AI for Sleep Clinics: Study Bookings, CPAP Follow-Up and Letters

A sleep clinic can safely hand AI the admin around the clinical work: chasing missing referral details, offering study slots, sending device pickup and return reminders, drafting CPAP follow-up messages and formatting letters from a clinician's own conclusions. It should never interpret a study, judge CPAP data or decide who is urgent.

The dividing line is simple to test. If a task would be done identically for every patient and a trained receptionist could do it from a checklist, AI can draft or run it. If it changes what happens to the patient's care, a clinician decides first and the AI only words the outcome.

Follow me on Instagram@sagnikteaches

Which sleep clinic jobs are admin and which are clinical

Sleep services have more moving parts than most outpatient clinics: a referral, a questionnaire, a home test or an overnight study, a device that has to go out and come back, a results appointment, and then months of follow-up on a therapy device. Most of the labour sits in the gaps between those steps, and that is where AI helps.

Connect on LinkedInSagnik Bhattacharya

Keep a copy of that table on the wall near the front desk. When a new idea comes up ("could it tell patients their apnoea score?"), find the row. If the answer sits in the right-hand column, the idea is out for now.

Subscribe on YouTube@sagnikteaches

A booking flow for home and in-lab studies, with the sums

Take a clinic that receives about 60 referrals a month and runs a mix of home sleep tests and overnight in-lab studies. This is an illustration with round numbers, not a benchmark. Before any automation, a receptionist spends roughly 14 minutes per referral: reading it, phoning or emailing for a missing date of birth or referrer detail, finding a slot, sending instructions, and chasing the questionnaire. That is 60 x 14 = 840 minutes, or 14 hours a month, and the figure does not include the patients who never answer the first call.

Here is the flow after a careful set-up, step by step.

1. Intake check. A new referral, once typed or scanned into your system, goes to an AI step with a checklist: patient name, date of birth, contact number, referrer name and contact, reason as written, test requested if any, and whether the patient needs an interpreter or accessibility support. The AI returns a list of what is missing, never a view on the referral itself.
2. Missing-detail request. The clinic coordinator reads the list, clicks approve, and a pre-written message goes to the referrer. This takes about 40 seconds instead of five minutes.
3. Triage by a clinician. The referral waits in a queue until a clinician marks the test type and priority. No AI is involved, and the system should not let the next step start until this is done.
4. Slot offer. The coordinator picks the test type; your booking tool offers the next three slots that fit your rules (for instance, in-lab studies Tuesday to Thursday, home device collection before 3 pm). The patient picks one by link or replies to a text.
5. Instructions pack. Confirmation and preparation instructions go out automatically, using wording a clinician approved once.
6. Chasing. If the questionnaire has not come back 48 hours before the study, a reminder goes out; if it is still missing 24 hours before, the coordinator gets a task instead of the patient getting a third message.

If the set-up works, the receptionist's time per referral might fall from 14 minutes to around 6: the steps that remain are the human ones, such as approving the missing-detail message and handling the patient who rings back. That would be 60 x 6 = 360 minutes, a saving of 8 hours a month. Treat those figures as a target to test over a month, because your referral quality and booking tool will move them in either direction.

The prompt for the intake check

Used on a business plan after removing any detail you do not need (see the data section below), this prompt keeps the AI inside its lane.

You are helping a sleep clinic coordinator check a referral for completeness.
You are not a clinician. Do not comment on the clinical content, the likely
diagnosis, urgency or the appropriate test.

From the referral text below, list:
1. Required fields present (name, date of birth, phone, referrer name, referrer contact, reason as written).
2. Required fields missing or unclear.
3. Anything that suggests the patient needs adjustments (interpreter, mobility, carer, hearing) exactly as written.
4. A two-sentence message to the referrer asking only for what is missing.

If the referral text is not clear enough to decide, say so. Do not guess.

Referral text:
[paste]

An illustrative output for a referral with no referrer phone number and a handwritten date of birth that reads two ways would list those two items, quote the handwriting, and draft: "Thanks for the referral for [first name]. Could you confirm the date of birth (we read it as 04/07 but the year is unclear) and give a direct number for your practice? We will book as soon as we have these." What you fix: the AI must never fill the gap with its best guess. If it does, tighten the "do not guess" line and test again.

Study-night reminders that people act on

A missed home test costs a device cycle; a missed in-lab night costs a bed and a technician for the evening. Reminders are the cheapest fix and the best place to start, because the content is the same for nearly every patient.

A workable sequence for an in-lab study looks like this (the timings are a starting point to adjust against your own no-show data):

• On booking: confirmation with date, arrival time, what to bring and how to cancel.
• Seven days before: a short message with the preparation instructions your clinicians approved, such as avoiding caffeine after a set time or bringing regular medication.
• Two days before: a "reply YES to confirm or call us to change" text. A non-reply creates a task for a person, not a third text.
• Day of the study: arrival details and a number to call.
• Morning after: a brief thank-you with what happens next.

The AI's role is to draft the wording once, vary it for tone and readability, and translate it where you serve other languages (with a native speaker checking). It does not need to be in the loop at send time. Sending is the job of your booking system or a text-message tool.

One practical detail saves money: texts longer than 160 characters are split and billed in segments by many providers, and a single curly apostrophe or an emoji can drop the limit for that message to 70 characters. When the AI drafts reminders, tell it to use plain straight quotes, no emoji and a hard limit of 160 characters, then paste the result into your sender and check the segment count.

Write four versions of a sleep study reminder text. Rules: under 160
characters, plain straight quotes, no emoji, no medical advice, ends with the
clinic phone number placeholder [clinic number]. Purpose: ask the patient to
reply YES to confirm their study on [date]. Audience: adults, reading age 12.

Device pickup and return for home tests

Home sleep test devices are expensive and easy to lose track of. Where a clinic posts or hands out devices, the booking record should carry two dates: when the device goes out and when it is due back. Three reminders cover most late returns: a "your device is due back tomorrow" message, a "we have not received it" message on the day after, and a task for a coordinator on day three.

Consider a four-person clinic with six reusable home devices. If two are overdue by a week at any time, a third of its capacity is idle. Putting the return date into the reminder rule, and having the AI draft three escalating but friendly wordings, is a small fix with a measurable result: count device-days idle per month before and after.

A before and after shows the tone. Before: "Your device is overdue. Please return it immediately." After: "Hi [first name], your sleep test device is due back by [date]. If anything went wrong with the test, just tell us; we would rather know than have you worry. Return options: [instructions]." The second keeps the facts and invites a problem report, which also catches patients who did a bad recording and are embarrassed to say so.

CPAP follow-up: the reminder ladder and the clinical hand-back

Therapy follow-up is where a clinic's workload grows quietly. Every patient started on CPAP creates a series of checks, and the early ones matter. How many and when depends on your clinic's protocol and any adherence rules set by your payer or commissioner, so do not copy anyone else's schedule. What AI can do is run whatever schedule you have, consistently.

1. Define the ladder with your clinical lead. For example: a "how is the mask feeling?" check a couple of days after set-up, another after a week, one at a month, and one nearer three months. Write down who owns each rung and what triggers a call.
2. Draft the messages. Each one asks one question and gives a reply route: a number to call for mask problems, a link to book a mask refit, a line to say "all fine".
3. Log replies in a defined way. "All fine", "mask problem", "machine problem", "other" and "no reply" become tags. The AI can suggest a tag for a free-text reply, but a person confirms any tag other than "all fine".
4. Route by tag. A mask problem creates a refit appointment task; a machine problem notifies the equipment technician; "other" goes to a nurse; "no reply" after two messages goes to a coordinator for a phone call.
5. Hand back to the clinician for anything clinical. If a reply says "I wake up gasping" or mentions chest pain, the rule is not an AI decision. The message tool should detect certain words and route the thread to a person within a defined time, and the auto-reply should say who to call urgently.

