Seven kinds of AI tool help a paediatric office with admin: a phone assistant that books visits and answers routine questions, a form-filling tool, reminder messaging, an AI scribe that drafts notes, a review-reply drafter, a newsletter writer and a rota planner. None should diagnose or triage symptoms, and each needs a person checking its output.
The catch is that each of these tools can drift into clinical territory without anyone noticing: a booking call turns into a fever question, a friendly review reply confirms a diagnosis in public. That is why every job here carries one safe-use rule you can paste straight into your practice policy, and why diagnosis and triage tools are left out entirely.
The data gate that comes before any tool
Before the seven jobs, one gate applies to all of them: anything that identifies a child or parent, or that links to a visit, counts as patient information. Do not put that into a general chat tool. A tool that handles it needs a signed business associate agreement (BAA), the contract that makes a vendor responsible for protecting health information on your behalf. Not every plan from the same vendor is covered.
- OpenAI: its help centre says the API can be covered by a BAA, with self-serve enrolment needing an established history of API usage; a BAA for a sales-managed ChatGPT Enterprise or Edu account comes through sales; and it does not offer a BAA for ChatGPT Business. Separately, ChatGPT for Clinicians lets an individually verified clinician sign a BAA for their own use; it is limited to certain licence types (physicians, nurse practitioners, physician assistants and pharmacists) and is offered in a single market only, and it does not cover the front-desk team. A BAA covering several users goes through ChatGPT for Healthcare, OpenAI's enterprise version.
- Anthropic: its BAA covers its health-ready services such as the first-party API and Enterprise plans, and a standard Enterprise plan is not covered until the Primary Owner turns on health-data compliance in the settings and accepts the BAA. Its commercial BAA article is separate from its guidance for personal plans, so do not assume Free, Pro or Max accounts are covered.
- Google Workspace: customers must review and accept a BAA in the admin console before using protected health information, and third-party add-ons are not covered by it.
Pages like these change, so read the vendor's current version and keep a copy of the signed agreement. The wider rules for sensitive records are in the vendor questions for client data. A good test is simple: if the vendor cannot show you a BAA that names the exact product and plan you will use, treat that product as off limits for anything about a child.
A working day at a four-doctor office, with numbers
Here is an illustrative main example to keep in mind through the seven jobs. A four-doctor paediatric office with two nurses and three front-desk staff takes about 220 calls on a typical weekday. The practice manager listens to a sample of recorded calls over a week and sorts them (all figures illustrative):
| Call type | Share | Calls a day | Average minutes | Desk minutes |
|---|---|---|---|---|
| Book, move or cancel a visit | 40% | 88 | 3 | 264 |
| Forms, letters and records requests | 12% | 26 | 4 | 104 |
| Opening hours, directions, "do you take my insurer" | 13% | 29 | 2 | 58 |
| Medication refills and results | 15% | 33 | 4 | 132 |
| Anything about symptoms or a worried parent | 15% | 33 | 6 | 198 |
| Billing questions | 5% | 11 | 5 | 55 |
The desk spends 811 minutes, about 13.5 hours, on phones daily. The first three rows (booking, forms, practical questions) total 426 minutes, about 7 hours, and are mostly routine. The "symptoms" row is the one that must stay with people. A safe plan lets a tool absorb a share of the first three rows, so perhaps half of those 426 minutes, roughly 3.5 hours, goes back to staff who can then answer the symptom calls faster. That is the kind of arithmetic to run on your own call log before choosing anything. Each job below slots into one of these rows.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy now1. Phones: an assistant for booking and practical questions
What it does. A voice assistant answers the phone, books, moves or cancels visits in your scheduling system, and answers questions you have pre-approved, such as hours, parking, and which visits need a fasting sample.
