A small healthcare practice can safely give an AI agent admin jobs first: appointment reminders and rebooking, intake form chasing, referral paperwork tracking, answering fixed FAQs from an approved fact sheet, filling cancelled slots from a waiting list, and sorting the shared inbox into drafts for staff. Clinical judgement stays human.
"Agent" simply means software that takes steps for you, such as sending a message or moving a booking, instead of only writing text. That is useful and also why each job needs a boundary: what it may do, what it may see, and who can stop it. A quick test of readiness: if you cannot write the job's boundary in one sentence ("it may do X using Y and never Z"), it is not ready to hand to an agent.
What stays human, and why it is not up for negotiation
Start with the list of things an agent must not do in a small practice, because it settles most arguments before they start.
- Triage and urgency. Deciding who needs to be seen today, or whether a symptom is serious.
- Interpreting results, images, readings or device data, or telling a patient what a result means.
- Medication decisions: doses, changes, refills that need a clinician's sign-off, or advice on interactions.
- Diagnosis or reassurance, including "that sounds nothing to worry about".
- Consent conversations for treatment, and decisions where a patient lacks capacity.
- Safeguarding and risk judgements, including reading between the lines of a message from a vulnerable person.
- Breaking bad news, handling complaints about care, and answering a distressed caller.
Pin that list next to the first agent's set-up notes. Each time someone proposes a new job, find the line it would cross. If it crosses one, the job is out for now, however clever the tool. Medical device regulators in many markets treat software intended for a medical purpose, such as informing a diagnosis or a treatment decision, as a medical device in its own right, and that is territory you stay out of by sticking to admin.
How to score a candidate job in five minutes
Rate each possible job from 1 (low) to 5 (high) on four questions. Add them up; lower is safer.
| Question | 1 means | 5 means |
|---|---|---|
| How identifiable is the data it needs? | None | Full record with clinical detail |
| How bad is a wrong action? | Mild annoyance | A patient could be harmed |
| Can the action be undone? | Easily, in a minute | Not at all |
| How much does it act without a person looking? | Drafts only | Sends or changes things alone |
A reminder text scores something like 2 + 1 + 2 + 3 = 8. An agent that reads a clinical inbox and replies to patients scores 5 + 5 + 4 + 4 = 18, and does not go live. A sensible rule for the first month: only jobs scoring 10 or below. The six below all do.
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. Appointment reminders, confirmations and rebooking
What it does. Sends confirmation and reminder messages, accepts "confirm", "cancel" or "move" replies, and offers the next free slots within rules you set (for example, same clinician, same appointment type, within two weeks).
An illustrative sum. A four-clinician practice runs about 2,400 appointments a month and loses 8 per cent to no-shows, which is 192 slots. If better reminders and one-tap rebooking cut that to 6 per cent, that is 144 lost slots, so 48 slots a month are recovered. At an assumed $60 average value per slot that is $2,880 a month of capacity, before counting staff time. These are invented figures to show the arithmetic; use your own no-show rate. The tutorial on using AI to cut missed appointments has the reminder timings and the measurement.
How to start. Switch on the reminder and online-rebooking features inside your existing booking or practice-management system, because the data stays where it already lives. Draft message wording once with AI, then lock it. Keep messages under 160 characters with straight apostrophes and no emoji, so a text is not split into several billed segments.
Boundaries and watch-outs. Messages never include the reason for the visit. A reply containing anything that is not confirm, cancel or move goes to a person. Check that opt-outs are honoured every time.
2. Chasing intake forms and checking they are complete
What it does. Sends the form link after booking, reminds the patient once or twice, and checks the returned form for missing fields (date of birth, contact number, emergency contact, consent boxes), producing a list for staff. It does not summarise the clinical content for the clinician in this first version.
Example. Forty new-patient forms a month, each taking a receptionist six minutes to check and chase by phone, is four hours. An agent that does the first chase and the completeness check might leave 15 forms needing a call at three minutes each, under an hour. Again an illustration: time your own checks and chases for a week to get the real figure before you buy anything.
How to start. Use the form tool already inside your practice system. If you must use a separate form builder, treat it as a processor of patient information and check its agreement before putting a real form through it.
Boundaries and watch-outs. The agent flags gaps only. It never fills in a missing answer, and it does not decide whether an answer is a clinical concern; any free-text "reason for visit" goes straight to a person.
3. Tracking referrals and drafting chase letters
What it does. Keeps a list of outgoing and incoming referrals with dates, flags any with no response after a set number of days, and drafts a polite chase message from a template for a person to approve. It can also format a clinician's finished text into your letter layout without altering it.
Example. A practice sends about 30 referrals a month and has a standing rule that anything unanswered after 14 days is chased. Today that rule depends on someone remembering. With a tracked list, a Monday digest names the six referrals past 14 days, each with a draft chase attached. The detailed routine is in referral tracking and chase letters.
Boundaries and watch-outs. The draft says what is needed, never why it matters clinically. Letters containing findings are written by the clinician, and AI only formats them. A sign-off step stays in place for every outgoing letter.
4. Answering fixed questions from an approved fact sheet
What it does. A website or text assistant answers repetitive questions: opening hours, how to register, where to park, how to order a repeat prescription, what to bring, how to cancel. It answers only from a fact sheet you write and approve, and hands over to a person for anything else.
Example. The fact sheet has 25 entries of one to three sentences each. A visitor asks "Do you do flu jabs and how do I book?" and gets the clinic's own paragraph and booking link. A visitor asks "I've had chest pain since this morning, what should I do?" and gets a fixed message that this is not something the assistant can help with, with the practice's urgent contact route, the local emergency number and a handover to staff. That second behaviour is the core of the safety design and needs testing with at least ten sample messages before go-live. See the dental fact-sheet chatbot for how to write and test the sheet.
Boundaries and watch-outs. The assistant never asks for a patient's name or condition in the chat, and its opening line says it cannot give medical advice. Review the chat log weekly and add any missing question to the sheet. Prices change; a fact sheet with an out-of-date fee is a fault, not a feature, so put a review date on each entry. The assistant should be set to say "I don't know" when the sheet has no answer. Costs vary from free tiers to a few hundred dollars a month; for instance, one AI receptionist vendor lists a free plan for 25 calls a month and then $3 a call, but check current prices and whether the vendor will sign an agreement for patient information before choosing.
5. Filling cancelled slots from a waiting list
What it does. When a slot opens, the agent offers it by text to people on a waiting list who fit the appointment type, in order, and books the first to say yes, then updates the diary.
Example. A physiotherapy-style clinic has 12 cancellations a month with under 48 hours' notice. If the waiting-list agent fills half of them, six extra appointments are held. The steps are described in filling cancellations from a waiting list.
Boundaries and watch-outs. Waiting-list order is a rule set by a person, and urgent or clinically prioritised patients are never skipped by the agent's offer. Agree an offer window (say 30 minutes) and what happens at the end of it. A slot offered twice is a classic failure, so test it with two simulated replies at once.
6. Sorting the shared inbox into labelled drafts
What it does. Reads incoming emails to the practice address, labels them (appointment, prescription, referral, billing, other), and for the routine ones drafts a reply for a person to send. Nothing leaves without a human click.
Example. Of 60 emails a day, perhaps 20 are routine appointment requests that have the same four-line answer, and 10 are billing queries. The labelling alone halves the time to find the urgent ones. The drafts are bonus.
Boundaries and watch-outs. This is the riskiest of the six, because emails contain clinical detail. Do it only in a mail system with built-in AI under your existing agreement, with a data processing contract in place, or in a tool that signs one. A clinical or urgent word triggers a "needs a person today" label and no draft. A rule I would add: no draft is created for any email that mentions a symptom, medication or result.
The checks every tool must pass before it touches the practice
For each of the six jobs, the tool that runs it needs the same checklist. Fill in one row per tool and keep it in the practice file.
| Check | Question to answer |
|---|---|
| Data seen | Which patient fields does it need? Can the job work with fewer? |
| Consent and notice | Does your privacy notice cover this use? Are patients told a message is automated and how to reach a person? |
| Agreement | A business associate agreement or data processing contract where your law requires one, signed for this product and plan |
| Training and retention | Is patient information excluded from model training? How long is it kept and where? |
| Access | Which staff can see logs? Does it use its own login with limited permissions? |
| Human override | Who can switch it off, and how fast? |
| Record of actions | Can you see what it sent or changed, with dates? |
The agreement line is where small practices slip. Never assume an agreement signed for one clinician's individual account covers the reception team or the practice; read whom it names. And a business-plan chat assistant (ChatGPT Business and Claude Team both list at $20 a seat a month annually, or $25 monthly, with a two-seat minimum) does not train on your content by default, but that is a different protection from a signed agreement for patient information. The vendors' own pages make the gap plain. OpenAI's list of products it offers with a business associate agreement, checked in November 2026, names ChatGPT for Healthcare, ChatGPT for Enterprise with a regulated workspace, ChatGPT for Clinicians and some API set-ups, and does not include ChatGPT Business; Claude's pricing page marks its healthcare-ready offering as unavailable on Team and available on Enterprise. If the vendor has not confirmed in writing that the product and plan are covered, treat patient data as not allowed in it. The wider clinic chatbot data and consent checklist expands each row.
