Professional Workbook · Healthcare · 3 lessons
Leave with an administrative process at your clinic specified in the Agent Architect’s Kit: patient data out of scope, confidentiality as a fixed restriction, and the agent handling scheduling, confirmations, and claim denials while the team takes care of people.
Lessons
Leave with the list of ten administrative processes and one mapped across six links, with clinical work separated from administrative work.
Leave with a clinic agent’s permissions matrix decided, with medical records and health data out of scope and confidentiality written as a restriction.
Leave with your clinic’s agent specification built in the kit from the healthcare preset and rehearsed in an AI chat.
Before you begin
The six links, the long-horizon request, the chain, the context inventory, the quality checklist, and the permissions matrix are explained in the course AI Work Architect. Here, they’re applied to the clinic. The exercise in lesson 3 uses the Agent Architect Kit, with the healthcare preset.
Health Workbook · Lesson 1
By the end of this lesson, you can list the ten administrative processes in your clinic that an agent can handle, distinguish the administrative from the clinical parts of each one, and map one of them across six links.
At a clinic, the AI conversation always gets stuck in the same place: “what if it makes a mistake with a patient?” The concern is legitimate and has a simple solution: the agent stays away from clinical decisions. It handles the work that currently takes up the receptionists’ and manager’s time: confirmation calls, claim-denial spreadsheets, private-pay billing, and administrative documents. That’s where a six-office clinic loses three hours a day, and that’s where the redesign begins.
↓ role to study
The rule that organizes this entire workbook fits in one sentence: the agent does the work administrative and never handles clinical. Administrative tasks are anything that would happen the same way if the clinic were a store: scheduling, confirming, billing, invoicing, filing, answering questions about where to park. Clinical tasks are anything that requires healthcare training: diagnosis, treatment decisions, prescriptions, symptom guidance, test interpretation.
The test for separating them is simple: could a well-trained receptionist do this without asking a health professional? If so, it’s administrative and can go to an agent. If they would need to ask, it’s clinical and stays out of scope, without exception—not even “just a summary.” This isn’t a limitation of the technology. It’s the same line the clinic already draws between the front desk and the exam room.
Renata, administrative manager of a medical clinic with six exam rooms, listed what the front desk handles in a day: 90 confirmation calls, 40 messages asking “what time is my appointment?”, 12 add-on appointments, the insurance claim denial spreadsheet, and three attendance certificates. All administrative. The two questions the front desk passes to nursing each day—“can I take my medication before the test?”—are clinical and stay with nursing.
Test yourself
A patient sends a message: "I have a headache. Can I move my appointment earlier?" What part of this can a clinic agent handle?
The core course asks for ten critical processes instead of dozens of chatbots. In a clinic or dental office, the ten are usually these: scheduling and confirmation of appointments; waitlist and appointment openings; claim denials from insurance (claims the insurer returns unpaid); billing to the health plan; billing for self-pay patients and treatment plans; administrative documents (attendance certificate, statement, receipt); onboarding for new patients (registration, consent form, arrival instructions); inventory of supplies; handoff to the professionals; and satisfaction survey post-appointment.
Notice that none of the ten requires healthcare training, and that all of them currently take time away from people who also serve patients at the front desk. Choose the first one using the core course’s criteria: what repeats every day, involves more than one person, and costs money or loses patients when it fails. For most clinics, that points to confirmations or claim denials.
Patrícia, manager of a dental practice with three dentists, thought her problem was something else: treatment plans that were presented but never accepted. It’s process five on the list: following up on treatment plans. Forty plans a month, twelve accepted, and no one had time to follow up on the other 28. An agent that follows up with a standard message and schedules a conversation with the dentist is administrative from start to finish.
Before
Two receptionists spend the afternoon calling to confirm the next day’s 90 patients. One-third don’t answer. The front desk is left unattended several times, and the no-show rate is 18%.
After
At 5 p.m., the agent sends the standard message, reads the replies, records the status in the schedule, and delivers the list at 8 a.m.: confirmed, no response, canceled. Reception calls only the 15 who didn’t respond.
Applying the base course's map to the most common process. Input: at 5:00 p.m., the next day's schedule is pulled from the clinic system. Rule decisions: patient hasn’t replied by 8 p.m.? Send a second message. Minor? Message only the registered responsible party. Judgment decision: release the slot of someone who hasn’t confirmed to the waitlist? That’s up to the manager because it depends on knowing the patient. Action: send the standard message and record the status.
Verification: the name, date, time, and provider match the schedule, and the message contains nothing beyond the standard text. Exception: the patient replies with a question ("can I reschedule?", "do I need to bring test results?"). The agent doesn’t reply; it forwards the message to the front desk. Result: at 8:00 a.m., the manager knows the status of every appointment that day. Notice what was left out: the reason for the appointment never appears in the message, and the agent never answers a question that isn't about scheduling.
At Renata’s clinic, the most common exception was the question “Do I need to fast?” It’s a clinic: it depends on the test. The written rule became: any message with a question that isn’t about scheduling goes to reception, which decides whether to answer or pass it to the nursing staff. The agent never reads the content beyond identifying that it’s a question.
Claim denial It’s the claim form the insurance plan returns without paying. The most common reason is a procedure code that doesn’t match the insurance plan’s table, and the appeal is a standard text citing the claim form, the correct code, and the amount. It’s repetitive, tedious work with clear rules and a short deadline—in other words, the ideal process for an agent. Input: the insurance provider’s weekly statement. Rule decision: denial due to a code? Prepare an appeal. Due to missing prior authorization? Set it aside for the manager.
Action: match the claim form against the insurance provider’s table and write the appeal using the template. Verification: the claim form, code, and amount match. Exception: a denial over 2 thousand reais or one repeated for the same patient, which may indicate a registration problem, goes to the manager. Result: the completed appeals for the manager to review and sign. The agent never sends anything to the insurance provider: sending is an external action, and the manager signs.
At Patrícia’s dental practice, claim denials happen for another reason: an X-ray is billed without the report attached. The rule is the same: denial due to a missing document? Attach the document from the administrative record and appeal. But the report is health data. The solution was for the agent to list which claims need a report and for the assistant to attach it. The agent prepares; a person handles the sensitive data.
The clinic’s agent never knows why the patient came in. It knows when, with whom, and whether they confirmed.
Practice now 0/4 done
In about 12 minutes, list your clinic’s ten administrative processes, choose one, and map it across six links, clearly marking where the clinical line runs.
Use your own paper or spreadsheet, without any patient names: you write about processes, not cases. Nothing changes at the clinic. If a process seems both administrative and clinical, write down both sides and continue: separating them is the exercise.
You’ve just done what holds most clinics back: clearly separated what an agent can handle from what only the healthcare team may touch. This map goes into the kit in lesson 3.
Summary
Health Workbook · Lesson 2
By the end of this lesson, you can decide on a clinic agent's permission matrix, keep medical records and health data out of reach, write confidentiality and LGPD into the constraints, and assign a named owner to each permission.
A clinic is where agent governance stops being a best practice and becomes an obligation. A registration record has a name, ID, insurance plan, and phone number; the medical record has the rest of a person’s life. The LGPD treats health data as sensitive, with stricter rules, and the Medical Council treats confidentiality as a professional duty, not a technology duty. The good news: the administrative agent doesn’t need any of that to do its job. This lesson shows where to draw the line and how to write it into a matrix the clinic owner can sign without fear.
↓ role to study
Every clinic stores information in two categories that often live in the same system. On one side, the administrative: registration (name, phone, insurance, responsible party), schedule (who, when, with whom), billing (forms, codes, amounts), and each insurance provider’s table. On the other side, the patient record: reason for the appointment, progress notes, tests, diagnosis, prescription. The administrative agent works only with the first set.