Example. A parent calls at 07:40 to move a 10-month check from Thursday to Friday. The assistant finds the right clinician's slots, offers two, confirms, and sends a text. The desk never touched it. The same assistant hears "she's been grunting when she breathes" and does not try to be clever: it says a nurse will call back, flags the call as urgent and warm-transfers to the nurse line if someone is there.
How to start. Write the red-flag list with your clinicians, not for them: the words and situations that must route to a person immediately. Typical entries might include difficulty breathing, a very young baby with a fever, not feeding, or "I'm worried" followed by silence. Your clinicians define the list; the tool only follows it. Then test with 30 scripted calls, including awkward ones, before a real parent hears it. The human and AI hybrid answering service tutorial goes deeper on how to split calls between staff, a nurse line and a tool.
Watch-outs. Siblings with the same surname, a parent calling about a child under another parent's name, and a "just a quick question" that is really a symptom.
Safe-use rule. The assistant books and informs; it never interprets a symptom. Any mention of symptoms, medication doses or "is this normal" goes to a nurse.
2. Forms: intake, school and camp paperwork
What it does. It turns a pile of forms into clean, checkable data. It reads a school physical request, pulls the child's details from fields you approve, drafts the form for a clinician to review, and tells the desk what is missing. Some practice systems offer this in-app; others do it with a separate document tool.
Example. In August a camp form arrives with a section on allergies, vaccination dates and emergency contacts. The tool pre-fills the parts that are plain facts already on file and leaves the clinical sections blank for the doctor. A before and after: previously a desk worker retyped 14 fields in about six minutes; now they verify pre-filled fields in about two, and the clinical block is still the doctor's. The vaccine recalls and school forms tutorial covers this job in detail.
How to start. Pick the single form you complete most often. List its fields. Mark each as "copy from record", "doctor only" or "parent to supply". Pilot with that one form for a month.
Watch-outs. Date-of-birth swaps, a form accidentally completed for a sibling, and the temptation to let a tool fill in a vaccine history from memory rather than from the record.
Safe-use rule. A tool may copy a fact from the chart. It never states a clinical fact, and a clinician signs every form that carries one.
3. Reminders: recalls, no-shows and waiting lists
What it does. It sends reminders by text or email at the right time, handles replies like "can we move it?" and offers cancelled slots to a waiting list. Many practice management and patient messaging systems already include reminders. The AI part is the reply handling and the wording.
Example, a quick sum. Say the office runs 90 visits a day and 6% are no-shows, which is 5.4 empty slots a day. If better reminders and a same-day waiting list fill a third of those, that is about 1.8 extra visits a day, or about 9 over a five-day week. These are illustrative figures, so measure your own no-show rate for four weeks before and four weeks after.
How to start. Keep the message short, with the child's first initial or "your child" and the time, and nothing about the reason for the visit in the text. Offer one-tap confirm, and one number to call. Before switching it on, send each reminder wording to a staff phone and look at what shows on the lock screen, because that preview is what a babysitter or older sibling may see.
Watch-outs. Texts that reveal a diagnosis on a lock screen, reminders sent to a number that now belongs to someone else, and over-messaging families with several children.
Safe-use rule. A reminder says when and where, never why. If a reply contains anything clinical, it stops being automated and goes to a human.
4. Notes: tidying a clinician's dictation, not writing it
What it does. An AI scribe listens to a visit, or a clinician dictates, and a draft note appears for review. This is the riskiest job of the seven because the output enters the medical record.
Example, a mistake and how it showed up. A draft note says "mother denies fever" when the parent said the fever had settled the night before. The error was caught because the clinician reads every note before signing and compares it with the visit. The note was corrected. The lesson is that the review step is the control, not the tool's accuracy claim. A short checklist for choosing one lives in the paediatric scribe checklist.
How to start. Pilot with one willing clinician, for two weeks, on well-child visits only. Record how long each note takes to review and correct. If correction takes longer than writing from scratch, stop.
Watch-outs. A parent's words attributed to the child, medication doses misheard, and a draft that sounds fluent while being wrong. Consent also matters: the consent wording tutorial has ready phrasing, and with children you often need the parent's agreement in the room.
