AI for Ophthalmology Practices: Recall Queues and Surgery Scheduling
How ophthalmology practices can use AI agents for recall queues, surgery calendar holds and pre-op paperwork, with clinical decisions left to physicians.

AI for an ophthalmology practice works best on the administrative edges of care. A well-scoped agent can keep recall queues moving, hold surgery calendar slots without double booking, and assemble pre-operative paperwork for staff to check, while every clinical judgment about diagnosis, timing and surgical candidacy stays with the physician. Be Human builds these agents with defined memory, tools and approval boundaries, so the practice can see what the agent did, why it did it and who approved the next step.
Key takeaways
- Start with recall queues or surgery calendar holds, because both are rule-based workflows where a missed step costs time, revenue and patient trust.
- Give the agent read access to the data it needs, and write access only to drafts and holds that named staff confirm.
- Keep diagnosis, surgical candidacy, urgency triage and consent with physicians and clinical staff, and write that boundary into both the agent's instructions and its permissions.
- Measure each task with one observable criterion, such as whether every due recall has a logged outcome by the end of the agreed window.
- Review logs, escalations and permissions on a schedule, not only after something goes wrong.
Where recall queues leak in eye practices
Recall work in an eye practice looks simple on paper. A patient is due for a follow-up, a routine check or a repeat visit, and someone makes the call. In practice the queue is shaped by several overlapping schedules. Some patients need follow-ups at intervals set by their treating physician, while others are due for routine refraction or contact lens checks. When the recall list lives in a spreadsheet, a practice management export and the memory of two front desk employees, items fall through the gaps between systems.
The failure points are usually predictable. A recall is created but the patient's phone number is out of date. A patient reschedules by phone and the original recall stays open. A physician changes a follow-up interval in the chart, but the recall list is never updated. An agent can take over the repetitive part: reading the due list each morning, checking for newer appointments or cancellations, drafting outreach in the practice's approved wording and recording each attempt. It should not decide that a patient can wait longer. That change belongs to the physician.
Dental offices run into the same pattern, and the structure transfers well. The recall agent approach for dental offices shows how a recall queue can be organized around outcomes rather than calls.
A step-by-step plan for a recall agent
Before connecting any tool, write the plan down in plain language and agree on it with the clinical lead and the office manager.
- List every recall type the practice runs, and name the person who sets the interval for each one.
- Map the data each recall needs: last visit, due date source, contact channel, language preference and any note that blocks outreach.
- Define what the agent may do alone: read the due list, check for newer appointments, draft a message from approved templates and log each attempt.
- Define what it must hand to staff: any reply that mentions a symptom, any request to change an interval, and any contact that fails twice.
- Set one success criterion, such as every due recall reaching a logged outcome (booked, declined, unreachable or escalated) within the agreed window.
- Run the agent against a copy of a past queue, compare its log with what staff actually did, and fix the gaps before it touches live patients.
- Schedule a recurring review of the log, the escalations and the message templates.
One rule deserves emphasis. If a patient mentions a change in vision, pain, flashes or a shadow in the visual field, the agent must not reassure the patient or suggest that the message can wait. It routes the message to clinical staff at once and records that it did so.
Surgery calendar holds without double booking
Surgery days depend on a chain of resources: the surgeon's block, the operating or procedure room slot, the anesthesia or sedation preference, the instrument set, the nursing staff and the patient's pre-op window. A hold is a temporary reservation that says these resources are likely needed but not yet confirmed. Holds fail when they are created in one system, released in another and nobody notices until a case has to move.
An agent can manage the mechanics of this calendar. It can propose holds that match the surgeon's blocks and the eligibility window the physician has documented, flag conflicts, expire unconfirmed holds at the deadline the practice sets and draft a message to the patient when a date changes. It should never choose which patient goes first on a shared operating day, and it should never move a case because a slot looks empty. Those are clinical and operational priorities that the surgeon and the coordinator set.
Illustrative scenario: a coordinator at a small practice sees two unconfirmed holds on a surgeon's block. The agent flags that one hold conflicts with an instrument set already booked for another case. It drafts a note to the coordinator with both options and leaves the decision in her queue. She chooses one, the agent updates the calendar, and it sends the patient a change message only after she approves the wording.
Pre-op paperwork: assembly, not judgment
Pre-operative administration is a stack of documents with different owners. Insurance authorization, consent forms, medication questions, transportation and escort details, and the list of tests the surgeon has ordered each come from a different source. Staff spend much of their time chasing these pieces and checking that each one matches the case. That assembly work suits an agent well: it collects what is on file, identifies what is missing, drafts requests to the patient or payer and produces a checklist for the pre-op nurse. Imaging centers deal with a similar authorization load, and the approach to scheduling and prior authorizations is a useful comparison for how the paperwork file should be organized.
The boundary here is clear. The agent does not decide whether a patient is a surgical candidate, does not interpret test results and does not sign, alter or pre-fill consent language beyond the approved form. If a result falls outside the range the surgeon documented, the agent flags it to the clinical team with the source document attached. Each pre-op file should show every document, when it arrived, what it was checked against and which staff member reviewed it.
Approval boundaries and who keeps the decisions
Every agent action should fall into one of three groups: act and log, draft and wait, or never do. Reading the queue and logging attempts are act-and-log actions. Drafting a patient message or a calendar change is draft-and-wait, and a named staff member must approve it. Changing a clinical interval, deciding candidacy, altering consent or sharing records through unapproved channels are never-do actions. The permission model should enforce these groups, not just the agent's instructions. OWASP's authorization guidance recommends checking access on every request rather than trusting a description of what the caller should be allowed to do, and the NIST AI Risk Management Framework treats human oversight, documentation and ongoing measurement as part of managing AI risk rather than a one-time setup task.
Be Human's managed approach puts these boundaries in place from the first engagement. The engagement defines one useful role, the memory and tools that role needs, the approval points and one observable success criterion. Ongoing operations then cover monitoring of the agent's logs, recurring evaluations on realistic queues, permission reviews whenever staff or connected systems change, and controlled improvements to templates and routing. The practice keeps the clinical decisions, and Be Human keeps the agent inside the lines the practice has defined.
Choosing the first task
Start with one queue, not with a broad front desk agent. Recalls or surgery holds are good first candidates because their rules are written down, their outcomes are easy to see and a mistake can be caught the same day. Give the agent one role, one set of tools and one success criterion, then widen its scope only after the log shows it behaving as intended. If you want a concrete starting point for your practice, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Can an AI agent make clinical decisions in an eye practice? No, and it should not be designed to. The agent handles queues, drafts, logs and assembly work, while physicians decide diagnosis, timing, candidacy and any change to a treatment plan. Routing rules send symptom reports and clinical questions to staff immediately.
Does the agent need access to the full patient chart? Usually not. Each task should receive only the fields it needs, such as the due date, the contact channel and the approved template. Practices should confirm the privacy obligations that apply where they operate before connecting any system.
How does a practice know the agent is doing its job correctly? Through the log, a recurring review and evaluations. Each week, staff compare a sample of the agent's entries with what actually happened, and the success criterion tells you whether every due item reached a documented outcome. Permission reviews and template updates are scheduled, not improvised.
Where should a practice start if the team is already busy? Start with the queue that creates the most avoidable follow-up work, usually recalls or surgery holds. Keep the first role narrow, put approvals in front of every patient-facing message and expand only after the logs show the work is reliable.
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.


