Dermatology Offices and AI Agents: Referral Tracking and Recall Reminders

A dermatology office can use an AI agent to track outgoing referrals, chase missing consultation notes, and send recall reminders for follow-up skin checks, all within rules the practice sets in advance. Clinical decisions, such as whether a lesion needs biopsy or whether a treatment plan should change, stay with the physician, who approves any action that touches care. The sensible start is the administrative loops, run under observation first, with the agent's role widened only when the records show it is reliable.

Key takeaways

• Referral tracking and recall reminders are repetitive, rule-based and easy to measure, which makes them a practical first job for an agent.
• The agent works only with administrative state: referral status, reminder queues, appointment requests and staff tasks.
• Diagnosis, treatment selection and any message that gives clinical advice remain under physician approval.
• Every agent action needs a named owner, a log entry and a way for staff to correct or reverse it.
• The role can grow over time, but each expansion should be a separate, reviewed change with its own success criterion.

Why referrals and recalls are the right first job

Dermatology runs on follow-up. A patient sent in for a changing lesion needs the referral to arrive, the consultation note to return, and the result to reach the referring clinician. A patient with a history of skin cancer or chronic inflammatory disease may need a skin check on a set interval. Both loops depend on dates, statuses and messages that are easy to lose when a front desk is busy. An agent does not replace the judgment behind these loops, but it can keep the queue honest: which referrals were sent, which were acknowledged, which notes are still missing, and which recalls fall due in the coming weeks.

These tasks suit an agent because their inputs and outputs are visible. A referral either has a confirmed appointment or it does not. A recall either went out or it did not. When that state lives in one place, a review takes minutes instead of an afternoon of checking inboxes and paper logs.

What the agent owns and what the physician keeps

The clearest boundary separates administrative state from clinical content. Administrative state includes dates, status fields, reminder schedules, document requests and staff tasks. Clinical content includes diagnosis, pathology interpretation, treatment choices, medication changes and any advice a patient might act on. Before any build, the practice should write this boundary down so that staff, physicians and the agent work from the same rules.

Use this guide for each proposed task:

• Does it change only a status, a date or a reminder? The agent may execute it under the configured rules.
• Does it require reading clinical content to decide? The agent may summarize the content for the physician, but the decision stays with the physician.
• Does it send a message to a patient or another clinician? The agent drafts it, and a named staff member or physician approves it before it goes out, until the practice has reviewed a track record for that message type.
• Does it touch a record outside the defined fields? Do not automate it. Route it to staff.
• Would a wrong action be hard to notice or hard to reverse? Keep it under approval, however simple it looks.

Building the referral tracker step by step

A referral tracker is a list with rules attached, not a clever model. Build it in this order:

• Define the referral states your office actually uses, such as sent, received, scheduled, seen, note returned and closed. Keep the list short enough that staff will use it.
• Name the source of truth for each state. If the practice management system holds the appointments, the agent reads from it rather than keeping a shadow copy.
• Set a response window for each state, such as a flag when a referral has no acknowledgment after the agreed period. Choose these windows with the physician, not with the software default.
• Write the escalation path: who receives a stalled referral, through which channel, and how quickly.
• Log every agent action with the record it touched, the rule it followed and the person who approved it where approval was required.
• Run the tracker in shadow mode for a short period. The agent proposes flags, staff compare them with their own list, and each disagreement is reviewed before any automation is enabled.

The AI Agent for a Chiropractic Practice: Intake and Visit Recalls article shows how visit recalls can be kept separate from clinical notes, and Superintelligent Agents for Orthodontists: Aligner Tracking and Recalls applies the same interval-based logic to scheduled care.

Recall reminders that respect clinical judgment

Recall reminders sit where clinical judgment meets administration. The interval for a follow-up skin check is a physician decision that may depend on personal history, prior findings and the patient's circumstances. The agent should not invent intervals. It should read the interval the physician has recorded, calculate the due date, and queue a reminder in approved wording.

