Skip to content
be-human.ai
← Back to the journal

AI by industry

Superintelligence Prep for Speech Therapy: Referrals and Authorizations

How an AI agent can track speech therapy referrals and insurance authorizations, which checks it runs, and where clinicians keep every decision on care.

Be Human typographic cover with the title: Superintelligence Prep for Speech Therapy: Referrals and Authorizations
Conceptual illustration created with AI · Be Human

An AI agent for a speech therapy practice can run the administrative tracking behind referrals and insurance authorizations: logging each referral when it arrives, checking that required documents are present, following pending payer requests, and flagging approvals or renewals before they lapse. It works from a structured record and a short list of permitted actions. Decisions about whether a referral fits a client, how much therapy is needed, and what a request should say clinically stay with the treating clinician.

Key takeaways

  • A referral agent is mainly a tracking system: one structured record per referral, with source, scope, documents, payer status and the next action.
  • Authorization work is largely assembly. The agent can find gaps and draft requests, but a clinician approves any clinical content before anything is sent.
  • Clinical decisions about eligibility, the service plan, frequency and justification stay with licensed clinicians and are never delegated to the agent.
  • Start with one referral type and one payer, with permissions limited to reading records and writing drafts and staff tasks.
  • Judge the agent by an observable criterion, such as whether every new referral has a complete file or a named open item by the end of intake day.

Why referrals and authorizations stall in speech therapy practices

Most delays in this workflow are visibility problems rather than effort problems. A referral can arrive by fax, portal upload, email, or a phone call from a physician's office, and each channel puts information in a different place. Staff then carry status in a spreadsheet, a shared inbox, and memory. The stalls tend to happen at predictable points: an order missing a required element, a coverage detail that does not match the intake form, a first visit booked before approval is confirmed, or a renewal request that should have gone out before the approved visits ran out.

The evaluation side of this work is covered in AI Agents for Speech Therapy: Evaluations and Insurance Authorizations. The point here is narrower: what the tracking layer should record, and what it should never decide on its own.

What an agent should track for each referral

Treat each referral as one record with a fixed set of fields. Free-text summaries are useful for reading, but staff need to see missing items at a glance. A workable record includes:

  • Referral source, date received, channel, and the ordering provider's contact details
  • Scope of the order: services named, frequency if stated, and any expiration on the order itself
  • Required documents, marked present or missing: signed order, evaluation report, intake form, coverage copy
  • Authorization status for each payer: not started, drafted, submitted, pending, approved, denied, or expired
  • Approved visits used compared with approved visits remaining, and the date a renewal request must be prepared
  • Open questions with an owner and a due date, such as a front desk call or a clinician sign-off

The status vocabulary matters more than it seems. If one person writes pending and another writes waiting on payer, the agent cannot count reliably. Agree on a short list of statuses with the clinical lead before anything is automated, and store that list in the agent's memory with the date it was last confirmed. For the session-notes side of the same practice, see AI for Speech Therapy Practices: Evaluation Scheduling and Session Notes.

Where clinicians keep decisions on care

The practical test for any step is whether it changes care or commits the practice to something. If it does, a licensed clinician decides. If it only moves information, the agent can act within limits. The guide below is a starting point for the practice's own policy, not a replacement for it.

  • Whether a referral is appropriate for this client: the clinician decides, and the agent only lists what is missing.
  • Which services, frequency, and goals go into the plan of care: the clinician decides and signs.
  • The clinical justification in an authorization request: the clinician writes or approves the final text, and the agent assembles the source notes.
  • Whether to submit a request: a named staff member submits after review until the practice has audit evidence to widen that permission.
  • Whether to accept a denial, appeal it, or change the plan: the clinician decides, and the agent prepares the packet and the deadline.
  • Routine reminders to families or the front desk once an approved plan exists: the agent may send them from approved templates and log each one.
  • Changes to diagnosis or billing codes: the clinician or billing lead decides, and the agent flags inconsistencies between documents.

Clinics that handle intake for other health services face the same boundary, as described in AI for Behavioral Health Clinics: Intake and Scheduling Limits. The NIST AI Risk Management Framework is a useful reference here, because it asks organizations to name who is accountable for each AI-supported function and how people oversee it. Write that accountability into the agent's instructions and into the practice's procedures, so a new staff member can see it without asking.

Illustrative scenario: a referral arrives by fax late in the afternoon. The agent creates the record, notes that the signed order is present, and marks the coverage copy as unreadable. It drafts a message to the front desk asking for a clearer copy and does not book a first visit. The next morning, the clinician reviews the scope, confirms it fits the client, and approves a draft authorization request, which staff submit through the payer portal. Later, the agent notes that the approved visits are nearly used and asks the clinician whether a renewal should be prepared. The clinician writes the progress summary, and the agent assembles the packet around it.