A realistic mistake: a clinic sets its message to "Your data looks great, keep it up!" for everyone who replied "fine". A patient with a leaking mask who thought the leak was normal gets praised for a poor night. The fix is to keep the message about the patient's experience ("Thanks for letting us know how it feels") and leave any comment on device data to the clinician's call.

Letters: what to draft and what to sign

Sleep services generate a steady stream of letters: acknowledgements to referrers, appointment letters, study preparation letters, results-appointment invitations, and update letters to the referring doctor after review. The risk is rarely typing. It is a sentence that says something clinical the clinician did not say.

Use two letter classes. Class A letters contain no clinical content: acknowledgements, booking confirmations, preparation instructions, "please contact us" chasers. A coordinator can approve these from a template filled by AI. Class B letters contain findings, diagnoses or therapy decisions. For those, the clinician writes or dictates the conclusions and the AI formats them into the letter layout, correcting grammar and setting out headings. It must not add to what they said.

Format the clinician's notes below into a letter to the referring doctor using
the headings: Reason for referral, Test performed, Clinician's conclusion,
Plan. Use only the content provided. Do not add, soften, explain or infer
anything. If a heading has no content, write "Not provided". Keep the
clinician's wording for conclusions. British English. Do not include any
information not in the notes.

Notes:
[clinician text]

Illustrative result: the clinician's notes say "Home test. Moderate obstructive sleep apnoea. Offered CPAP. Review at 3 months." The letter comes back with the four headings filled and, if the prompt is working, "Not provided" under reason for referral because the notes do not say. A weaker prompt fills that gap with a plausible reason, and that is the failure to test for. Send the same notes through three times with one heading deliberately blank. If any run invents content, the prompt is not ready.

The review checklist that stops a wrong letter leaving

• Is the patient's name, date of birth and reference correct on every page?
• Does every clinical statement appear in the clinician's own notes, word for word or close paraphrase?
• Is any number (hours, events, dates) copied from the source, not rewritten?
• Does the plan match what the clinician recorded, including who will do the follow-up?
• Has the clinician initialled it as approved?

Pair this with the habit described in the human review without slowing down tutorial: one named reviewer, one checklist and a record of the corrections. After a month the corrections log tells you which letter types are safe to review more lightly.

Patient data: what goes into an AI tool and what stays out

A sleep clinic holds health data, and a date of birth plus a referral reason is enough to identify someone. Three rules keep the set-up defensible.

1. Prefer tools inside systems you already have. If your booking or practice-management software has built-in drafting and a data agreement in place, use that before adding another vendor.
2. Remove identifiers when you can. A letter-formatting task needs the clinician's text, but not necessarily the full name or patient number. Replace names with "[patient]" and add identifiers back in the final document. The anonymising tutorial shows the replace-and-restore routine.
3. Do not use free consumer plans for anything identifying. Business plans of the main chat assistants state that your content is not used for training by default, and both ChatGPT Business and Claude Team list at $20 a seat a month on annual billing ($25 monthly) with a two-seat minimum. Where your law requires a business associate agreement or a data processing contract, get it signed for the specific product before patient information goes in, and keep a copy. The standard team plans are usually not the covered ones: as of November 2026, OpenAI's list of products available with a business associate agreement names ChatGPT for Healthcare, ChatGPT for Enterprise with a regulated workspace and ChatGPT for Clinicians, not ChatGPT Business, and Claude's pricing page shows its healthcare-ready offering on Enterprise plans, not Team. If you are not sure whether your plan is covered, treat the answer as no.

The data and consent checklist for clinic chatbots covers the vendor questions in more detail, and the BAA explainer sets out what an individual clinician's agreement does and does not extend to.

