AI for ABA Therapy Practices: Seven Admin Jobs and Three to Keep Human

An ABA therapy practice can safely hand seven admin jobs to AI: scheduling, authorisation tracking, parent logistics messages, intake paperwork, billing checks, staff credential reminders and enquiry replies. Treatment decisions, clinical judgement on data, and conversations about risk or progress stay with the BCBA. Every client detail needs a signed business associate agreement first.

Behavioural practices run on paperwork that is easy to get wrong and expensive when it is: units used against units approved, notes matching timesheets, schedules that survive a cancelled session. That makes admin a good target. It also makes the dividing line important, because the same tool that spots an expiring authorisation could, if misused, suggest a change to a child's programme. So the gate comes before any of the jobs: no client detail moves until the vendor has signed the right agreement for the exact plan you use.

Follow me on Instagram@sagnikteaches

The gate: a business associate agreement before any client data

If a tool will handle information that identifies a client or their family, a practice covered by HIPAA needs a business associate agreement (BAA) with that vendor: a contract making the vendor responsible for protecting the information on your behalf. The agreement is specific to a product and plan, which is where practices get caught.

Connect on LinkedInSagnik Bhattacharya

Here is where the three vendors most practices meet stand, from their own help and pricing pages:

Subscribe on YouTube@sagnikteaches

Vendors change these pages, so read the current version on the day you decide and keep a copy of the signed agreement. A practical rule: a tool without a BAA for your exact plan gets no client names, initials with dates of birth, funder details, diagnoses, or session content. It can still do green-zone work, such as drafting a job advert or a policy. For a list of vendors that sign, see the tutorial on which vendors sign a BAA.

An illustrative practice to hold in mind

For the examples below, picture a practice with 38 clients, 12 RBTs (registered behaviour technicians, who deliver most sessions), 3 BCBAs (board-certified behaviour analysts, who design and supervise each programme) and 2 office staff. Sessions run after school and on weekends, most clients have a funder authorisation covering a set number of 15-minute units over six months, and parents get a short update after sessions. All figures are made up for illustration. The jobs follow in order of how much time they usually return for the least risk.

Job 1: scheduling, cancellations and make-up sessions

An ABA schedule is a matching problem: client availability, therapist availability, travel or room limits, and continuity (a child should see the same RBT where possible). Cancellations leave holes that are costly to fill. A scheduling tool or a chat assistant used on non-identifying codes can propose a weekly grid, list which sessions lost cover, and suggest make-up slots.

Quick sum. With 38 clients averaging 12 sessions a month and a 10 per cent cancellation rate, there are about 46 cancelled sessions a month. If a tool helps fill a third of them from a make-up list, that is roughly 15 recovered sessions, around 30 hours of therapy at two hours each. Measure your own rates first; these are illustrations.

Illustrative prompt and output. "Here are weekly availability windows for six RBTs, coded R1 to R6, and eight client slots coded C1 to C8, each with a required two-hour block and a continuity preference. Propose a grid for Monday to Friday, list any slot with no cover and any client moved from their preferred RBT." The output is a table plus two lines: "C5 Thursday 16:00 has no cover; C3 moved from R2 to R4 on Tuesday." Those two lines are what the coordinator acts on. What you fix: any pairing the BCBA has ruled out for clinical reasons, and any slot that ignores travel time between homes.

Safe-use rule. Use client codes, not names, and keep the schedule's clinical continuity rules (who may work with whom) written by the BCBA. The shift scheduling tutorial shows the general method.

Job 2: authorisation tracking and the burn-rate alert

Funders approve a set number of units for a period. Running out early leaves a client without therapy while a reauthorisation is requested; running late wastes capacity. This job is a sum, and the sum is the whole point. The prior authorisation tracking tutorial for physical therapy uses the same logic for visit counts.

Worked example (illustrative numbers). A client is authorised for 2,600 units (15 minutes each) over 26 weeks, about 100 units a week. At week 18, the log shows 1,900 units used.

1. Average use so far: 1,900 ÷ 18 = about 105.6 units a week.
2. Units left: 2,600 − 1,900 = 700. Weeks left: 8.
3. At the current pace, 8 × 105.6 = about 844 units needed, so the shortfall is about 144 units, roughly 36 hours.

The 700 units left cover about 6.6 weeks at the current pace, so the alert for the office is "client X will run out about a week and a half early; start the reauthorisation or discuss reducing sessions". It does not say what the BCBA should recommend. The BCBA decides whether to request more units, with the clinical justification only she can write. A spreadsheet with a few formulas does this better than a chat assistant; ask the assistant to write and explain the formulas, then verify them on three clients by hand.

Safe-use rule. Calculations live in a spreadsheet or your practice system, not in a chat window. A person checks a sample of alerts each month against the funder's own portal.

Job 3: parent messages about logistics and routines

Parents want to know where the RBT is, whether a session is on, and what to prepare. Those messages are logistics, and a model can draft them: late arrivals, reschedules, reminders about materials, welcome messages for a new family. It should not describe the session, comment on behaviour or reassure about progress, because it did not see the session.