Why not start with a general-purpose agent
General agents that browse and click, such as ChatGPT Work or Claude in Chrome, are impressive and tempting. For a small practice's first job they are the wrong tool. They act with whatever accounts you let them reach, so they can see far more than the task needs, and the action log is not designed for a regulated environment. Keep them for non-patient work, such as researching suppliers or tidying a public rota. The checklist in what can go wrong when AI agents take actions for you explains the risks in more depth.
A four-week plan for the first job
- Week 1: choose and score. Pick one of the six, score it, and write its boundary in a sentence: "It may do X using Y and never Z."
- Week 2: set up and test with fake data. Use invented patients. Try ten odd replies, such as a symptom, a rude message, an unclear "maybe", a wrong number and a patient who says "stop".
- Week 3: pilot with a small group. One clinician's list, or one appointment type. A person reads every message sent for the first three days.
- Week 4: review and decide. Count messages, replies that needed a person, errors and patient complaints. Keep, change or stop. Only then pick the second job.
The practice manager owns the plan; the clinical lead owns the boundary list. Having both names written down avoids the common situation where everyone assumed the other had checked.
A worked first month in a six-person practice
Here is how the plan might run, with invented but plausible numbers. The practice has two clinicians, three reception and admin staff and a practice manager, and about 1,100 appointments a month. The manager scores the six jobs and picks reminders and rebooking (score 8) as the first, with the intake-form check as the planned second.
Week 1. She records the baseline: 9 per cent no-shows, or about 99 missed slots, and reception spending around ten hours a week on reminder calls. She writes the boundary: "The agent may send confirmation and reminder texts and offer two alternative slots using the booking system. It never mentions the reason for a visit, never answers anything other than confirm, cancel or move, and passes every other reply to reception." The clinical lead initials it.
Week 2. She switches on the reminder feature in the booking system and uses AI to draft three versions of each message. They test with six invented patients and ten awkward replies. Two failures appear: a reply of "can't come, feeling awful" is treated as a plain cancellation and a slot is offered to the waiting list without any human noticing, and a "stop" reply does not stop the next reminder. Both are fixed before any real patient is involved: any reply containing words from a short list (unwell, pain, worse, bleeding, emergency) now goes to a person and pauses the rebooking, and the opt-out is checked with three more tests.
Week 3. One clinician's list goes live. Reception reads every message sent for three days, and the log shows 140 messages, 61 confirmations, 9 cancellations, 7 rebookings and 5 replies passed to staff. One text arrived in a patient's name but on a shared family phone; the team adds a line to the notice about messages going to the number on record.
Week 4. After the first full week with both clinicians, the no-show rate for the pilot list is 6 per cent against a 9 per cent baseline, and reception reports two hours less phone time. She does not conclude the change is permanent from a single month of data, so the plan is to run another month before extending to the second clinician's list and starting job two. That patience is the point of the four-week structure: errors found at 140 messages are cheap, errors found at 1,100 are not.
A final point on order. If you can only afford one job this quarter, the choice usually comes down to where the time goes today. A practice drowning in phone calls starts with reminders and the fact-sheet assistant; one whose receptionists spend afternoons chasing paperwork starts with intake forms and referral tracking; one with long gaps in the diary starts with the waiting-list job. Count the hours for a week before deciding. A sheet with six rows and a tally of minutes beats any opinion in the staff room, and it gives you the baseline you need to prove the first job worked.
Wording for patients: the notice and the first message
Patients should know when a message is automated and how to reach a person. Two short pieces of wording cover most of it. Ask your data-protection adviser to check them for your situation.