Safe-use rule. Nothing enters the record until the clinician has read and signed it. The scribe drafts; it never files.
5. Reviews: replying to families without leaking a diagnosis
What it does. It drafts replies to online reviews and prompts happy families to leave one. A language model is good at tone; it is not good at knowing what it must not say.
Example, before and after. A review says "Dr [surname] was brilliant with my son's ear infection." A careless reply thanks the family and mentions the ear infection, which confirms a patient relationship and a diagnosis in public. The better reply, drafted from a template, reads: "Thank you for taking the time to write. We are glad your visit went well and will pass your kind words to the team." It confirms nothing.
How to start. Give the model a fixed template with the rules written in it.
Draft a reply to the review below for a paediatric office.
Rules: do not confirm that the reviewer is a patient, do not mention any
condition, visit, medication or date, do not use the child's name. Thank them,
keep to three sentences, invite them to call the office for anything specific.
Use plain, warm English. Review: [paste]
Watch-outs. Replying to a negative review with the facts of the visit "to set the record straight" is the classic way to breach privacy. A model will happily do it if asked.
Safe-use rule. A review reply never confirms that someone is a patient. The practice manager approves every reply before it posts.
6. Newsletter: seasonal reminders written once, checked by a doctor
What it does. It drafts the monthly newsletter and social posts: opening hours around holidays, back-to-school forms, how to book a check, a note on which clinic days have capacity.
Example. A summer newsletter has four blocks: opening hours, how to book school physicals, what to bring to a visit, and a short "when to call us" reminder. The first three are admin and can be drafted from your own notes. The last block is health content: a doctor writes or signs off that wording and the assistant only formats it. In an illustrative month the practice manager spends 25 minutes pasting notes, 10 reviewing the draft, and a doctor spends 5 on the health block.
How to start. Keep a "facts sheet" of approved wording (hours, how to reach a nurse, what the office does and does not offer by phone) and have the tool write only from it. Ask it to flag any sentence not traceable to the sheet.
Watch-outs. Invented health tips, confident but outdated advice, and "evidence" the tool cannot cite. Treat any health claim as unverified until a clinician confirms it.
Safe-use rule. Admin text may be drafted from your facts sheet; health content is written or approved by a clinician and nothing else.
7. Staff rota: drafting cover and clinic sessions
What it does. It proposes a rota from your sessions, leave, and part-time patterns, and rewrites the message when someone calls in sick. This job involves no patient data, which makes it a good, low-risk place to start.
Example, a quick sum. The office runs ten half-day clinics a week (mornings and afternoons, Monday to Friday) with four doctors, two nurses and three desk staff. A nurse's two days of leave next month leave four half-days without their usual nurse. The tool lists which half-days lose cover and suggests two swaps that keep every half-day at its minimum of one nurse and two desk staff. A person approves the swap and messages the team. For the general approach on small teams, read the AI shift scheduling versus a spreadsheet tutorial.
How to start. Write your minimum cover rules as plain sentences, such as "every half-day clinic needs one nurse on site and two at the desk". Give the tool names or initials of staff only, and keep leave reasons out.
Watch-outs. Rules nobody wrote down (who can lock up, who holds the vaccine-fridge key), and staff health details entered as a leave reason.
Safe-use rule. The rota tool never sees patient data or staff medical reasons, and a manager approves each change before it is sent.
