AI Agents for Med Spas: Consult Requests, Packages and Follow-Up
How a med spa owner can use an AI admin agent for consult requests, package questions and follow-up, and when to book a scoping call.

An AI agent for a med spa is a software worker that handles consult requests, answers package questions within rules you set, and runs follow-up after treatments. It works from a company memory made of your services, booking rules, aftercare notes and escalation limits. Start with one admin role, build that memory from documents your team already uses, and define one success measure you can check every week.
Key takeaways
- Give the agent one job first, such as consult request triage or post-treatment follow-up, not the whole front desk.
- The company memory is the asset you own: services, package rules, aftercare language, screening questions and escalation rules, kept current by a named person.
- Separate what the agent may read and draft from what it must never decide, such as medical eligibility, exceptions or refunds.
- Measure success with a few observable checks reviewed on a schedule, not with impressions or general enthusiasm.
- Book a scoping call once the memory draft and the first task are clear enough to test against real requests.
What a med spa admin agent actually does
Most med spas lose time in the same places. A consult request arrives through a web form, a direct message or a voicemail, and someone has to decide whether it is a fit, which provider should see it and what to say next. Package questions come in repeatedly with slight variations. After a treatment, the team intends to check in, confirm aftercare and suggest the next visit, but the day fills up and the follow-up slips.
An admin agent can take over the repetitive parts of that work. It can acknowledge a request, collect the details your team needs, explain which packages exist and what they include, propose a time from the calendar you allow it to see, and send a planned follow-up message after a treatment. It can also flag requests that need a human, such as anyone who mentions a medical condition, a medication, a recent procedure or a complaint.
What it should not do is equally important. It does not diagnose, does not decide whether someone is a candidate for a treatment, and does not promise results. Those decisions belong to licensed clinicians under your own clinical policies. A good agent makes sure the right person sees the right request quickly, with the context already written down.
Build the company memory before choosing tools
Most owners start by comparing software. The better order is to write down what the business knows. An agent is only as useful as the information it can reach and the boundaries it respects. Anthropic's guidance on context engineering makes a practical point: the context an agent receives is a finite resource, and more text is not automatically better. Curated, current and well-labeled material beats a dump of every folder.
There is also a design choice between teaching a model your business through retraining and giving it retrieval from documents you control. The AWS guidance comparing retrieval and fine-tuning describes the tradeoffs. For most small operators, retrieval from maintained documents is the more practical path, because you can update a service description the same day a protocol changes, and you can see which source the agent used. Your memory then lives in documents you own, not inside a model you rent.
This is the part that gets more valuable over time. When a better model becomes available, the memory stays the same and the agent can usually do more with it. Teams that rebuild their knowledge for each tool start from zero every time.
A step-by-step plan for the memory
Work through these steps before any agent sees a real client message. Each one is small enough for an operations lead to finish in a few focused sessions.
- Step 1: List the services you offer, with a plain description of what each one is for, who should not be referred for it without clinician review, and what the aftercare steps are.
- Step 2: Write the package rules in full sentences: what a package includes, how sessions are used, how long it is valid, and what happens when a client misses a visit. Leave out any amount that changes often and link it to the official source instead.
- Step 3: Draft the screening questions your clinicians already ask, then mark which answers must go straight to a human.
- Step 4: Write the escalation rules: who gets each type of request, how fast they should respond and what the client is told while they wait.
- Step 5: Write the follow-up schedule for each treatment type, including the message wording your team approves and the messages that must always be reviewed first.
- Step 6: Name one owner for each document and a date-free review rhythm, such as a standing monthly check.
- Step 7: Collect ten or fifteen real past requests, anonymized, and mark what a good response would have been. These become your first test set.
The last step is the one most teams skip. Without real examples, you cannot tell whether the memory works.
Consult requests and packages: where the agent stops
A decision guide helps everyone agree on the limits before launch. Write it as a short table in your head, then as a document your staff can read.
- The agent may answer: what a service is generally for, what a package includes, which open times exist, and what to bring to a visit.
- The agent may draft but a person must send: replies that mention prior treatments, anything about aftercare for an unusual reaction, and any reply to a complaint.
- The agent must route to a clinician: any mention of pregnancy, medication, a recent procedure, a skin condition or a request for a specific outcome.
