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

AI by industry

AI for Dermatology Referrals: Packets, Recalls and Prior Auth Checks

An AI agent can assemble dermatology referral packets, run recall lists and check prior authorization status. Clinical judgment stays with clinicians.

Be Human typographic cover with the title: AI for Dermatology Referrals: Packets, Recalls and Prior Auth Checks
Conceptual illustration created with AI · Be Human

An AI agent for a dermatology practice is most useful on the paperwork edges of the front desk: assembling referral packets from documents the office already receives, generating recall lists from scheduling data, and checking where prior authorization requests stand with the payer. It should not decide how urgent a lesion is, which referral is clinically appropriate, or what a patient should hear about a result. Those judgments stay with clinicians, and the agent's job is to make their queue cleaner and their handoffs faster.

Key takeaways

  • Start with three tasks that are document-heavy, repetitive and easy to check: referral packet assembly, recall list generation and prior authorization status checks.
  • Give the agent read access to the data it needs and write access only to drafts and task queues, not to clinical notes or payer portals without a named reviewer.
  • Write clinical triage, diagnosis language and patient-facing medical advice into the out-of-scope list before the first run.
  • Measure success with checks staff can perform without reading the agent's reasoning, such as packet gaps reviewed at intake and recall lists reconciled against the schedule.
  • Treat payer rules and office conventions as memory that needs an owner, a review date and a retirement process, not as a permanent knowledge base.

Why the dermatology front desk is a good place to start

Dermatology front offices handle inputs that rarely arrive in the same shape. A referral may come as a fax, an emailed PDF, a portal upload or a phone message from a primary care office. Photographs may be attached without dates or body-site labels. Follow-up needs live in the physician's plan, but someone has to turn that plan into a reminder by checking the schedule. This mix of varied inputs, repeated rules and frequent checking is where a well-bounded agent can save staff time without touching judgment.

The reason to begin here is that errors are visible. A missing insurance card copy is easy to see. A recall that was never sent shows up in the next reconciliation. A clinical mistake, by contrast, can hide inside a neat summary. Starting with work where mistakes are cheap to detect lets the practice learn how the agent behaves before anything closer to care is considered.

Referral packets: what the agent can assemble

A referral packet is the set of documents a dermatologist needs before the first visit, plus the administrative data the scheduler needs to book it. An agent can collect these items from the inbound referral, the registration record and any documents already on file, then compare them with a checklist written for each referral type. It produces a draft packet and a short list of gaps. It does not turn the referral into a diagnosis, and it does not decide whether the practice should accept the referral.

A practical checklist, owned by the practice manager and a physician reviewer, might include:

  • Referring provider name, return contact route and the date the referral was received
  • Patient identifiers matched to the registration record, with any mismatch flagged for staff
  • Insurance card copy and eligibility status as last verified by a staff member
  • Reason for referral in the referring provider's own words, copied without editing
  • Prior notes, labs or pathology reports already in the practice's possession
  • Dated clinical photographs with body-site labels, or a clear note that none were sent
  • Current medication and allergy lists as provided by the referring office
  • Gaps listed as staff tasks, never filled by the agent from inference

Two rules keep this safe. The agent copies what the referral says and labels what is missing. It never writes an interpretation into the packet that the referring provider did not write. A packet with visible gaps is safer than one that looks complete but contains an invented summary.

Recall lists: rules the agent can apply

Recall intervals in dermatology come from the physician: return visits after a procedure, skin checks for patients with a documented history that calls for surveillance, check-ins for patients on certain long-term therapies, and seasonal maintenance visits. The clinical content of those intervals belongs to the physician. The agent's job is to apply approved rules consistently against the schedule, produce the list, and remove people who should not be contacted.

A simple decision guide helps staff agree on the boundary:

  • Agent acts alone: generates the weekly list from a physician-approved interval, excludes patients with an upcoming appointment, and skips anyone with an opt-out on file.
  • Agent prepares, staff approves: outreach messages, a changed interval for one patient, and any recall that depends on a result not yet in the chart.
  • Clinician decides: whether a patient should return sooner than the standard interval, and any reply to a patient who reports a new or changing lesion.

The list is only as good as the rule set behind it. Each interval should name its owner and the date it was last reviewed. Dental offices face a similar split between routine reminders and clinical exceptions, and the same discipline applies, as described in the dental office recall approach.