Illustrative prompt: "Draft a text for a parent saying the RBT will be 15 minutes late today because of a delay, asking them to reply if the new start time is a problem. Warm, two sentences, use [first name] and [RBT first name]." The sample output is exactly that. What you fix: any line promising a make-up session before you have checked the schedule.

Before and after for a BCBA's summary. The BCBA writes: "[first name] manded (asked) appropriately for breaks 8 of 10 opportunities today; two tantrums, both during transitions; we will keep the current plan." A tool reworded for a parent might produce: "[first name] asked for breaks nicely most of the time today and found two transitions tricky. We are continuing with the current plan." The BCBA checks that "most of the time" fairly reflects 8 of 10 and that "tricky" does not understate or dramatise. If the rewrite adds "great progress", she deletes it, because that is a judgement she did not make.

Safe-use rule. Messages about what happened in a session, or how a child is doing, come from the clinician. A tool may reword a BCBA's own summary in plain language, and the BCBA reads the rewrite before it is sent. The autism assessment clinic tutorial has a parent-update pattern for the waiting stage that you can adapt.

Job 4: intake paperwork and records requests

New families bring forms: consent, funder details, previous reports, school contacts. An assistant can build a checklist of what each family has returned, draft a polite chase email, and produce a cover note for a records request. It can also summarise which fields are missing from a form. The contents of reports and diagnoses are another matter: those go only into a contracted tool.

Example, a filled-in checklist. Family A has returned consent and the funder letter; still missing: previous assessment and school contact. The chase email lists those two items, the date they are needed and a reminder that therapy cannot start without the first. A before and after: the office used to track this in a notebook and send chases when someone remembered; now each new client has a seven-item list and a weekly sweep that drafts the chases, and the coordinator sends them.

Safe-use rule. The checklist uses client codes unless the tool is under a BAA. Reports and assessments are never pasted into an uncontracted tool.

Job 5: billing checks, notes against timesheets

Claims go wrong in boring ways: a note without a signature, units that do not match the timesheet, a session billed to the wrong authorisation. An AI-assisted check can compare exported data and flag mismatches. The principle is the same as in the physical therapy claim checks tutorial: let software find inconsistencies, and let a person decide what to do about them.

Illustrative mistake. An RBT's timesheet shows a session from 15:00 to 17:00 (8 units) and the note's end time is 16:45 (7 units). The check flags the one-unit difference, the clinician corrects the note after confirming with the RBT, and the claim goes out correct. Across 600 sessions a month, a one-per-cent error rate would be six wrong claims, each a possible denial or a repayment request.

Safe-use rule. A model never edits a note or changes units. It flags; the clinician or biller fixes the source record. Do arithmetic in a spreadsheet, and have a person verify a sample of flags each week.

Job 6: staff credential, supervision and training reminders

Practices carry a lot of dates: credential renewals, background-check renewals, supervision contact schedules, annual training, first-aid certificates. A tracker with alerts at 60, 30 and 7 days, plus a drafted reminder message, prevents the embarrassing day a therapist cannot work because a document lapsed. This job holds no client data at all, which makes it a good first job.

Example. An office manager pastes a list of staff initials and expiry dates into a spreadsheet and asks a chat assistant to write the alert formulas and a reminder template. The result: a column that turns amber at 60 days and red at 30, and a message that reads "Your [document] expires on [date]. Please upload the renewed copy by [date]." What you fix: anything that states requirements you have not confirmed; the dates and rules come from your credentialing body's own pages.

Safe-use rule. Requirements are checked against the official source. The tool writes reminders; it does not tell you what a credential requires.

Job 7: enquiry replies and waitlist messages

Families searching for ABA services often write to several practices. A prompt reply, a plain explanation of how intake works and an honest waitlist message keep families engaged. A chat tool can draft the templates; you supply the facts: funders you work with, how long intake usually takes, what happens next.

Before and after, illustrative. Before: "Thanks for your enquiry, we will be in touch," sent two days later. After, drafted from the facts sheet the same morning: "Thank you for contacting us about [first name]. We have a waitlist for after-school sessions at the moment. Intake starts with a phone call with our clinical team, then an assessment once funding is confirmed. Could we call you on [day] at [time]?" What you fix: any waitlist length the tool guessed, any funder it named that you do not work with, and any phrase that sounds like a promise about hours or outcomes. If the parent's message described the child's needs, keep those details in your inbox and give the tool only the question.

Safe-use rule. No promises about start dates, outcomes or funding, and no clinical advice. The reply invites a call; a person answers anything about a child's needs. Keep the approved facts on one sheet and write every reply from it.

The three to keep human

Seven jobs are good to give away. Three are better kept, and the practice should write them into its policy.