Privacy notice addition (adapt):
"We use automated messaging to send appointment reminders and to offer
alternative times. These messages use the contact number we hold for you
and do not include the reason for your visit. You can opt out at any time by
replying STOP or telling reception. A member of staff reads any reply that
is not a simple confirm, cancel or move."
First reminder text (under 160 characters):
"Reminder: your appointment at [clinic] is on [day] at [time]. Reply C to
confirm, M to move, X to cancel. Questions: call [number]."Notice what the reminder leaves out: the clinician's name, the type of appointment and anything about the reason. A text that lands on a shared or lost phone should reveal as little as possible.
Measuring whether it helped
Choose one number per job and record it before you start. For reminders, the no-show rate. For intake, the share of forms complete before the visit. For referrals, the number past 14 days without a reply. For the FAQ assistant, the number of routine calls per week. For the waiting list, slots filled from cancellations. For the inbox, the median time to a first reply. After a month, compare. If the number did not move, the job was not the bottleneck, or the set-up needs a change; either is useful to know before you add the next one.
Also record errors, because that number has a veto. One wrong appointment message in 500 may be tolerable if it is fixable; one message that reveals the reason for a visit to the wrong person may be a reportable data breach, and the agent stops until the cause is fixed. Decide now what you will count as an incident and who needs to know within what time.
Where small practices go wrong with agents
- Buying the platform before choosing the job. A tool with ten features invites ten experiments. Choose one job and a tool that does that job.
- Treating a demo as a test. Vendors show tidy cases. Test your own awkward ones.
- Skipping the patient notice. People should know when a message is automated and how to reach a human. Update your privacy notice before launch, not after.
- Forgetting the out-of-hours case. An agent that replies at 11pm to a worried message must say what to do in an emergency, every time.
- No owner. Someone has to read the weekly log. If nobody will, do not switch it on.
If you want a second pair of eyes on which job to start with, that is what my AI implementation consultation covers, though the scoring table and the checks above are enough to make a safe first choice yourself.
Questions practice owners ask before using an AI agent
What is the difference between an AI agent and a chatbot?
A chatbot answers; an agent also acts, such as sending a text, moving an appointment or updating a record. That extra step is why an agent needs tighter rules: limited permissions, a list of allowed actions and a person who can see and stop what it did.
Can an AI agent triage patients?
Not in a small practice's first deployment. Deciding urgency is a clinical judgement with real harm if wrong. An agent can collect the same fixed questions your staff already ask and pass the answers to a person, but it should not rank or reassure.
Do we need a contract with every AI tool?
If a tool touches information that identifies a patient, yes, check before use: a business associate agreement or data processing contract as your law requires, plus the vendor's retention and training terms. If a tool sees no patient information at all, you may need nothing beyond normal vendor terms.
What is the cheapest safe job to start with?
Reminder and rebooking messages inside the booking system you already pay for, because the data stays in that system. The next cheapest is a fact-sheet FAQ answerer that never sees a patient record.
Further reads
- 10 Admin Tasks a Small Clinic Can Hand to AI This Month — Ten more admin tasks a small clinic can hand over.
- How a Small Clinic Can Use AI at the Front Desk Safely — The front-desk version of this first-job thinking.
- How to Prepare Your Small Business for AI Agents — Prepare your processes before any agent goes live.
- Virtual Receptionist for a Small Clinic: Live, AI or Hybrid? — Live, AI or hybrid cover for the phones.
- AI for Sleep Clinics: Study Bookings, CPAP Follow-Up and Letters — A specialty example: bookings, reminders and letters.
- Patient Satisfaction Survey With AI: Questions and Comment Summaries — Add a short survey and a comment-summary routine.
- Patient Intake Automation in a Small Practice: Forms, AI Summaries — Design the intake form the chasing job depends on.
- ChatGPT for Clinicians: Does the Individual BAA Cover Your Practice? — What an individual clinician's agreement does and does not cover.
- Best AI for Healthcare Providers: What to Buy First by Pain Point — A buying order for a small practice: price your missed calls, late notes and denied claims, then pick the phone agent, scribe or claim check.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: ChatGPT Business, ChatGPT and Claude pricing pages; OpenAI's help article on products eligible for a business associate agreement; Smith.ai AI receptionist pricing; the International Medical Device Regulators Forum definition of software as a medical device; Twilio's SMS character limit documentation; the practice figures and sample messages are illustrative.