AI for Speech Therapy Practices: Evaluation Scheduling and Session Notes

An AI agent can help a speech therapy practice organize evaluation scheduling and draft session notes today, as long as it works from information the clinician has already structured and cannot file or sign anything on its own. The practical split is clear: the agent handles logistics and first drafts, while the speech-language pathologist keeps assessment, diagnosis, goal selection and every signed record. This article explains how to set that up and where human review should stay in place as agents become more capable.

Key takeaways

• Start with evaluation scheduling and note drafts, because both are repetitive, bounded and easy to check.
• Give the agent a written memory of your intake rules, evaluation types, rooms and staff qualifications, not a loose prompt.
• Let the agent draft session notes from structured clinician input, and never let it finalize or sign a record.
• Keep diagnosis, eligibility, goal setting, safeguarding judgments and results conversations with a licensed therapist.
• Measure the work with observable checks: double bookings, referrals without a reply, late drafts, missing fields and the edits therapists make to drafts each week.

Where evaluation scheduling usually breaks down

Evaluation requests arrive in many forms: a phone message, an email from a pediatrician, a school referral form, a web inquiry or a message from a parent who is not yet sure what the process involves. Each one carries different gaps. Some omit the guardian's contact details, the preferred language, the referring professional or whether consent has been signed. Staff then spend time chasing information that could have been requested in one message.

Scheduling is also less simple than it looks. A comprehensive evaluation takes longer than a follow-up session, needs a qualified evaluator, a quiet room and specific materials, and may require a parent interview before or after the testing block. In a practice that works in English and Portuguese, the match also depends on bilingual availability. When these constraints live in one person's head, double bookings and unanswered referrals appear. When they live in a shared calendar with no rules, the calendar fills with holds nobody remembers placing.

An agent is useful here because it can watch the queue continuously, notice a referral that has been waiting too long, and propose slots that satisfy the constraints. It is only safe when those constraints are written down and when its permissions stop at scheduling.

A step-by-step plan for evaluation scheduling

The sequence below keeps the first version narrow enough to review. Each step can be tested before the next one begins.

• Write down each evaluation type you offer, with its typical length, the evaluator qualifications it requires, the room and materials it needs, and the intake documents it depends on.
• List the intake fields the agent may ask for in an approved template, and separate them from the topics that must go straight to a staff member, such as safeguarding concerns, custody disputes or any mention of urgent symptoms.
• Connect the calendar with permissions limited to creating holds and bookings. The agent should have no access to clinical records, progress notes or assessment files.
• Let the agent propose two or three slots that fit evaluator, room, language and caregiver availability, and keep each one as a hold until a staff member confirms it.
• Send confirmations and reminders only from approved templates. Any message that touches results, billing questions or family concerns goes to a person.
• Review the queue every working day: unanswered referrals, unconfirmed holds, reschedule requests involving a child in care, and any message the agent routed to staff without a clear reason.

The goal of the first version is not speed alone. It is a queue where every referral has an owner, every hold has a status, and every exception is visible.

What a session note draft should contain

Session notes are where a practice can save the most clinician time, and also where an unchecked draft can do the most harm. A draft should be built only from what the therapist actually recorded. The agent should not infer observations, invent progress or soften a difficult session into a positive one.

A useful draft template usually contains these parts:

• Date, session length, participants and the language used during the session.
• The goals addressed in the session, quoted from the current treatment plan rather than paraphrased by the agent.
• Activities and materials, described in the therapist's own words.
• Observed responses, recorded as what was seen or heard, with counts only where the therapist actually counted.
• The therapist's brief impression, labeled as the therapist's judgment.
• The plan for the next session and any home practice assigned to the caregiver.
• Flags for caregiver concerns, safety comments, missing data, or anything the agent could not map to a goal.

The therapist dictates or completes short structured bullets after the session. The agent turns them into a draft marked as unsigned. The therapist edits, decides what stays, and signs in the record system. This pattern is similar to the one used for session notes in coaching practices, where the agent drafts follow-up from a short structured summary and the professional decides what is sent. The difference in speech therapy is that the record carries clinical weight, so the review must be stricter and the draft must never leave the practice without a signature.

