AI for an ABA Therapy Clinic: Authorization Tracking and Session Notes

An AI agent for an ABA therapy center can track authorization balances, flag renewals before units run out, and draft session notes from data staff already record. The clinical content stays with the licensed clinicians who delivered the service, so the agent prepares, reminds and routes while people decide and sign. This playbook explains where to draw those lines and which workflow to hand over first.

Key takeaways

• Start with authorization tracking, because exhausted or expired units create gaps that show up in the schedule long before they show up in billing.
• Treat session notes as drafts. The agent assembles structured data into a readable draft, and the clinician who ran the session edits and signs it.
• Write approval rules before connecting any tool, listing what the agent may draft, what needs clinical sign-off, and what it must never do.
• Treat uploaded payer letters and referral documents as data, not instructions, and keep a record of every action the agent takes.
• Define one checkable success criterion for the first role, then review it on a schedule instead of trusting the agent's own summary.

Why authorization tracking is the first workflow to scope

In most ABA centers, an authorization is a document, a register entry and a set of dates that several people touch. A payer approves a number of units for a service type over a defined period. Staff schedule against that approval, deliver sessions, and eventually request a renewal or an increase. When the balance runs low or the end date passes without a request, the center either pauses services or delivers them without coverage, and both outcomes are expensive to repair.

Authorization tracking is a good first role because the rules are explicit, most inputs are already structured, and a mistake can be caught by a person before anything leaves the building. The agent's useful work is arithmetic and attention. It compares approved units with scheduled and delivered units, projects when each balance will run out at the current caseload, and lists authorizations that need a renewal packet inside a set window. It does not decide whether a client needs more hours. That is a clinical and payer conversation, and it belongs to people. Teams that already handle similar records in intake work often find the pattern familiar, as described in AI for behavioral health clinics: intake and scheduling limits.

What the agent reads, writes and remembers

Define the access boundary before writing any instructions. For this role, the agent needs read access to the authorization register, exports of scheduled and completed sessions, and the folder where approved payer letters are stored. Its write access should be limited to a staff task queue, draft renewal packets saved for review, and a weekly exceptions report. Clinical records should stay read-only.

• Reads: the authorization register, with approved service types, unit counts and date ranges.
• Reads: scheduled and completed sessions by service type, so balances reflect delivery and not only booking.
• Reads: payer approval letters and requirements that staff have filed for the agent to use.
• Writes: task queue items for the scheduler, the billing lead and the assigned clinician.
• Writes: draft renewal packets in a review folder, never directly into a payer portal.
• Remembers: the center's written procedure for each payer, maintained by staff and versioned, so the agent works from documented rules rather than guesses.

The memory layer matters most here. If a payer requires a specific attachment for a reauthorization, that requirement should live in a procedure document the billing lead owns, not inside a prompt that nobody reviews.

Session notes: drafting from data without deciding the content

A session note is a clinical record and a billing support document. It normally identifies the client, date, time and provider, names the service delivered and the goals addressed, summarizes the data collected, and records the clinician's assessment of progress and next steps. Much of that is structured data entered during or after the session. An agent can assemble those parts into a clear draft and reduce the formatting work the clinician faces at the end of a long day.

What the agent should not do is interpret behavior, infer progress, or fill a gap with a plausible sentence. Missing fields should appear as visible prompts, not as smoothed-over prose.

Illustrative scenario: A clinician finishes a one-to-one session and saves a data sheet in which two planned trials have no entries. The agent drafts the note with the completed trials, marks the two blanks as needing entry, and adds a task to the clinician's queue. The clinician fills in what happened, corrects anything wrong, and signs. The agent never writes a reason for the blank trials.

Approval rules that keep clinical judgment with licensed staff

Approval rules are the most important part of this playbook. Write them as a decision guide the team can read in a few minutes, and name an owner for each category. Apply least privilege: the agent should hold only the permissions its role requires, and those permissions should be reviewed whenever the role changes, consistent with the [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html). The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) is a useful reference for documenting how these risks are identified, owned and reviewed.

