AI Secretary for Doctors: Referral Letters, Inbox, Insurer Forms

A solo doctor can use an AI secretary for three jobs: referral letters from a fixed template, inbox sorting into labelled piles, and first-draft insurer forms. Start with letters. Choose a tool with a signed data agreement, give it a template, and read every draft before anything leaves the practice.

The hard part is the tool choice, not the prompts. Most "AI secretary" offers are virtual-secretary services that answer the phone and book appointments. The paperwork that piles up after clinic ends is a different job, and it puts identifiable patient details in front of whatever tool drafts it. So decide what each tool may see before you decide which one you like, write templates with placeholders, and keep a no-paste list beside the screen.

Follow me on Instagram@sagnikteaches

Choose the tool by what it is allowed to see

The order of decisions matters. Pick the data rule first and the product second, because the product you like best may not be allowed to see patient information at all.

Connect on LinkedInSagnik Bhattacharya

There are three tiers of risk, and each job in this tutorial fits one of them:

Subscribe on YouTube@sagnikteaches

A sensible path is to start at tier A to write the templates, then move to B or C for live use. If you cannot get an agreement for a tool, it stays at tier A, whatever its other merits. Ask each vendor three questions in writing: will you sign an agreement covering patient information for this plan; are my inputs used to train models; and where is the data stored and for how long.

Business plans from the big assistants generally do not train on business content by default, but a contract for patient data is a separate matter and is often limited to certain plans or to the vendor's API. OpenAI shows the gap clearly. ChatGPT Business is listed at $20 a user a month billed annually or $25 billed monthly, for teams from two users, and OpenAI does not train on Business workspace content by default. Yet its help centre states that it does not offer a business associate agreement for ChatGPT Business. Its agreements cover the API, sales-managed Enterprise and Edu accounts, and a separate, free ChatGPT for Clinicians plan that verified individual clinicians can join where OpenAI offers it. So a two-seat Business plan stays a tier A tool in a medical practice, however good its letters are. Check the current terms for the exact product before you paste anything identifiable.

What never goes into a consumer tool

Write this list on a card and put it where your staff will see it. The rule is the same whichever product you choose:

• Full names, dates of birth, addresses, phone numbers, email addresses and record numbers, in any combination with clinical details.
• Photographs, scanned letters, test results and screenshots of the patient record.
• Audio of consultations or voicemail messages.
• Anything about a child, a safeguarding concern, mental-health history, sexual health or substance use, unless the tool is contracted for it.
• Insurer policy numbers, claim references or bank details.
• Another clinician's name alongside a patient's case.

Two habits make the list workable. First, write templates and prompts with placeholders such as [patient initials] and [date of birth], and fill them in only inside a system that is allowed to hold them. Second, keep one browser profile or one app for approved tools and nothing else, so a consumer chatbot is never one click from a patient's letter.

Job 1: referral letters from a fixed template

Referral letters are the best first job. They have a predictable shape, a human reads the output before sending, and the saved time is obvious. Say a solo general practitioner writes about eight referrals a week, ten minutes each. That is roughly 80 minutes a week, and drafting to a template might bring each one to about four minutes of checking and editing, saving an hour or so a week. These are illustrative numbers; time your own for a week first.

Step 1: build the template at tier A

Use an assistant with no patient data to write the letter skeleton you want. A prompt for the skeleton:

Write a referral letter template for a general practitioner to a specialist.
Sections, in this order, each with a placeholder in square brackets:
1. Reason for referral and urgency (routine, soon, urgent)
2. Relevant history (three lines maximum)
3. Current medicines and allergies
4. Examination and investigations so far
5. What I would like the specialist to do (assess, advise, treat, scan)
6. What the patient has been told
Keep it to one page. Plain language. No diagnosis unless I enter one.
Use "[ ]" for any detail I must supply. Never invent clinical details.

Edit the result until it sounds like you. Save it as a text file in your practice folder, one per specialty if the receiving departments ask for different things (cardiology wants an ECG mention, dermatology wants lesion description and photographs).

Step 2: draft from your own shorthand

For live use at tier B or C, give the tool the template and your quick notes, not the whole record. Here is the pattern, with a sample of the notes and what comes back, labelled as illustrative:

Your notes:
Pt [initials], 61. 3 months intermittent chest tightness on stairs, no rest pain.
ECG normal today. On ramipril 5 mg, atorvastatin 20 mg. NKDA. Ex-smoker.
Want exercise tolerance test / cardiology opinion. Routine. Told him it's likely angina until proved otherwise but not an emergency.

Draft back:
Reason for referral: Three months of intermittent exertional chest tightness, no rest symptoms. Routine referral requested for assessment, including consideration of exercise tolerance testing.
Relevant history: Ex-smoker.
Medicines: Ramipril 5 mg, atorvastatin 20 mg. No known drug allergies.
Investigations: Resting ECG normal today.
Request: Cardiology opinion and exercise tolerance test.
Patient informed: Symptoms may represent angina; not an emergency.

What you would fix: the draft wrote "may represent angina" in the patient-informed line, which is softer than your "likely angina until proved otherwise", and it left out the age. Both are the kind of drift to expect: tone softening and small omissions. You adjust the wording, add "61-year-old man", and sign.

Step 3: the five-point check before sending

1. Is the patient correct (name, date of birth, record number)?
2. Do the medicines and doses match the prescribing screen?
3. Is the urgency what I decided?
4. Is every clinical statement something I said or recorded, with nothing added?
5. Does the ask match what the specialist can actually do?

For the error types that creep into AI drafts, the same list used for notes applies: omissions, invented details, wrong doses, flipped negatives, mixed-up people and carried-over text. See six note errors to catch before you sign for how each looks in practice.

Chasing the reply

A letter sent is not a referral completed. A simple spreadsheet with the date sent, specialty and the date you expect a reply lets you draft a chase letter when a referral is overdue. The workflow is covered in referral tracking and chase letters for clinics, and it needs only initials and dates, not clinical detail.

Job 2: sorting the inbox into labelled piles

An inbox is where good intentions meet a hundred messages a day. The job for AI here is sorting and summarising, never answering clinical questions. The goal is a short, ordered list that tells you what to read first.

Pick five labels and one urgent rule

Illustratively, a solo practitioner's sort might use these labels:

The urgent rule is the safety net: any message mentioning a list of red-flag words (chest pain, breathless, bleeding, suicidal, severe, allergic reaction) goes to "Urgent clinical" regardless of what else the sorter decides, and anything it is unsure about goes there too. Make the sorter lean towards urgent. A false alarm costs you thirty seconds; a missed one costs far more.

Run it on de-identified samples first

Before connecting a tool to a live mailbox, test the rules on ten to twenty old messages with the names and dates removed. Here is a prompt and an illustrative result:

You sort messages for a doctor's inbox into five labels: Urgent clinical, Results and letters in, Prescription and sick-note requests, Admin and appointments, Marketing and vendors.
If a message mentions chest pain, breathlessness, bleeding, suicidal thoughts, a severe reaction, or you are unsure, choose Urgent clinical.
Do not answer the message. Output one line per message: label, then a summary of no more than 12 words.

Message 1: "Hi, I took the new tablets and my lips are swelling a bit, should I stop them?"
Message 2: "Can I get a repeat of my usual blood pressure tablets before Friday?"
Message 3: "Exclusive offer on practice management software, book a demo."

Expected output: Message 1, Urgent clinical, "Possible allergic reaction to new tablets, lip swelling". Message 2, Prescription and sick-note requests, "Repeat blood pressure tablets wanted before Friday". Message 3, Marketing and vendors, "Software vendor demo offer". If the sorter files message 1 anywhere else, the rule has failed and the tool is not ready.

Where the sorter gets it wrong

Three realistic failures to test for. Short, polite messages that hide urgency ("Just checking whether I should be worried about the lump"). Messages in a second language or with heavy abbreviation. And one message holding two requests, such as a repeat prescription and a new symptom, which a single label cannot capture. For the last, tell the sorter to use the most urgent label that applies.

Keep an eye on drift: once a month, pull ten messages the sorter labelled as admin or marketing and check them yourself. If you find a single clinical one in the wrong pile, tighten the rule.

Job 3: insurer forms and medical reports

Forms are the least pleasant job and the most variable. The same insurer questions come back with different wording, and most of the answer is already in the record. AI helps in two places: mapping the form's questions to the record, and writing the narrative boxes.

Build a form-answer bank

Keep a document of approved wording for the questions that recur: employment status of the patient, dates of first consultation, the standard description of a long-term condition, treatment plans in plain language. Each entry has a date of last review. When a form arrives, the assistant (at tier B or C) matches questions to the bank and flags the ones with no stored answer.

Sample task, labelled illustrative. An income-protection insurer asks: "Describe the patient's current treatment and expected time to return to work." Your note says: physiotherapy twice weekly for six weeks, review at week six, lifting restrictions, expected return in 6 to 8 weeks if progress continues. The assistant's draft:

Patient is attending physiotherapy twice weekly for six weeks, with review at week six. Lifting restrictions apply during this period. If progress continues, a return to work is expected in approximately six to eight weeks. This estimate will be reviewed at the week-six appointment.

