AI Agent for Speech Therapy Practice: Evaluations and Referral Paperwork

An AI agent for a speech therapy practice can coordinate evaluation scheduling and referral paperwork by reading incoming referrals, listing what is missing, proposing evaluation windows that fit approved clinician availability, and drafting the packet a clinician reviews. Every proposed booking and every outbound message waits for clinician approval before it is confirmed or sent. The agent handles routing and follow-up, while clinical judgment and the final commitment stay with the people responsible for the child or adult being evaluated.

Key takeaways

• Start with one bounded task, usually evaluation scheduling for new referrals, rather than the whole front office.
• Let the agent propose schedules and draft paperwork, but require clinician approval before any booking, family message or referral reply goes out.
• Give the agent only the memory, tools and permissions the task needs, and review them on a fixed schedule.
• Define one observable success criterion, such as how many referrals reach a clinician with a complete packet on first review.
• Keep a log of every proposal, approval and edit so the practice can explain how any plan was made.

Where an evaluation scheduling agent fits

Evaluation scheduling is a chain of small steps: a referral arrives, required fields are checked, consent is confirmed, a clinician with the right caseload is identified, a window is found, the family is contacted and reminders are sent. Each step is simple. The difficulty is that a busy practice handles them across inboxes, phone notes and calendars, and a referral can stall for days because one field was missing. An agent is useful at the coordination layer, where it keeps the chain moving and makes the gaps visible.

The division of work matters. Whether an evaluation is needed, which assessments apply, and which clinician takes the case are clinical decisions, and they stay with the clinical team. The agent prepares the information those decisions depend on. A deeper look at intake and scheduling for this kind of practice is in AI for Speech Therapy Practices: Evaluation Scheduling and Session Notes.

A simple decision guide helps staff agree on the boundary before anything is connected:

• The agent may match a referral to candidate windows drawn from a calendar the practice has approved for new evaluations.
• The agent may list missing referral fields and draft a request to the referring office for staff to review.
• The agent may prepare a short referral summary that states what is known, what is missing and what is unverified.
• Staff must decide whether an evaluation is appropriate, which clinician takes the case and whether the priority changes.
• Staff must handle anything touching diagnosis, coverage decisions, consent exceptions or clinical urgency.

Which referral paperwork can the agent prepare?

Paperwork is where most delays hide, so the agent's memory should hold the structure of a complete packet rather than the full contents of every file. A practical checklist for a new evaluation referral:

• Referral source, date received and the referring clinician's contact details
• The reason for referral, quoted as the referring office wrote it
• Consent and guardian authorization status, marked as received, missing or pending
• Coverage information, marked as provided, pending or not applicable, and confirmed by staff
• A list of prior records on file, identified by name, without copying their contents into the agent's memory
• Preferred language, communication needs and scheduling constraints stated by the family
• Open questions for the referring office, drafted for staff to send

Keeping prior records out of the agent's memory is a deliberate choice. The agent needs to know that a report exists and where it is stored. It rarely needs the full text of that report to offer a window or chase a missing signature. A smaller memory means less exposure if something goes wrong, and it is easier to explain to a privacy officer.

How do you set the approval boundary?

The approval boundary is the most important design decision in this kind of work. For every action, it answers one question: can the agent do this alone, can it prepare this for a person, or must it never do this at all. A workable process looks like this:

• List every action the agent could take, including reading, drafting, booking, sending and changing records.
• Tag each action as read only, draft or commit, and grant read access broadly while keeping commit access narrow.
• Require clinician approval for every draft that goes to a family, a referring office or a payer.
• Require a second approval for any booking that changes another clinician's calendar.
• Route exceptions, such as urgent referrals, unclear consent or conflicting requests, to a named staff member with a stated response expectation.
• Record who approved each plan, what was edited and when, so the log can answer an audit question without reconstruction.

Be careful that approval does not become a quick click. A clinician who approves dozens of drafts without reading them has not provided oversight. Keep batches small, show the source referral next to each draft, and make the differences from the agent's original proposal visible.