Five small jobs worth trying, in the order I would try them

1. Pre-study instructions rewritten for readability. Give the AI your current instruction sheet and ask for a version at a reading age of 12 with the same content. A clinician checks that no instruction changed. Quick win, no patient data.
2. A "what to expect" text for first-time patients. Patients who know what the sensor feels like and how long set-up takes are less likely to cancel on the night.
3. Late-questionnaire chaser. One message, three tones, picked by the coordinator.
4. Mask-refit invitation. A reply tagged "mask problem" triggers a template inviting a refit, with the real booking link.
5. Weekly admin digest. Paste an anonymised list of open tasks (referrals awaiting triage, devices out, questionnaires missing) and ask for it grouped by owner and age. The coordinator starts Monday knowing what is stuck.

How to tell it is working after a month

Pick four numbers before you start and write them down: referral-to-booking days, no-show rate for studies, devices overdue by more than three days, and coordinator minutes per referral. After four weeks compare. If referral-to-booking time improves but no-shows do not, the reminders need rewording or retiming. If coordinator minutes do not drop, find the step still done by hand; it is usually the missing-detail chase or exception handling.

Also keep a corrections log: every time a person changes an AI-drafted message or letter, tick a reason (wrong fact, tone, missing item, invented content). Invented content on even one patient letter means pause that letter type and tighten the prompt before continuing. Related routines such as cutting missed appointments and using AI at the front desk safely use the same measurement habit.

Where this goes wrong in a sleep service

• Urgent symptoms hiding in a reply. A patient answers a routine check-in with something serious. The fix is a keyword route to a human and an auto-reply that names an urgent contact. Test it with ten invented replies before go-live.
• A wrong slot offered. The booking rules do not know that bariatric beds exist only on certain nights. Write exceptions into the rules, not into memory.
• Reminder fatigue. Five texts for one study feels like harassment. Cap it at three before the study and none after the morning-after message unless there is a reason.
• A letter template quietly edited. Version your templates; a changed paragraph that reaches every letter is the AI-era version of a stale form.
• Consent gaps. Contact preferences collected at referral should decide whether text, email or post is used. If you have no record, post is the safe default.

If you would rather map the whole referral-to-follow-up chain with someone, that is the kind of job my AI implementation consultation is built for, but the steps above are enough to start on your own this week.

Questions sleep clinic managers ask before starting

Can AI tell a patient whether their CPAP data looks good?

It should not. Whether usage, leak or residual events are acceptable is a clinical judgement. The safe version has the clinician mark a patient as on track or needing a call, and the AI only drafts the message that matches that decision.

Do we need patient consent to send automated reminders?

Usually you need a lawful basis and a clear way to opt out, and many clinics also record contact preferences at booking. Rules differ by jurisdiction, so ask your data-protection adviser which basis and wording your clinic should use before switching reminders on.

Which plan should the front desk use for drafting letters?

Use a business or team plan whose terms say your content is not used for training, and get a data-processing or business associate agreement in place where your law requires one. Free consumer plans are the wrong place for anything that identifies a patient.

How many letters should a person review before sending?

All of them at first. Once a letter type has gone through twenty or more without a correction, you might move routine acknowledgements to spot checks of one in five. Anything that mentions a diagnosis or a result stays at one hundred per cent review.

Further reads

• AI for Healthcare Clinics: Referral Tracking and Chase Letters — A chase-letter routine for referrals that stall.
• Patient Intake Automation in a Small Practice: Forms, AI Summaries — Build the pre-study questionnaire as a digital form.
• Telling Patients You Use AI Note-Taking: Consent Wording That Works — Consent wording you can adapt for AI-assisted admin.
• Is an AI Scribe Worth It for a Small Private Clinic? — If the clinical notes themselves are your next question.
• 10 Admin Tasks a Small Clinic Can Hand to AI This Month — Ten more admin jobs a small clinic can hand over.
• AI Agents for Healthcare: Six Safe First Jobs for a Small Practice — Safe first jobs for an AI agent across a small practice.
• Healthcare Appointment Scheduling Automation for a Small Practice — Reminders, confirmations, waitlist backfill and online booking a small clinic can automate safely, the privacy limit for each, and the line AI must not cross.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Twilio's SMS character limit documentation; ChatGPT Business and Claude pricing pages; OpenAI's help article on products eligible for a business associate agreement; the clinic workflow and figures are illustrative.