AI for Eye Doctor Offices: Surgery Scheduling and Recall Admin
Which eye practice admin queues an AI agent can run, including surgery scheduling and recall outreach, and which clinical questions must route to clinicians.

An AI agent for an eye doctor office can run the administrative queues that repeat every day: surgery date holds, pre-operative logistics reminders, recall outreach for routine exams, reschedule requests, and status checks on referral and authorization paperwork. Clinical questions, such as whether a new visual symptom needs same-day attention or whether a patient is ready for surgery, belong to clinicians and should reach them with a clear summary. The practical first step is to choose one queue, write down what the agent may and may not do, and measure whether that queue gets cleared.
Key takeaways
- Start with one administrative queue, such as surgery date holds or recall outreach, rather than the whole front desk.
- Write the clinical routing rules with the clinicians before the agent sends a single patient message.
- Give the agent only the memory, tools and permissions that its queue requires.
- Judge the agent on a success criterion that staff can check each week, such as fewer recalls left open at month end.
- Change scripts, permissions and integrations only through a reviewed process, never ad hoc.
Why eye practices suit a narrowly scoped agent
Eye practices run on due dates. Glaucoma monitoring, injection follow-ups, post-operative checks, contact lens refits and routine refraction visits each create a future obligation that someone must notice, offer and confirm. Many offices track these in practice management reports, spreadsheets or the memory of a long-serving coordinator. When the volume grows, the work slips quietly, and patients drift out of care without anyone deciding that they should.
That pattern is well suited to a scoped agent. The rules for who is due and which slots may be offered can be written down, and the work is repetitive. The risk sits in a different place: any message that sounds like medical advice. So the design question is not whether an AI can talk to patients. It is which messages carry no clinical judgment and can therefore be handled by a process the practice controls.
Which administrative queues an agent can run
A well-scoped deployment usually starts with some of the following queues:
- Surgery date holds: offering dates from an approved list, recording a tentative hold with an expiry, and tracking whether the patient confirms.
- Pre-operative paperwork: checking which consent forms and instruction acknowledgments are on file and sending the approved reminder text for what is missing.
- Recall outreach: identifying patients whose clinician-defined follow-up interval has passed, offering open slots, and logging each outcome.
- Reschedule and cancellation intake: capturing the request, proposing approved alternatives, and notifying the surgical coordinator the same day.
- Referral and authorization status: reporting which items are missing from a file and drafting requests that staff review before anything is sent.
- Waitlist and no-show follow-up for routine visits.
Each of these has a clear input, a bounded set of actions and an output that a coordinator can verify. That is what makes them suitable for delegation. Anything that requires reading a chart and deciding what it means falls outside this list.
What stays with clinicians
The routing rule should be written in plain language and signed off by the clinicians who own the patient relationship. A practical starting list includes the following triggers:
- Any new or changing visual symptom, including flashes, a new curtain or shadow, sudden blur, or pain with redness.
- Questions about eye drops, dose changes or interactions with other medication.
- Post-operative complaints of any kind, including discomfort that the patient describes as worse than expected.
- Questions about whether a patient is a candidate for a procedure or what a result means.
- Requests for records, diagnoses or test interpretations.
- Any message that mentions an emergency, in which case the agent gives the approved emergency wording and does not continue the administrative flow.
When a trigger fires, the agent should stop the administrative task, send the approved safety text, and notify the right staff member with a short summary of what the patient said. It should not try to reassure the patient or guess at an answer. Clinicians should review the routing list on a schedule, because new services and new protocols change what counts as clinical.
Surgery scheduling: holds, preparation and cancellations
Surgery scheduling is where administrative errors are most expensive, because a wrong date or a missing form can cancel a full room. The agent's role is to keep the chain of steps moving and to show staff exactly where each case stands.
Illustrative scenario: a patient is tentatively held for a cataract surgery slot. The agent confirms the date by message, sends the approved pre-operative instruction text, and sees that the acknowledgment and one consent form are still missing. Two days later it sends the approved reminder, and the coordinator receives a short list of cases with missing items. If the patient replies that they cannot attend, the agent records the reason, notifies the coordinator, and offers only the alternative dates that the surgeon's team has already approved.
A step-by-step plan for this queue looks like this:
- Clinicians define the rules for each surgeon, room and day, including the slots that may be offered and how long a tentative hold lasts.
- The agent offers only open slots from that list and records each hold with its expiry.