An illustrative scenario and how Be Human approaches it

Illustrative scenario: a referral arrives by email from a pediatric office, asking for a speech and language evaluation. The agent reads it and finds that the reason for referral is clear, but the consent field is blank and the coverage field says only that it is pending. It drafts a request asking the referring office for both items and proposes two candidate windows with the clinician whose caseload has been approved for new evaluations. The clinician edits the request, approves it and sends it. The agent marks the referral as waiting for consent and does not offer a window to the family until staff confirms. When the consent arrives, the agent updates the packet and asks the clinician to approve the final booking.

Be Human approaches this problem by defining one useful role first, then mapping the memory and tools that role needs, setting approval boundaries and naming an observable success criterion before anything touches a live calendar. In managed operations, the same team monitors the integrations, reviews permissions on a recurring schedule, responds to incidents and proposes controlled changes to the workflow, which the practice approves before they take effect.

Which permissions and safeguards matter most?

Access should follow the task. An agent that schedules evaluations needs read and write access to one calendar or a defined set of calendars, not to the practice's full records system. Delegated access should be limited to the actions the role needs, and those limits should be reviewed whenever the role changes. The OWASP authorization guidance is a useful checklist for this review: who can do what, on which resources, and under which conditions. Clinics that also run intake for other services face similar limits, which are described in AI for Behavioral Health Clinics: Intake and Scheduling Limits.

Referral documents are also untrusted input. A referral letter, an email attachment or a pasted note can contain text that tries to change how the agent behaves, whether by accident or on purpose. The agent should treat instructions inside a referral as content to be summarized, not as commands. The NIST AI Risk Management Framework offers a structure for identifying and managing these risks across the life of a system, from design through operation. Practices should also ask any vendor how it handles logs that contain personal information and who is able to read them.

Finally, the agent should never be the only record of a decision. Scheduling state belongs in the practice's own calendar and records system, with the agent's log as a supplement. If the agent is paused, the front desk must still be able to work from the same source of truth.

How do you know the agent is working?

Choose the success criterion before the agent is connected. For this workflow, a useful one is the number of referrals that reach the clinician with a complete packet on first review, counted each week. Supporting measures can include how often drafts are heavily edited before approval, how many bookings had to be corrected after confirmation, and how long approvals wait in the queue.

Build a small set of past referrals, de-identified, with the correct outcome recorded by a clinician. Run the agent against that set before any change, and again after each change to its instructions, tools or memory. Anthropic's guidance on evals for agents describes this pattern of defined tasks, graded outcomes and repeated runs, and it is a sound starting point even for a small practice.

Recurring review matters more than a single test. Referral forms change, calendars shift and staff habits evolve. A monthly review of the log, the success criterion and a sample of approved plans will show drift early. Include one clinician, one front desk lead and whoever operates the system in that review, so that decisions about changes are made by the people who will live with them. If you want to find the first task in your practice worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent book speech therapy evaluations without clinician approval? In a well-designed setup, it should not. The agent proposes windows from a calendar the practice has approved, and a clinician or designated staff member confirms each booking before the family hears about it. Any action that changes another clinician's schedule needs a second approval, so the agent reduces coordination effort while the practice keeps control of commitments.

Does the agent need access to clinical records to schedule an evaluation? Usually it needs far less than the full record. It needs to know which referral fields are present, whether consent is on file and which calendar windows are open. Keeping prior reports out of its memory and referring to them by name reduces exposure and makes privacy review simpler.

What happens when a referral is incomplete or the family cannot be reached? The agent should mark the referral as waiting, record what is missing and who was asked, and route the item to a named staff member. It should not invent missing information or keep contacting people beyond the limits the practice sets. A clear queue with a stated response expectation prevents referrals from sitting unnoticed.

How should a practice choose its first agent role? Pick a task that recurs, has a clear finish and can be checked against past examples. Evaluation scheduling for new referrals often qualifies, but the right first role depends on where your team loses the most time. Write the success criterion down before you connect any system, so the first review has a clear standard to judge against.

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