Prior authorization status checks without arguing the case

Prior authorization is where front desks lose the most time to waiting, and where the temptation to over-automate is strongest. An agent can track the administrative state of a request reliably: the submission date, the payer reference number, the last status entered by staff or retrieved from an authorized portal, and the next follow-up date. It can prepare a call sheet or a task listing what staff should ask. It should not write medical necessity arguments, choose clinical codes on its own, or interpret a denial beyond routing it to the physician who owns the decision. Imaging centers face the same boundary, as outlined in the imaging prior auth guide.

Illustrative scenario: a referral for a procedure has been waiting for payer approval for more than a week. The agent flags the request on the morning status list, shows the submission date and last update, and creates a call task for the coordinator with the reference number and the question to ask. If the payer requests a supporting note, the physician writes it. The agent does not draft the clinical justification.

Portal access needs its own boundary. Use a dedicated account with only the permissions needed to view status, not to submit or change requests, and review that access on a fixed schedule. The OWASP authorization guidance is a useful reference for writing down who can do what and checking that the written rule still matches reality.

What stays with clinicians, and how to keep it there

Some work is clinical no matter how simple it looks. Judging urgency from a patient message, interpreting a photograph, deciding whether a pathology result needs a call, choosing a treatment and explaining a result all belong to licensed clinicians. An agent that receives a message containing words such as bleeding, spreading or new pain should route it to a clinical queue with the original text attached, and it should send no reassurance.

Clinical boundaries are easiest to keep when they are written down before the first run:

  • Triage of symptoms or lesions, including any message that describes a change over time
  • Diagnosis, treatment selection, or interpretation of pathology reports or photographs
  • Medical necessity statements, appeals and clinical letters
  • Any reply to a patient about a result before a clinician has reviewed it
  • Changes to medication lists, allergy records or problem lists in the chart

Patient-submitted text is also untrusted input. A message or pasted document can contain instructions meant to change the agent's behavior. The OWASP guidance on prompt injection prevention explains why instructions found inside data should be treated as data. In practice, the agent's tools should be limited so that even a successful manipulation cannot send a message to a patient or change a record. The NIST AI Risk Management Framework offers a broader structure for mapping these risks to owners and controls.

A practical plan for the first deployment

Start with one workflow, not the whole front desk. Keep the first run small enough that a person can review every output:

  • Choose one referral type and one recall rule set that a physician has already approved in writing.
  • Write the packet checklist and the recall exclusion list with the practice manager and one physician reviewer.
  • Grant read access to registration and schedule data, and write access only to a draft folder and a task queue.
  • Define the approval boundary for each output in one sentence, naming the staff role that approves it.
  • Set a success criterion staff can check, such as packets with no unflagged gaps at review or recall lists that match the schedule after reconciliation.
  • Log every flag, correction and escalation, and review the log on a fixed schedule.
  • Retire any rule that keeps producing corrections until its owner has reviewed it.

Be Human designs and operates this kind of setup as a managed service. In a first engagement, we define one useful role for the front desk, the memory it needs (such as approved recall rules and payer contact notes), the tools it can use, the approval boundaries and an observable success criterion. In managed operations, we monitor outputs, run recurring evaluations against the checklists, review permissions, maintain integrations with scheduling and registration systems, and handle incidents when something goes wrong. The clinical boundaries are written into the setup and reviewed with the physicians, rather than left for the model to remember.

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

Frequently asked questions

Can an AI agent handle prior authorization for dermatology procedures on its own? No, and it should not be set up that way. The agent can track administrative status, prepare the questions staff should ask the payer, and route requests that need a clinical note to the physician. The clinical justification and any appeal stay with the clinical team.

Does the agent need access to the full electronic health record? It should not. Start with the registration, scheduling and referral data required for the specific workflow, and add access only when a named reviewer has approved a specific need. Access should be reviewed on a schedule, because needs change as payer rules and office workflows change.

How do we keep the agent from giving patients clinical advice? Write the out-of-scope list into the setup, route any clinical language to a human queue, and make sure the agent cannot send clinical messages to patients on its own. Test it during evaluation with realistic messages, including messages that try to change its behavior. Review logs for any output that reads like advice.

What should be in place before the agent goes live? You need a written list of approved recall rules with owners and review dates, a packet checklist agreed by a physician and the practice manager, defined access and approval boundaries, and a success criterion staff can check. Logging and a review schedule should be running before the first real referral arrives.

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 →