- The agent checks the required forms and acknowledgments and marks each one as received or missing.
- The agent sends approved preparation text and logs when the patient acknowledges it.
- For cancellations or reschedules, the agent captures the request, notifies the coordinator, and releases the slot according to the written rule.
- Staff review a daily list of open items, which is the only output they need to act on.
Recall admin: due lists without clinical judgment
Recall work depends on two inputs that must come from the clinical side: the follow-up interval for each service and the message wording that patients will receive. The agent should not decide an interval based on what a patient describes. If a patient's reply to a recall message mentions a symptom, the reply moves to the clinical routing queue.
Within those limits, recall outreach is repetitive and easy to lose. The agent can identify who is due, send the approved message through the channel the patient prefers, offer open slots, and record every outcome, including refusals. Opt-outs should take effect immediately and be stored where every future message checks them first. Staff then see a short weekly view of who was contacted, who booked, and who still needs a call.
This is a good place to compare how other specialty offices structure recall work. The dental and chiropractic versions of the same problem share the same core rule: outreach follows a clinician-defined schedule, and anything clinical is handed back. The dental office recall workflow and the chiropractic intake and recall approach show how the same boundary is drawn in adjacent fields.
Permissions, memory and an audit trail
An agent that touches schedules and patient messages needs tight permissions. Give it read access to the schedule and the queue it works, write access only to holds and approved message templates, and no access to diagnoses or clinical notes. Authorization should be enforced by the systems the agent calls, not by instructions in its prompt. The OWASP authorization guidance describes the same principle: check permissions on the server side for every action, and grant only what a task needs.
Memory deserves the same discipline. The agent should remember operational facts such as preferred channel, opt-out status, open tasks and the date of the last approved message. It should not accumulate clinical narrative that its queue does not need. Patient messages are also untrusted input. A message can contain text that tries to change the agent's instructions, so the OWASP prompt injection guidance applies: treat patient content as data to classify, never as commands to follow.
Every action should leave a record: which rule was applied, which template was sent, which slot was held and who was notified. The NIST AI Risk Management Framework is a useful reference for mapping these risks, assigning owners and monitoring behavior over time. Evaluation matters as well. A practical test set includes ordinary requests, edge cases and messages that should trigger clinical routing, and it should be rerun whenever the rules or the underlying tools change.
Be Human approaches this kind of deployment as a defined role before any automation is built. The first engagement names the queue, the memory and tools it needs, the approval boundaries and a success criterion the practice can observe. Managed operations then cover monitoring, recurring evaluations, integration maintenance, permission reviews when staff or systems change, and controlled improvements to scripts and rules. The clinical routing list stays under the practice's clinicians, who decide what the agent may say.
Where to start with a first engagement
A first engagement should be small enough to finish and specific enough to measure. Before anything is built, the practice should be able to answer these questions:
- Which queue has a visible backlog and a named staff owner?
- Which messages are administrative, and which must route to clinicians?
- What does success look like in a form a coordinator can check each week?
- Which systems may the agent read, which may it write, and which must it never touch?
- Who reviews the evaluation results, and how often?
If the answers are unclear, the right first step is to map the queues rather than to automate them. If you want to find the first task in your office worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Can an AI agent book eye surgery on its own? It can offer approved dates, record a tentative hold and track the forms and acknowledgments that a surgical file needs. Confirming candidacy, changing a surgical plan and deciding whether a case is ready remain clinical decisions. The agent should hand those points to the surgical coordinator or surgeon with a clear summary.
Is it safe for an AI agent to message patients about recalls? It can be, when the recall intervals and message wording come from clinicians and the agent is limited to approved templates and slots. Any reply that mentions a symptom, a medication question or a complaint must leave the recall flow and go to clinical staff. Opt-outs should be honored immediately and checked before every message.
How does an agent know when a patient message is clinical? It cannot rely on intuition, so the practice must define triggers in writing, including symptom language, medication questions, post-operative complaints and emergency wording. The agent classifies incoming messages against that list and routes any match with a short summary. The list should be reviewed on a schedule and tested with realistic messages, including unusual phrasing.
What should a practice measure after deploying an agent for administrative work? Choose a measure that staff can verify, such as how many recalls remain open at the end of the week, how many surgery holds expire without confirmation, or how many cancellations reach the coordinator the same day. Review the audit log alongside the measure, because a good number can hide a message that should have been routed. Adjust rules through the same controlled process each time.
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.