Several practical rules help. Keep reminder templates neutral: they say a follow-up is due and how to book it, and they never suggest what the team suspects. Give patients a reply path that routes to staff, not to an automated clinical answer. If a patient reports a new symptom, the message must reach a person promptly, and the agent should not attempt to triage it. Patients who opt out, have changed phone numbers or are overdue should appear on a review list rather than disappearing from the queue.

Illustrative scenario: a referral that stalls

Illustrative scenario, not a real case: A general practitioner refers a patient with a changing mole to a dermatology office. The referral is sent, but the office never confirms receipt. The tracker flags it as sent without acknowledgment once the window the physician set has passed. The coordinator sees the flag at the morning review, calls the receiving desk, and finds the referral sitting unread in a queue. The appointment is booked that day. The agent made no clinical decision. It noticed that a state had not changed and sent the record to a person who could act on it.

Access, privacy and the audit trail

Health data calls for narrower access than most business data. The agent should hold only the permissions its role needs: read appointment and referral status, write reminder queue entries, and create staff tasks. It should not have blanket access to clinical notes or billing. Credentials should be scoped and revocable, and access should be reviewed on a schedule rather than only after something goes wrong. The OWASP authorization guidance is a useful reference for checking that each action is authorized at the moment it happens, not assumed from an earlier step.

Inbound text also needs care. A referral letter or patient message is content the agent reads, and content can contain instructions meant to change its behavior. Treat inbound documents as data, never as commands, and keep the agent's permitted actions in configuration that a message cannot alter. The health privacy rules that apply where the practice operates are a matter for the practice and its counsel, but the design should make the required records easy to produce. The NIST AI Risk Management Framework offers a structure for naming these risks and their owners before launch.

How Be Human approaches the first engagement

Be Human's managed approach starts with one role. For a dermatology office, that usually means the referral tracker and the recall queue. The engagement defines the memory the agent may use (the states, intervals and templates the physician approved), a short list of tools and permissions, a clear approval boundary for patient-facing messages, and one observable success criterion, such as the share of referrals with confirmed receipt inside the agreed window. Monitoring, recurring evaluations, permission reviews and incident response follow launch, so every change to the role is controlled and recorded. The service is remote, in English or Portuguese.

Choosing the first task and what to measure

Pick the first task by asking which loop depends on one person remembering to check something. Measure it before you automate: how many referrals sit without acknowledgment, how many recalls go out late, and how often staff correct a reminder before it is sent. These baselines show whether the agent is worth the effort and give you a fixed point for comparison after launch. Expand the role only when the record shows reliability and a physician has approved the new boundary. The free Be Human AI diagnostic can help you find the first task worth handing to an agent in your practice.

Frequently asked questions

Can an AI agent send recall reminders in a dermatology practice without physician review? Yes, for administrative reminders that use wording the physician approved and intervals already recorded in the chart. Messages that answer clinical questions, change an interval or respond to a new symptom should go to a person, and the agent should route them rather than answer. Many practices begin with a draft-and-approve step for every patient message and relax it only for message types with a clean review history.

Does referral tracking replace the front desk? No. The agent keeps the status of each referral current and flags the ones that stall, but staff still make the calls, confirm the appointments and handle the patients who need a person. In most offices the change is that the coordinator starts the day from a short exception list instead of searching through inboxes and paper logs.

What should a practice measure before and after launch? Record a baseline for the loop you plan to automate, such as referrals without acknowledgment, recalls sent late and reminders corrected by staff. Keep the same measurements after launch so the comparison is fair. A single success criterion, tied to the role's defined purpose, is easier to review than a long list of metrics.

Which parts of dermatology work should stay with the physician? Diagnosis, the interpretation of findings, treatment decisions, medication changes and any clinical advice given to patients should remain with the physician. The agent can prepare administrative context, such as the date of the last visit or the status of a requested note, but it should not decide what that context means for care. Any expansion of the agent's role into clinical content should be a separate decision, documented and approved before it is built.

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