The LGPD classifies health data as sensitive data, with stricter rules than those for standard registration. This makes the decision simpler, not more complicated: the agent cannot access sensitive data, period. It’s not "with approval," and it’s not "just for summarizing." If the administrative work comes out the same without the medical record—and it does—the medical record stays out of reach.
Renata made the agent’s context inventory for confirmations and claim denials: schedule, patient registration with guardians for minors, standard message, insurance statement, procedure fee schedule. Five sources, none of them the medical record. When the system vendor asked, “Do you want the agent to have access to the clinical module too, since it’s being integrated?” the answer was no, and it was documented.
As of September 2026, a clinic agent runs in three ways, from temporary to permanent. In the AI chat with project folder (ChatGPT, Claude, Copilot), you keep the standard message, the insurance agreement table, and the appeal templates, and attach the statement each week, without patient names. In the the clinic's messaging platform, many already offer automatic confirmation based on rules. And in the integration with the clinic system through the service door, the agent reads the schedule on its own, set up by whoever maintains the system, under an agreement that covers personal data.
The first approach works for learning and claim denials. For confirmations involving patient phone numbers, prefer the second or third approach from the start: a generic AI chat is a service outside the clinic, and a phone number with a name is personal data. This may seem bureaucratic, but it works like the rule against leaving the patient list on the printer: it’s not about the tool; it’s about where the data passes.
Patrícia started with approach one for X-ray claim denials: a project folder with the insurance fee schedule and appeal template, a statement attached each week, and only the claim number and code. For following up on treatment plans, which involves patient names and phone numbers, she waited for integration with the practice management system, which already had a service interface.
Common mistake
Export the day’s schedule with the "reason for visit" column and paste it into an AI chat to "test the confirmation." This happens because the system's default export includes every column and no one removes them. The reason for an appointment is health data leaving the clinic for an external service. Before any test, export only the four columns used in the confirmation: name, date, time, provider. And replace the names in the tests.
Applying the base course's seven actions to the appointment confirmation and claim-denial agent: read scheduling, administrative registration, and statement: allowed; medical records: never; write schedule status and appeal draft, allowed; change scheduling (reschedule, open a slot), with reception approval; send the standard confirmation message is allowed because the text is fixed and approved; sending any other text or filing an appeal with the health plan requires approval; purchase, delete e approve, it’s not allowed.
The only line that departs from the base course's ladder is "send the standard confirmation: allowed." This is defensible because the text doesn't change, the four fields are verified, and nothing beyond them is included. Any message that requires new text goes back to "with approval." And every permission has a named owner: the manager for reading and writing, the front desk coordinator for changing the schedule, and the provider for signing appeals and certificates.
At Renata’s clinic, the matrix has eight rows and fits on a sheet posted in the manager’s office. The line that sparked discussion was “release an opening to the waitlist”: reception wanted “allowed” to speed things up. It stayed “with approval” for the first two months; after the test cases passed and the log stayed clean, it became “allowed,” with a limit of two openings per day.
Three things must be put in writing before any clinic agent runs, and none of them is technology. Confidentiality: fixed constraint in the request, "no information beyond name, date, time, and provider in any message". Consent and channel: the patient authorized messages to that phone number, and for minors, messages go only to the registered responsible party. Record: every message sent and every prepared appeal are logged with the date, content, and approver, so you can answer any question from a patient, insurance plan, or board in one minute.
The detail that changes over time is the regulation, not the principle. Resolutions from the Federal Councils of Medicine and Dentistry on patient communication and technology use are updated; Brazil’s data protection law has new regulations from the national authority. That’s why this step is labeled as a tool, with a date: the principle (confidentiality, consent, logging) is foundational; check the exact wording with whoever handles the clinic’s legal matters when it’s time to sign.
Patrícia added a new field to the practice’s patient registration form: “accepts confirmation and treatment plan follow-up messages: yes/no,” completed at the first visit using the administrative consent form. The agent only contacts people marked “yes.” It was the cheapest change in the whole project, and the one the dentist who owns the practice liked most: the question “can it message everyone?” no longer came up.
Clinic Agent Governance Checklist
Practice now 0/4 done
In about 12 minutes, read the case, decide on the matrix and which data stays off-limits, compare with the answer key, then fill in your clinic agent’s matrix.
This workbook is intentionally on paper: permission to access patient data is the most consequential decision in this workbook, so you practice on a case before making your own decision. No system is configured. Keep your matrix and bring it to the clinic owner and whoever manages the system.
The case. Clínica Horizonte has eight exam rooms, four insurance plans, and 120 appointments per day. The manager, Mariana, designed an agent that confirms appointments, reads replies, records statuses, and prepares claim-denial appeals. The system vendor also offered to integrate the clinical module "so the agent can remind patients to bring test results." The front desk wants the agent to reschedule appointments on its own when a patient asks. For each patient, the system has: name, CPF, phone number, insurance plan, and member ID, responsible party, reason for the appointment, medical record, attached test results.
You’ve just decided, in writing, what an agent can see and do at your clinic, with the medical record kept out of reach. This is the document the clinic owner needs to say yes.
Summary
Health Workbook · Lesson 3
By the end of this lesson, you can build your clinic agent's specification in the Agent Architect Kit using the healthcare preset, then rehearse it in an AI chat with a fictional case.
You’ve mapped the process (lesson 1) and decided on the matrix (lesson 2). Now you need to bring them together on one page the clinic owner can approve in eight minutes and the person who manages the system can build from. The kit puts it together, and the healthcare preset comes prefilled with Renata’s clinic in all seven sections, with this workbook’s rules built in: medical records kept out of reach, confidentiality as a constraint, claim deadlines, and consent. Your job is to replace what’s specific to your clinic and rehearse it.
↓ role to study
The kit’s healthcare preset is Renata’s clinic, which you followed in lessons 1 and 2, now filled out across the seven blocks: next-day confirmations and appeals by denial code, six offices, four insurance plans. The rules in this workbook are already written into it: reason for the visit and medical records out of scope, confidentiality and LGPD as fixed restrictions, minors only through their guardian, appeals to the insurer only with approval, and the exception “patient asks a question → reception.”
What changes from one clinic to another is predictable: the insurance plans and their tables, the send time, the approved standard message, who approves what, and the chosen process. A dental office replaces claim denials with treatment plans. The framework stays the same; the content is yours.
Renata loaded the preset and changed six things: the four insurance plans, the time (her clinic sends messages at 4 p.m.), the standard message, the name of the front desk coordinator as the owner of “change schedule,” the claim denial threshold that triggers a call (she uses 1,500 reais), and the test cases, which she built from fifteen real appointments and claim denials from March, with the names changed. Patrícia changed more: the entire process became treatment plan follow-up, but the matrix and out-of-scope data stayed almost the same.
At your clinic, the first process is usually
In the kit, the “load example by area” selector has a healthcare option. When you load it, the “area” field is set to healthcare, turning on three checks the kit runs automatically, in addition to the general checks in the core course: if health data (medical records, diagnosis, report, test) isn’t on the out-of-scope list, a gap appears; if any action or goal includes a clinical term (diagnosis, prescription, care plan), another gap appears; if LGPD, consent, or confidentiality don’t appear anywhere, a third gap appears.
This may look like a screen full of fields, but it works like a patient intake form: each field is a question you answer once, and the kit builds the page automatically. If you accidentally delete a confidentiality restriction, the gap appears below the specification before you hand it over.
When adapting the preset for treatment plans, Patrícia wrote in the objective: “follow up with patients who were presented with a plan and recommend the pending treatment.” The kit flagged it: “some action or objective seems clinical.” She changed it to “follow up with patients who were presented with a plan and schedule a conversation with the dentist.” The dentist is the one who discusses the pending treatment. That was the gap she was most grateful to catch.