- The agent must never decide: candidacy, exceptions to package rules, refunds, discounts outside policy or any medical claim.
Illustrative scenario: a med spa receives a web form asking about a package for a series of skin treatments and mentions that the person has recently started a new prescription. The agent does not answer the eligibility part. It confirms the request was received, explains that a clinician reviews new prescriptions before booking, lists the package options as general information, and places the request in the clinician queue with the prescription note highlighted. The owner sees a clear case, and the client gets a useful reply rather than a generic one.
Form text and messages from outside your business can also contain instructions meant to manipulate an agent. The OWASP prompt injection guidance recommends treating inbound content as data, not commands, and keeping the agent's permissions narrow. That means the agent can read a request but cannot change your package rules because a message asked it to.
Follow-up that stays accurate and kind
Follow-up is where many med spas find the quickest return on structured work. A treatment creates a natural moment for a message: confirm how the client is doing, repeat the aftercare points, and offer the next visit or a patch check if your protocol includes one. The agent can schedule these messages from the treatment record, but only if the record is accurate and the wording was approved.
Write the follow-up in your clinic's voice, not in a generic automated tone. Keep messages short and specific. A client who had a treatment on the left side of the face on one day should not receive a message about a different area, so the agent must read the actual record. If a client replies with a symptom, the message goes to a clinician at once, not into a sequence.
The same logic applies to rebooking. For a comparable example outside med spas, see how a lash and brow studio can structure fills and patch tests in AI for a Lash and Brow Studio: Fills, Rebooks and Patch Tests. Packages in wellness settings raise similar questions about intake and follow-up cadence, which are covered in AI for Acupuncture Clinics: Intake, Packages and Follow-Ups.
How permissions and evaluations keep the agent honest
Every agent needs a permission list. Use the principle of least privilege: the agent gets read access to the documents it needs, write access only to the draft folder or the request queue, and no access to payment systems or clinical charts unless a specific, reviewed task requires it. The OWASP authorization guidance describes this pattern of checking access at each action, not assuming it from the start of a session. Review the permission list once a month, and remove anything that is no longer used.
Evaluations are the second safeguard. Anthropic's guidance on evals for AI agents describes building a set of realistic tasks, scoring the outcomes and rerunning them after any change. For a med spa, that means keeping your test set of past requests, running the agent against it after every memory update or model change, and checking whether it routed the sensitive cases correctly. A failing case is more useful than a confident demo.
Be Human's managed approach handles this as an ongoing operation. The engagement starts by defining one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion. Managed operations then add monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, so that changes to packages, staff or models are tested before they reach clients.
When to book a call to scope the first admin agent
A call makes sense when you have the core of the memory drafted, you can name the one task you want handled first, and you know who on your team will own the approval rules. You do not need every document finished. You do need enough clarity to say what success looks like. A reasonable success criterion might be that every consult request receives a first reply with the required intake fields collected, and that every sensitive case reaches a clinician queue on the same day it arrives. Those are checks you can see.
Book a call when the team agrees on the decision guide above, when you have your first set of anonymized past requests, and when you want to decide whether the first role should be consult triage or post-treatment follow-up. Do not book one simply to ask what the software can do. The first scope should be narrow enough to test in a few weeks and wide enough to matter to your calendar.
If you are not sure where to start, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Can an AI agent answer medical questions for a med spa? It should not give medical advice or decide whether someone is a candidate for a treatment. Its role is to provide general information about services, collect intake details and route anything clinical to a licensed clinician under your own policies. Anything that involves symptoms, medications or a recent procedure should go straight to a person.
What should go into the company memory first? Start with the services you offer, the package rules written in full sentences, your screening questions, your escalation rules and the aftercare steps for each treatment. Add a set of anonymized past requests with the response you would have wanted, because that set becomes the test you use to judge the agent. Keep one named owner for every document.
Why does the memory matter more than the model? Models change, and a memory built from maintained documents can move with them. When you update a package or a protocol, the agent reads the new version without retraining, and you can see which source it used. A memory tied to one tool or model has to be rebuilt each time you switch.
How do we know the agent is working? Check it against a fixed set of realistic requests after every change, and track whether sensitive cases reached a clinician and whether routine requests received accurate, timely replies. Review permissions monthly and look at every incident. If the checks pass and the team trusts the queue, you can widen the role carefully.
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.