A first rollout plan for referral tracking

A first rollout should be small enough for people to check by hand. The sequence below keeps the scope narrow and the review human.

  • Map every channel through which referrals reach the practice, and name the person who handles each one today.
  • Choose one referral type and one payer to start with, not the whole intake queue.
  • Agree on record fields and status vocabulary with the clinical lead and the front desk lead, then write them into the agent's memory with an owner and a review date.
  • Give the agent read access to the referral folder and the scheduling calendar, and write access only to draft records and staff tasks.
  • Write the approval rules as a short list of actions, each marked automatic, draft for review, or clinician only.
  • Replay a set of past referrals whose outcomes are known, and compare the agent's flags with what staff actually found.
  • Turn on live tracking for the chosen referral type, review every exception each day while the workflow settles, and change one rule at a time.

Set the success criterion before going live. A useful one is observable and small: every new referral has either a complete file or a named open item by the end of intake day. Avoid criteria that need a dashboard nobody reads.

How should the agent handle permissions, memory, and tools?

Referral documents come from outside the practice, and their text is uncontrolled. A faxed order or an uploaded PDF can contain sentences that look like instructions to a model. The OWASP guidance on prompt injection recommends treating retrieved content as data, separating it from instructions, and limiting what a model can do even when it has been manipulated. In practice, the agent should summarize document text and never act on it as a command.

Authorization belongs in the system, not in the prompt. The OWASP authorization guidance favors explicit rules over implicit trust, so each action the agent can take should map to a specific permission held by a specific role, with the account owner reviewing those permissions on a schedule. Access to client records should follow the minimum the task requires.

Memory needs the same discipline. Payer requirements, document checklists, and the practice's own rules change, and an agent that remembers an old requirement as current produces errors that look correct. Store each rule with its source, the date it was last verified, and the person responsible for it. Connecting tools through a defined interface, such as the Model Context Protocol, makes the list of available actions explicit and reviewable, which is what a permission review needs.

Every action should leave a record: which client file it touched, what it did, and who approved it. Observability concepts from OpenTelemetry, such as traces and structured logs, are a useful model, because a later review has to reconstruct the sequence of events rather than rely on memory.

How do you measure whether the agent is doing its job?

Evaluation for a tracking agent is mostly about exceptions. Build a small set of past referrals with known outcomes, including incomplete ones, duplicates, referrals with conflicting coverage details, and renewals that were late. Run the agent against that set after every change to its instructions, tools, or memory. The OpenAI evals guide describes the general pattern of fixed tasks, graded outcomes, and repeated runs, which suits this kind of check.

In production, track three things: referrals that reached a first visit before authorization was confirmed, renewals prepared after the approved visits ran out, and drafts the clinician rejected along with the reason given. The first two point to workflow failures. The third shows where the agent misreads how the practice actually works. Review these in a fixed meeting, and change the rules instead of asking staff to remember the workaround.

How Be Human runs this under managed operations

Be Human designs the first engagement around one useful role, here a referral and authorization tracking agent. That covers the record fields, the memory for payer rules and document checklists, the tool permissions, the approval boundaries, and one observable success criterion. Managed operations then include monitoring of exceptions, recurring evaluations on the practice's own past cases, maintenance when a portal or scheduling integration changes, permission reviews, incident response if something is submitted in error, and controlled changes when a payer or the practice's policy changes. Clinical decisions remain with the practice's clinicians throughout.

If you want to find the first referral or authorization task worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent submit authorization requests to a payer on its own? Some portal integrations make this technically possible, but a speech therapy practice should start with drafting and staff submission. Submitting is an action that commits the practice, so keep it behind a named person until the audit trail and evaluation results support a wider permission. Widen one action at a time and record each change in the permission review.

What should happen when a referral is missing a required document? The agent should mark the gap on the record, name the owner who can resolve it, and draft the request to the referring office or the family. It should not guess the missing content or book a first visit that depends on it. The item stays open until a person confirms it is resolved.

How should the practice protect client privacy when an agent reads referral documents? Limit access to the referral folder and calendar, store only the fields the workflow needs, and keep an audit log of who and what accessed each record. Check the obligations that apply where the practice operates, including Brazil's data protection law administered by ANPD for practices in that country. Complete the privacy review before the first live record, not after.

How long before we know whether the agent is working? Define the success criterion before launch, so the answer does not depend on how the first weeks feel. Run the replay of past referrals first, then review the exception list daily while the workflow settles. The pace depends on referral volume and on how many payers are in scope, so start with one referral type and one payer.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

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

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →