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

AI by industry

AI Agents for Eye Surgery Practices: Referral Intake and Scheduling Admin

How an eye surgery practice can use an AI agent for referral intake and scheduling admin, what it must never do, and what to prepare before booking a call.

Be Human typographic cover with the title: AI Agents for Eye Surgery Practices: Referral Intake and Scheduling Admin
Conceptual illustration created with AI · Be Human

An AI agent for an eye surgery practice is most useful when it handles administrative referral intake and scheduling admin: reading incoming referrals, requesting missing details, proposing surgery slots and routing anything clinical to staff. It should not triage symptoms, interpret examination findings or choose surgical dates on its own. Before a practice books a call, it needs written rules for what the agent may read, write and must never touch.

Key takeaways

  • Limit the first agent to referral intake and scheduling admin, with written rules for what it may never answer.
  • Map each referral type and each surgeon's booking rules before connecting any tool.
  • Start with read access, then allow hold-only booking, and keep final confirmation with named staff.
  • Define a fixed escalation route for urgent symptoms so the agent never improvises clinical judgment.
  • Choose one observable success criterion, such as whether referrals reach a coordinator complete and without duplicate holds.

Why referral intake is the right first role

Referrals to an eye surgery practice arrive in many shapes: faxed letters, PDFs, portal messages, emails from outside optometrists and general physicians, and phone notes taken by whoever answered. The front desk then spends its time chasing the same gaps: the patient's date of birth, which eye is affected, insurance identifiers, the referring clinician's callback number, and the stated reason for the referral. These tasks are repetitive, reviewable and bounded, which is what makes them suitable for an agent.

The clinical question of whether and when surgery is indicated belongs to the surgeon. The agent's job is to make each file complete and routable, not to judge it. A practice that starts with this boundary usually finds the first project easier to approve, because the risk is narrow and easy to audit. Our notes on how limits shape intake for behavioral health clinics follow the same logic in a different specialty.

Where the agent's authority should stop

A boundary sheet should answer these questions in writing before anyone configures a tool:

  • Which referral sources and formats the agent reads, and which it passes to a person without reading.
  • Which fields it may request from the referring office, and which it may never estimate or fill in from memory.
  • Which calendar blocks it can propose, and which only a named coordinator can place or release.
  • Which words, symptoms or messages trigger an immediate handoff, with the phone route written out.
  • What it keeps in memory, for how long, and who can delete it.
  • Which questions receive a fixed reply and which are forwarded without an answer.

Keep the agent away from questions about outcomes, medication changes, expected recovery or post-operative symptoms. A patient who writes that a painful, red eye has worsened after surgery should receive the fixed escalation reply and a phone route, not a reassurance composed by a language model. The boundary is not a technical detail. It is the part of the design that the practice's surgeons and compliance owner should approve in writing.

What to settle before booking a call

A practice that cannot answer the following yet is not ready for an agent. A good first call is one that helps produce these answers:

  • An inventory of referral channels: fax line, shared inbox, portal, voicemail and phone notes.
  • A list of referral types, each with the fields it requires before a surgeon can review it.
  • Each surgeon's scheduling rules: lead times, protected sessions, procedures that need a second approval and cancellation practice.
  • Named owners for approvals, exceptions and escalations, by role and by person.
  • The calendar and records systems in use, and who administers access to each.
  • A data map showing which personal and health information appears in referrals, where it is stored, and which privacy rules apply where the practice operates.
  • A baseline from sampling recent files: how long referrals take to reach a coordinator and how often they come back for missing details.

The baseline matters most. Without it, the practice cannot tell whether the agent improved anything, and the first review becomes a matter of opinion.

Scheduling admin without clinical decisions

The scheduling part of the role is narrower than it sounds. The agent may check availability in designated blocks, propose a tentative hold for an evaluation or a pre-operative visit, and draft the message to the patient. It may not place a surgical slot that requires surgeon sign-off, change the order of a waiting list, or move an existing patient. Permissions should reflect this: read access to the calendar, write access only to hold status in named blocks, and no ability to delete or reassign.

OWASP's authorization guidance is a useful reference here. It recommends granting only the actions each step needs and checking them on every request, rather than trusting the agent's own description of what it is allowed to do.

The agent proposes, the coordinator confirms, and the surgeon's team decides the clinical next step.

Illustrative scenario: a referral arrives by email from an outside optometrist asking for a cataract evaluation. The attached PDF lacks the patient's date of birth and does not say which eye is affected. The agent asks the referring office for those two fields, records the request in the file and places nothing on the calendar. When the answers arrive, it prepares a short summary for the coordinator and proposes a tentative hold in the evaluation block. The coordinator confirms the hold, and the surgeon's team reviews the referral in the usual way.

Referral content is also untrusted input. A PDF or email can contain text that looks like an instruction to the agent, whether by accident or by design. OWASP's prompt injection guidance treats external content as data rather than commands, which means narrow tools, approval steps for any write outside holds and a complete log of every action the agent takes.

How Be Human runs this as a managed service

Be Human designs the agent around one role, here referral intake and scheduling admin, and defines its memory, tools, channels, permissions and approval boundaries in writing before anything goes live. The managed part covers monitoring of the referral queue, recurring evaluations of the agent's drafts and holds against the boundary sheet, maintenance when a booking system or intake channel changes, periodic review of access rights and a defined response when the agent acts outside its scope. Changes are made in controlled steps, and the practice's staff keep the final say on every booking. The NIST AI Risk Management Framework offers a neutral structure for mapping these risks and assigning owners, and it is a sensible reference when the practice writes its own review routine.

How to judge whether the first role is working

Choose one observable criterion for the first role and review it on a fixed schedule. Useful options include whether referrals reach a coordinator complete, whether any surgical hold was created without a named approver, and whether urgent language was routed through the written escalation path. Review these weekly at the start and monthly once the pattern is stable. At this stage, a log that shows every exception is more useful than a target number, because it tells the practice where the boundary is being tested.

If you want to find the first task worth handing to an agent in your practice, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent answer patient questions about eye symptoms? No. The agent should not interpret symptoms, reassure patients or advise on urgency. It should recognize defined trigger language, send the fixed escalation reply with the phone route, and alert a named staff member, so that clinical judgment stays with people who are qualified to exercise it.

Does the agent need access to our full electronic health record on day one? Usually not. A first deployment can work with read access to the calendar and a narrow set of intake fields, with write access limited to tentative holds. Broader access can be considered later, after the boundary sheet has been tested and the access review has been done.

What should we have ready before the first call with Be Human? Bring a list of your referral channels, the fields each referral type requires, each surgeon's scheduling rules, the names of people who approve exceptions and a short description of the systems you use. A sample of recent referrals, with identifying details removed, makes the first conversation far more concrete.

How is this different from a general chatbot on our website? A website chatbot answers public questions, while an intake agent works on the practice's own files, calendar and channels, under permissions and approval steps that the practice defines. It keeps its memory and operating context under human control and is reviewed on a schedule, so its behavior can be explained and corrected.

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 →