

Intake and waitlist admin is the next area many practices look at, and intake, waitlist and referral admin jobs sets out which parts are safe. If you want a sounding board for the sequence itself, my AI implementation consultation covers choosing the first jobs and what the approved list should say. The steps above are enough to run a careful first quarter on your own.

Questions practice owners ask before rolling out

Should every clinician be allowed to use AI?

No. Start with one willing clinician and a short list of admin uses. Others join after the pilot review. Clinicians who prefer not to use AI should be able to opt out without any penalty, and clients should be able to opt out of any use that touches their sessions.

Is a signed consent form always needed?

It depends on the use. Drafting a newsletter needs none. Anything that records, transcribes or summarises a session needs explicit client agreement, written down. Ask your professional body or an adviser for the rule that applies to your clients and your insurance.

Can an AI tool write session notes?

Some products are built for this and some practices use them, but it is the highest-risk use in this plan. Treat it as a later pilot with its own consent, a signed data agreement from the vendor, and the clinician reading and signing every note.

What if a client asks whether we use AI?

Answer plainly and have the answer ready: which tools, for what, what never goes in, and how to opt out. A one-paragraph client notice on your website and intake pack stops most worry before it starts.

How do we stop clinicians pasting notes into free chatbots?

Give them a better route and a clear rule. Provide an approved tool for the admin jobs they want help with, explain why free accounts are off limits, and ask them to report slips without blame. Rules alone rarely hold.

Further reads

• AI for Pharmacies: The First 90 Days for an Independent Owner — A month-by-month plan in another regulated small business.
• AI Training for Pharmacists and Counter Staff: A Two-Session Plan — A two-session training plan you can adapt for clinicians.
• Therapist Termination Letter to Client: Wording and an AI Draft — Template wording that clinicians can adapt and review.
• AI for Care Homes: A Registered Manager's 90-Day Implementation Plan — A parallel 90-day plan from another care setting.
• AI Software for Psychologists: Practice Systems With AI Built In, Compared — SimplePractice, Carepatron and Jane compared for a solo psychologist: plan prices, what the AI adds, and the real monthly total for one clinician and three.
• AI Scribe for Counsellors: A Safe Trial Plan With Consent and Opt-Out — A six-week scribe trial for counsellors: who to include, consent and opt-out wording, comparing drafts with your own notes, and when to stop.
• AI for Psychotherapists: Plan and Fill a New Therapy Group — A planning and referral workflow for a new therapy group: outline, screening questions, invite emails, waitlist, and the fee-versus-capacity sums.
• AI in Psychotherapy Practice: Enquiry Replies and Fee Policies — An admin-only template workflow for first enquiries, fees and cancellations, with a crisis hand-off rule.
• Fertility Clinic Intake Form: Digital Version, AI Summary — A fertility clinic intake form built in stages: fields, consent, a clinician-ready AI summary prompt, and the data you should never collect.
• OpenAI for Medical Providers: Which Product Fits a Small Practice? — Which OpenAI products will sign a BAA, which will not, and how a solo clinician, small practice or developer should choose between them.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT Business pricing page, Claude pricing page, OpenAI data controls FAQ, Claude privacy centre article on model training, Microsoft Learn privacy page for Microsoft 365 Copilot and the Google Workspace generative AI privacy hub (all checked 4 November 2026).