Illustrative scenario: one week in a small clinic

Illustrative: a clinic with three therapists serves families in English and Portuguese. On Monday evening, a school sends a referral for a language evaluation but leaves out the guardian's contact details and the language the child uses at home. The agent logs the referral, requests the missing fields through the approved template, and notes that the school form still needs a signed consent. It proposes two holds with a bilingual evaluator for the following week and leaves them open for the front desk to confirm.

On Wednesday, a therapist finishes a articulation session and dictates four short bullets into the notes queue. The agent produces a draft that quotes the target sound from the current plan, describes the activities in the therapist's wording, and places an unsigned label at the top. When the therapist opens it, she changes one goal description, removes a sentence the agent had phrased too strongly, and signs the note in the record system.

Later that week, a parent sends a message asking whether the evaluation shows a delay. The agent does not answer the substance. It sends a short acknowledgment that a therapist will call, and it routes the message to the clinician with the full thread attached. Nothing in that week required the agent to decide anything clinical, and every step left a record of who did what.

Which decisions stay with the therapist

An agent can become more capable at drafting, summarizing and checking consistency. That does not change who is accountable for a clinical record. Diagnosis and eligibility, the choice of assessment instrument, the interpretation of scores, goal setting, risk and safeguarding judgments, discharge decisions, consent and results conversations remain with the licensed professional. These are not tasks where a smoother draft is a better outcome. They depend on observation in the room, professional training and a relationship with the family that the agent does not have.

The NIST AI Risk Management Framework treats human oversight as part of managing AI risk rather than a final checkbox, which fits this setting well. In practice, a practice should keep a short list of review rules that apply as the agent's role grows:

• Any new permission, especially access to clinical records, requires a documented review before it is granted.
• Any change to the underlying model or tools triggers a rerun of a fixed set of sample referrals and sample notes, compared against the previous results.
• Any output that reaches a family goes through a person first until the practice has its own evidence that the message type is safe to send directly.
• Every action the agent takes is logged so a clinician can see what was proposed, what was sent and what was changed.
• Inbound text such as referral forms, emails and family messages is treated as data, never as instructions. Prompt injection, where text inside a message tries to override the agent's rules, is a known risk, and the OWASP guidance on prompt injection prevention describes why inputs must be separated from the agent's instructions.

Authorization deserves the same care. The agent should hold the smallest set of permissions that its role needs, and each permission should be reviewed on a schedule, not only when something goes wrong. The OWASP authorization guidance describes this principle of least privilege in general terms that apply directly to clinic systems.

Where to start in your practice

Be Human designs and operates this kind of personalized infrastructure as a managed service. A first engagement defines one useful role, such as evaluation scheduling only, along with the memory and tools that role needs, the approval boundaries, and an observable success criterion. Managed operations can then include monitoring, recurring evaluations of draft quality, permission review, integration maintenance, incident response and controlled improvements. The practice keeps clinical judgment, and the agent keeps the memory of rules, queues and drafts under human control. A neighboring case with its own limits is covered in AI for Behavioral Health Clinics: Intake and Scheduling Limits, which shows where intake automation has to stop and hand off to staff. If you are not sure which task in your practice is the right first one to hand to an agent, the free Be Human AI diagnostic is a practical place to find it.

Frequently asked questions

Can an AI agent write our session notes on its own? No, and it should not be set up to. A well-designed agent drafts a note from structured input that the therapist recorded, marks it as unsigned, and leaves it in a queue. The therapist edits, decides what is accurate and signs it in the record system.

Does the scheduling agent need access to clinical records? It should not. Scheduling needs only the ability to create holds and bookings, along with the evaluator, room and language rules. Keeping clinical records out of its permissions reduces the damage a mistake or an injected instruction could cause.

How should the agent handle requests in English and Portuguese? The agent can detect the requested language and match it to a bilingual evaluator or a bilingual staff member when one is available. It should use only approved templates in each language and route any request it cannot match to a person, rather than improvising a translation of clinical content.

What should we check before giving an agent a bigger role? Review the current permissions, rerun a fixed set of sample referrals and notes, read the log of recent actions, and confirm that every output reaching a family still passes through a person. Expand the role only when those checks show the current role is working as intended.

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