The adaptation follows an order that helps you avoid forgetting the essentials. Start with block 1: your clinic's name and the outcome in your own words. In block 2, replace the rule decisions with your limits (when to send the second message, the claim-denial amount that sets the cutoff). In blocks 3 to 5, add the standard text, insurance plans, actual systems, and inventory with your owners. In blocks 6 and 7, add your stop-and-escalate thresholds, your set of cases, and people's names.
Think twice before changing the out-of-scope data list, the confidentiality restriction, “appeal to insurer: with approval,” or “approve: not allowed.” Those lines are why the clinic owner says yes. If you need to loosen any of them, do it after a month of clean logs, one step at a time, as the core course teaches.
Renata spent twenty minutes adapting it, and the longest part was building the test set: choosing fifteen appointments and claim denials from March that covered the tricky cases (minor without a guardian, patient with two records, repeated denial). She didn’t copy any names into the kit: she wrote “Tuesday’s appointment 7, 2 p.m., Dr. Amaral” and kept the actual reference in the clinic’s spreadsheet.
Order for adapting the healthcare preset
When you paste the specification into the chat with the kit’s simulation prompt, the chat plays the coordinator and runs a fictional case. In clinics, three gaps come up often. First: the patient replies with a question, and the specification says to route it to the front desk but doesn’t say how the front desk is notified. Second: a claim is denied for a reason that isn’t on the list, and the coordinator doesn’t know whether to stop. Third: a patient has two records, and no one said which one to use.
Each one becomes a new line. "Notify the front desk using the 8 a.m. list, not a one-off message." "Claim denial for a reason outside the list: stop and escalate to the manager." Run three cases: a routine one, a difficult one, and the exception. Once all three pass, the page is ready for the clinic owner and the person who manages the system.
During Renata’s rehearsal, the chat, acting as a coordinator, asked: “The 17-year-old patient has no guardian on file; should I send it to her phone?” The specification said “minors only through their guardian,” but didn’t say what to do when there was no guardian on file. She added: “Minor with no guardian on file: don’t send; flag for reception to call.” Ten minutes of rehearsal, and one line that would prevent a message to a teenager without authorization.
Practice now 0/4 done
In about 15 minutes, load the healthcare preset into the kit, adapt it to your clinic, copy the practice prompt, and run a fictional case in an AI chat to get the list of gaps.
The specification contains no patient data: it has rules, system names, and criteria. The rehearsal case is made up (name, age, claim form, amount). No agent is built, no system is touched, and no message is sent. If the chat "sends" something in the simulation without asking for approval, that's a finding about your matrix: record it and press.
Passo a passo (o prompt de ensaio vem pronto do kit): 1. Abra https://inematds.github.io/kit-arquiteto-agentes/ 2. "carregar exemplo por área" → saúde. Confira que "área: saúde (administrativo)" ficou marcada. 3. Adapte os sete blocos na ordem do passo 03 desta aula. Zero lacunas de saúde na lista abaixo da especificação. 4. Em "Prompt de ensaio simulado", escreva o caso fictício, por exemplo: <Paciente Lúcia F. (fictícia), 17 anos, consulta quinta 14h com a Dra. Amaral, sem resposta até as 20h, responsável não cadastrado; guia 88213 glosada pelo convênio Vida por código divergente, 380 reais> 5. "copiar prompt com a especificação" → cole numa conversa nova do seu chat de IA. 6. Leia a narração: o coordenador deve parar pelo menos uma vez (menor sem responsável, recurso com aprovação). Anote os itens de "o que faltou".
You’ve just produced a specification for an agent at your clinic, with the medical record kept out of reach, and rehearsed it before building anything. This page is what separates "let’s use AI" from a project the owner will sign off on.
Summary