Superintelligence Prep for Occupational Therapy: Referral Intake Admin
Define the referral intake admin an AI agent can handle now in an occupational therapy clinic, what staff keep, and when a scoping call makes sense.

An AI agent for an occupational therapy clinic can handle referral intake admin today when the work is bounded: logging each incoming referral, checking that required fields, consents and supporting documents are present, drafting acknowledgments and scheduling options, and routing anything that needs a person. Clinical judgment about eligibility, goals, priority and clinical fit stays with licensed staff. The practical first move is to define one intake lane with a named owner, written approval boundaries and a success check, rather than a general assistant for the whole practice.
Key takeaways
- Start with one intake lane, such as new referrals from a single source, and leave the rest of the clinic unchanged until that lane works.
- Let the agent collect, check, draft and route. Clinicians decide eligibility, priority, clinical fit and anything that changes care.
- Write the approval boundaries before any build: what the agent may send, what it may book, and what always waits for a staff member.
- Treat referral documents as untrusted input, because they can contain text that tries to steer an automated tool, so permissions must stay narrow.
- Book a scoping call when the workflow spans several systems, payer rules differ by referral source, or the clinical sign-off cannot yet be described in writing.
What an intake agent can hold today
Referral intake is mostly structured repetition with a few judgment points. A referral arrives by fax, portal, email or phone note. Someone opens it, identifies the patient and the referring provider, checks that the order is complete, confirms contact details, looks for the consent and demographic forms the clinic requires, and decides which queue it belongs in. Most of these steps are rules the clinic already applies. An agent can apply them consistently if the rules are written down and the systems it touches are reachable through permissions the clinic controls.
What an agent should not decide is whether a patient needs occupational therapy, how urgent the case is in clinical terms, or whether a referral fits the caseload of a particular therapist. Those are professional judgments tied to licensure and to the clinic's responsibility for care. Design the agent so it presents the facts a clinician needs, not a recommendation dressed up as a decision.
Good early tasks share three traits: the input is predictable, the output can be checked quickly, and an error can be caught before a patient is affected. Acknowledgment messages, missing-document requests and held scheduling options usually meet those conditions. Clinical summaries, priority labels and anything that explains what treatment will involve usually do not, at least in a first version. For a neighboring example of insurance-related admin work, see AI agents for speech therapy evaluations and insurance authorizations.
Separating admin from clinical judgment
A useful exercise is to split every intake step into three columns: work the agent may complete, work the agent may prepare for staff, and work only a licensed person may decide. Label the columns in the clinic's own words so that front desk staff and therapists read the same document.
The first column usually includes logging, extracting fields, checking completeness against a list, sending approved acknowledgments and updating queue status. The middle column includes drafting a scheduling offer, summarizing referral text for the therapist without interpreting it, and proposing a queue based on rules the clinic has approved. The third column includes eligibility, priority, goal setting, declining a referral and any message that explains clinical reasoning.
Edge cases need explicit rules. A referral naming a condition the clinic does not treat should go to a person, not into a standard decline template. A note saying the family is distressed or asking a clinical question should leave the automated path immediately. Writing these rules before the build is what makes the agent safe to run, and it is also what lets staff trust the output. The same separation between administrative flow and clinical decisions appears in AI for behavioral health clinics, where intake and scheduling need clear limits.
A step-by-step plan to define the intake lane
Use this sequence before anyone configures a tool. Each step produces a written artifact that a manager can review and change.
- Step 1: Choose one referral source or one referral type. Name the person who owns that lane and the person who approves changes to its rules.
- Step 2: List every field a complete referral must contain, and every document required before scheduling, with the system that holds the source of truth for each.
- Step 3: Map the current path from arrival to first appointment or first decision, and mark where referrals wait and why.
- Step 4: Mark each step as agent, agent draft, or staff only, using the three-column split above.
- Step 5: Define the systems the agent may read, the actions it may take, and the actions it may never take, such as sending clinical content or changing a priority.
- Step 6: Define a success check you can observe without the agent, such as how many referrals are sent back to the referrer for a missing item, compared with the same measure before the change.
- Step 7: Run the lane in draft mode, where the agent proposes and staff approve, and review every disagreement before any permission expands.
Treat every referral as untrusted text. A referral can include instructions, pasted messages or odd formatting that an automated tool might read as commands. The agent should be able to read that text, but it should not gain new permissions because of it. The [OWASP guidance on prompt injection prevention](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) describes the failure mode in detail, and the answer is architectural: narrow tools, explicit approval steps and no free-form actions on patient records.
Illustrative scenario: a referral that arrives incomplete
This scenario is illustrative and does not describe a real clinic. A hand therapy referral arrives through the clinic's portal with the referring provider's name and a diagnosis, but no contact phone number and no signed consent form. The agent logs the referral, records the two missing items, and sends the approved request to the referring office. It places the referral in a held queue with the missing items listed, and it offers no appointment times, because the clinic's rules require a complete file before scheduling.
Later, the referring office replies with a phone number and a note saying the patient would like a call from a therapist before booking. The agent does not interpret that note or answer it. It flags the message to the intake coordinator with the original text, the time it arrived and the reason it left the standard path. The coordinator calls the family, and the therapist decides whether the referral fits the caseload. The agent's job ended at accurate routing and a clear handoff, which is the outcome the plan was built to produce.
How Be Human's managed approach handles intake
Be Human starts with one intake lane and documents it as the fields the agent reads, the checks it runs, the drafts it may prepare and the actions that stay with staff. The agent works through scoped permissions to the referral inbox, scheduling tool and records system, and each permission is logged and reviewed on a fixed schedule. Monitoring watches for failed checks and unusual routing, incident response has a named person, and recurring evaluations compare the agent's outputs with the clinic's rules when forms or referral sources change. Each change to what the agent may do is made in a controlled step and approved by the clinic.
When a scoping call makes sense
A scoping call is worth booking when the referral flow crosses several systems, when authorization or payer requirements differ by source, when more than one clinician needs different intake criteria, or when nobody on staff can yet write the approval boundaries. The call should produce the lane definition: the first task, the owner, the boundaries and the success check. If you can already answer the seven steps above, you may be ready sooner than you expect. If you want a starting point, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Can an AI agent decide which referrals an occupational therapist accepts? No. The agent can check that a referral is complete, log it, and route it to the queue the clinic has defined, but acceptance depends on eligibility, clinical fit and caseload, which are professional decisions. Those decisions stay with licensed staff, and the agent should present facts for review rather than a verdict.
Does an agent replace the front desk or the intake coordinator? Usually not. It takes over repetitive checks, acknowledgments and status updates, which frees staff for calls, exceptions and conversations that need judgment. The people who handle those exceptions remain the owners of the lane.
What should a clinic write down before starting? Write the referral fields and required documents, the three-column split of agent, draft and staff-only work, the systems the agent may access, and the success check. Also name who approves rule changes and who responds when something goes wrong. Those documents are useful even if you never automate.
How do we know the agent is working as intended? Compare its output with the success check you defined before the build, and review disagreements between the agent's proposals and staff decisions each week at first. Monitoring should flag failed checks and unusual routing, and recurring evaluations should confirm that behavior still matches the written rules after forms or referral sources change.
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.