1. Treatment decisions. Setting and changing goals, selecting interventions, reducing or intensifying hours, deciding when to fade or discharge, and interpreting session data. These are the BCBA's responsibility. A model has not met the child and cannot be accountable. The tutorial on whether AI can write behaviour support plans shows how a regulator can view the same question, and the answer is cautious.
2. Clinical records of what happened. Session notes record what a clinician observed, did and measured. A tool may help with wording after the clinician has dictated, but the author reads and signs every line, and nothing is filed that the clinician did not witness.
3. Conversations about risk, incidents and setbacks. A challenging-behaviour incident, a safety concern, a regression, a discharge. These conversations need judgement, empathy and accountability, and sometimes a mandated process. A template may help a nervous clinician prepare notes; the conversation and the report are theirs.

The signal for all three is the same: if a wrong sentence could change a child's programme, safety or a family's trust, a trained person writes it.

Choosing the first two jobs

A practice should not turn on seven jobs together. Score them on three things: client data involved, minutes saved, and cost of a mistake.

Authorisation tracking comes early despite its risk, because the tool is a spreadsheet and the data is codes and numbers. If you want a wider view of what practice software offers, the practice software and seat-cost tutorials in the further reading are good places to compare.

A one-page rule sheet for the team

1. No client name, date of birth, funder number, diagnosis or session content goes into a tool without a signed BAA for that product and plan.
2. AI drafts admin. It never writes goals, interventions or clinical conclusions.
3. Every clinical record is written or fully reviewed and signed by the clinician who did the work.
4. Calculations (units, hours, dates) are done in a spreadsheet or practice system and sampled by a person.
5. Parents are told in plain words when a tool is used for messages, and can ask for a person.
6. Mistakes are logged and reviewed monthly.

What to do when a tool gets something wrong

Expect errors in the first month and plan for them. Keep a simple log with four columns: date, job, what the tool got wrong, and what caught it. A wrong make-up session offered to a family, a unit miscount flagged by the biller, a parent message that over-promised: each goes in the log. Review it monthly with the BCBAs in the room. If a mistake involved client information in the wrong place, follow your incident procedure the same day and check the vendor's deletion options. The log has a second use: it shows which instructions to tighten. In the third month, a practice that started with five errors a week and is now at one has evidence the controls work, and evidence is what lets you extend a tool to another job with confidence rather than hope.

How to tell after a month whether it worked

Measure the hours and the errors. For scheduling, count sessions recovered from cancellations. For authorisation tracking, count alerts raised more than two weeks before a gap. For billing checks, count mismatches found before claims went out. For parent messages, count replies asking "what does this mean?", a sign of unclear wording. For credentials, count lapses, which should be zero. If a job shows no improvement after four weeks, switch it off and revisit the instructions. A tool that saves a coordinator two hours but sends one wrong message a week has not paid for itself, so keep an error column beside the hours column.

Questions ABA practice owners ask

Can AI write treatment plans or session notes?

It should not write treatment plans, set goals or change interventions, which are the BCBA's decisions. Session notes record what a clinician observed and did, so the clinician must author them. A tool may tidy wording a clinician dictated, but a person checks every line before it is filed.

Do I need a business associate agreement?

Yes, before any tool handles client information. Without a signed agreement naming the product and plan, nothing that identifies a client or their family should go in. Many standard chat plans are not covered, so check each vendor's current page.

Which job should an ABA practice automate first?

Authorisation tracking pays back fastest because running out of approved units is costly and predictable. A spreadsheet with a simple forecast, or a rule in your practice software, catches it before it becomes a gap in therapy.

Can AI draft parent updates?

It can draft scheduling and logistics messages, and reword a BCBA's own summary in plainer language. It must not invent progress, describe behaviour it did not see or reassure a parent about outcomes.

Can a chat assistant calculate billing units?

It can help build a check, but do the arithmetic in a spreadsheet or your billing system and have a person verify a sample. Language models make arithmetic and rounding errors that cost real money on claims.

Further reads

• Best AI Software for Therapists in a Group Practice: Seat Costs — Seat costs of AI software for a group practice.
• Therapy Practice Management Software: What the AI Add-Ons Cost — What practice management AI add-ons cost.
• Medical Billing AR Follow-Up: Prioritise Aged Claims With AI — Prioritising aged claims once units and notes line up.
• Starting a Speech Therapy Private Practice: AI and the Admin Stack — An admin stack for a neighbouring therapy practice.
• Telling Patients You Use AI Note-Taking: Consent Wording That Works — Consent wording when a tool listens or drafts notes.
• Is an AI Scribe Worth It for a Small Private Clinic? — Judging a scribe on cost, time and risk.
• ABA Session Notes With AI: From Raw Session Data to a Draft Note — A practice-owner workflow for turning trial data into a draft ABA note: de-identify or use a BAA tool, a prompt, a worked example and sign-off.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Help Center article on business associate agreements; Anthropic privacy centre article on BAAs for commercial customers; Google Workspace help on health-data compliance. Unit and authorisation figures are illustrations, not payer rules.