• The agent may draft, remind, calculate, compare and flag without approval, because these actions do not change the clinical or billing record.
• A licensed clinician must review, edit and sign any session note content, goal wording, progress summary or change to a treatment plan.
• A billing lead or clinical director must approve any payer submission, unit request, code selection or appeal before it leaves the center.
• Family questions about clinical matters go to a licensed staff member. The agent may route them but never answers them as final.
• The agent may never sign a note, unlock a locked note, or edit a record after the clinician has signed it.

Log each approval with the name of the person and the time it was given. When payer requirements change, update the decision guide the same week, not at the next annual review.

Approval rules belong in a document a clinician can read, not only in a configuration file an engineer can edit.

Handling documents safely and keeping a record of every action

Payer letters, referral packets and intake forms are written by outside parties. Some of them may contain text that reads like an instruction, such as a line telling a system to skip a review step. Prompt injection is a known risk for any agent that reads documents. The [OWASP prompt injection prevention guidance](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) recommends treating document content as untrusted input and keeping the agent's operating instructions separate from it. In practice, the agent extracts fields from a document and never acts on directives written inside one.

Keep a record for every action: what the agent read, what it drafted, which task it created, and which person approved it. Logs of each step make it possible to answer what happened and why without reconstructing events from memory. The same discipline applies to access, documentation and handoffs in other regulated settings, such as the record controls described in AI agents for clinical research sites.

Minimize the personal health information the agent sees. A renewal task needs the client identifier, service type, unit balance and date range, not the full clinical history. Confirm the access list with your privacy or compliance lead before connecting any system, and follow the privacy rules that apply where your center operates.

How a managed approach handles this in practice

Be Human scopes a first engagement around one useful role, such as authorization tracking or session-note drafting. The engagement defines the memory and tools that role needs, the approval boundaries, and one observable success criterion. The agent's identity, memory, channels, tools, permissions and operating context stay under the center's control. Managed operations can then include monitoring of the task queue, recurring evaluations of drafts against a rubric written by clinicians, permission reviews, integration maintenance, incident response and controlled changes to procedures. The service is delivered remotely, in English or Portuguese.

Measuring whether the agent is working

A success criterion should be something a manager can check without trusting the agent's own report. For authorization tracking, that might mean every authorization with a renewal window appears in the weekly report before the window opens, and the billing lead confirms the list matches the register. For notes, it might mean a sample of drafts each week is scored by clinicians for accuracy against the session data, with every error recorded as a test case. The [OpenAI evals guide](https://platform.openai.com/docs/guides/evals) describes this general pattern of testing a system against a fixed set of cases, which suits clinical work where small errors carry weight.

Review the approval rules on the same schedule. If clinicians keep editing the same sentence in every draft, the template needs work. If the billing lead reports that renewal packets arrive without a required attachment, the payer procedure document needs an update before the next cycle.

Choosing the first task

Choose the task with the clearest inputs, the most visible cost of delay and the easiest human check. For many ABA centers that is authorization tracking, with session-note drafting as the second role once the team trusts the review loop. Write the approval rules before anything is connected, name the owner of each rule, and agree on the success criterion before the first draft is produced. Take the Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Can an AI agent submit authorization requests to insurance payers? In the design described here, no. The agent prepares renewal packets and places them in a review folder, and a billing lead or clinical director submits after checking them. This keeps the submission decision and its accountability with named people.

Can an AI agent write session notes for ABA therapists? It can assemble a draft from structured session data, but the clinician who delivered the service must review, correct and sign the note. The agent should not invent behavior descriptions or progress statements to fill gaps. Missing fields should appear as visible prompts for the clinician to complete.

What data should we give an agent before starting? Give it the minimum needed for the first role: the authorization register fields, scheduled and completed session counts by service type, and the written procedure for each payer. Avoid sharing full charts when a task needs only a few fields. Confirm the access list with your privacy or compliance lead before connecting any system.

How do we know whether the agent is reliable enough to keep using? Judge it against a success criterion you can check independently, such as whether every renewal window appeared in the report on time. Review a sample of drafts each week against the session data and record every error. Re-review permissions whenever the role or your payer requirements change.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.