How to choose the first two
Seven jobs are too many to launch at once. Rank them on two things: how much patient information the job touches, and how many staff minutes it saves. An illustrative scoring for the office above:
| Job | Patient data touched | Minutes saved a day | Start order |
|---|---|---|---|
| Rota | None | 10 | First, as a warm-up |
| Newsletter | None | 5 | Second, monthly |
| Reminders | Name and time only | 40 | Within the first quarter |
| Review replies | None if rules are kept | 5 | Alongside reminders |
| Forms | Chart facts | 50 | Second quarter, one form first |
| Phones | Voice, identity, symptoms | 160 | Only with a BAA and a tested hand-off |
| Notes | Everything | Varies by clinician | Last, with a signed BAA and a pilot |
The phone and forms figures are half of the booking, practical-question and forms minutes from the call table, the same share the worked example assumed. The order is not "easiest first" but "least harm first, biggest saving as soon as the controls exist". The phone assistant saves the most time and also has the most ways to go wrong. If you want help deciding on those two, my AI implementation consultation can cover mapping your own call log.
A one-page policy to paste into your staff handbook
Whatever you adopt, write down the rules in plain language. Here is a version to adapt:
- We use AI tools only for tasks listed in this policy.
- No patient or family information goes into any tool without a signed business associate agreement for that product and plan.
- No tool gives medical advice or interprets symptoms. Anything clinical goes to a nurse.
- A named person approves anything a tool writes before families see it: review replies, newsletters, forms, notes.
- We tell families when a tool is involved in a call, message or note.
- We log every error a family or staff member spots, and the practice manager reviews the log monthly.
Six lines are enough for a small practice. The monthly error log is what turns the policy into improvement: if the same mistake shows up three times, change the tool's instructions, not the staff.
Checking after a month that it works
For each job you switched on, keep one number. Phones: calls handled without a person, and calls handed off within a minute. Reminders: no-show rate. Forms: minutes per form. Notes: review time per note. Reviews: replies posted within two days. Newsletter: minutes to produce. Rota: time to rebuild after a sick call. If a number does not improve after four weeks, switch the job off and fix the instructions before trying again. A tool that makes a front desk faster but sends one wrong message a week is not working, so count errors alongside minutes.
What practice managers ask before choosing
Do these tools diagnose or triage children?
No. Everything here is front-office admin: calls, forms, reminders, notes tidy-up, reviews, newsletters and rotas. Any message that describes symptoms should reach a nurse or clinician quickly, and the tool should be set to hand it over rather than answer.
Can I paste a parent's message into ChatGPT?
Not if it identifies a child. A free or standard chat plan is not a place for patient information. Use a tool under a signed business associate agreement, or strip every identifier and keep the clinical detail out entirely.
What should I start with?
Reminders and the newsletter carry the least risk and show results in a few weeks. Phones and notes carry the most, so add them later, with a human review step and a clear hand-off rule.
How do I tell families a tool is involved?
Say so plainly on the phone greeting, the booking page and the consent form. A short sentence such as 'our assistant can book and answer practice questions; a nurse handles anything about symptoms' is enough.
Who is responsible if the tool gets something wrong?
The practice. A vendor contract can share the cost, but the office still owns the message that went out. That is why each job below has a person who checks the output before families rely on it.
Further reads
- What to Do When an AI Receptionist Gets a Booking Wrong — What to do when a phone assistant books the wrong slot.
- Is an AI Scribe Worth It for a Small Private Clinic? — A cost-and-time check for a scribe in a small clinic.
- Medical Billing AR Follow-Up: Prioritise Aged Claims With AI — Chasing aged claims once the front desk is faster.
- AI for Nurse Practitioners: Prior-Auth Letters and Practice Policies — Letters and policies for another small clinical practice.
- How to Roll Out an AI Scribe Without Losing Patient Trust — Wording and steps for telling families a tool is in use.
- Is It Safe to Put Customer Data Into ChatGPT? — What consumer and business plans do with what you type.
- How Dental Practices Use AI for Recalls and Appointment Reminders — Recall and reminder tactics from another appointment-led practice.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI Help Center articles on business associate agreements and ChatGPT for Clinicians (checked November 2026); Anthropic privacy centre article on business associate agreements for commercial customers; Google Workspace help on health-data compliance and the business associate amendment. Check each vendor's current terms before use.