What you would check: that "approximately six to eight weeks" is your estimate and not a promise, that the restrictions match the sick note or work certificate you issued, and that nothing implies a diagnosis you have not made. Forms are certifications; you are attesting every sentence.

What stays blank for you

Anything where you must give an opinion on causation, prognosis, capacity or fitness for work is your judgement. Let the assistant draft the facts and leave those boxes empty for you to fill in. Likewise, never let a tool tick consent boxes or sign on your behalf.

Walk through: a three-form afternoon

Say Friday afternoon holds three forms: a life-insurance attending-physician report, a fit-for-travel certificate and a disability-benefit questionnaire. Without help, each takes 15 to 25 minutes. With the bank, you paste the de-identified question list into the tool, get answers matched to bank wording, and fill the identifiable parts in your own system. The first form takes ten minutes as you build the bank; the second and third take five to eight. You have saved perhaps half an hour and, more usefully, got consistent wording across forms for the same patient.

A Monday inbox of sixty messages

To see the sorter in motion, say the weekend has left a solo practice with 60 unread messages. Illustratively, the sort might land like this: 4 urgent clinical, 11 results and letters in, 14 prescription and sick-note requests, 18 admin and appointments, and 13 marketing. Without a sorter you read all 60 in the order they arrived, and the possible allergic reaction is message 43. With one, the four urgent items sit at the top with one-line summaries, you spend ten minutes on them, hand the 14 prescription requests and 18 admin items to staff, and skim the 11 inbound results in a batch.

The saving is not mainly time. It is that the order of reading is now decided by risk rather than by arrival time. Check the result by counting: if the labels do not add up to 60, something was dropped, and you want to know that before you trust the tool with a quiet week.

Replies to patients: draft, never send

A fourth use sits beside the three jobs: drafting replies to routine patient messages. Keep it to administrative answers such as opening hours, how to request a repeat prescription, what to bring to a new-patient appointment and how long a form takes. A sample prompt, with placeholders only:

Draft a short, warm reply to a patient asking how long an insurance form takes to complete.
Our policy: forms are completed within 10 working days, there is a fee that the front desk will confirm, and the patient should book a short appointment if the form needs an examination.
Do not give medical advice. Sign off as "[practice name] team". Address the patient as "[first name]".

The tool should return four or five plain sentences. Staff review them, fill in the name and send from your usual mail system. Anything that asks about symptoms, results or medicines is not a routine reply: it goes to the urgent pile for you. If the draft starts offering advice on a symptom, delete it and tighten the prompt.

Three numbers to track in the first month of drafting

Track three numbers for the first month, on paper if you like: minutes per referral letter, number of edits you made per draft (a rough tally is enough), and the number of messages the sorter put in the wrong pile when you spot-checked. If minutes per letter have not fallen after two weeks, your notes are too sparse or your template too loose. If edits stay high, the template, not the tool, is usually at fault. If the sorter misses even one urgent message in a spot-check, pause it and fix the rule before continuing.

Reading the draft: a one-minute review step

Every AI draft passes through the same short routine before it leaves. Make it physical: print the checklist, or put it as the first line of every template, so it is not skipped on a busy day.

• Who: the right patient and the right recipient.
• What: every medicine, dose, date and number against the record.
• Nothing extra: any sentence you cannot trace to your notes comes out.
• Tone: your degree of certainty, not a more confident or softer one.
• Sign: you sign, not the tool.

A realistic mistake and how it showed up: a draft referral for a patient with two similar medicines listed both at the dose of the first. Reading the medicines line against the prescribing screen caught it. The slip came from the tool tidying two list items into a pattern. It is the reason the medicines line gets its own look.

Consent, retention and being honest with patients

If a tool handles identifiable information, your privacy notice should say so in plain words. Say what the tool does (drafts letters you review), what it does not do (make clinical decisions), and who the vendor is. Check how long the vendor keeps inputs and whether you can delete them. A patient who asks "does AI write my letters?" should get an honest answer: "It drafts the wording, and I read and sign every one."

Keep a one-page register: tool name, plan, what it sees, the date you obtained the agreement, who may use it and the date of your last review. When a vendor changes its terms, as vendors do, the register tells you what to re-check. For a wider view of which tasks should never be handed to a tool in a health setting, see which pharmacy tasks should never be handed to AI, much of which carries over.

A four-week rollout for one person

1. Week 1: templates at tier A. Build three referral templates and the form-answer bank. No patient data involved.
2. Week 2: tool and agreement. Ask for the agreement, read the retention terms, fill in the register. If no agreement is available, stop at tier A or pick another tool.
3. Week 3: letters live. Draft every referral with the tool, using the five-point check. Time each one.
4. Week 4: add the sorter or forms. One at a time. Test the sorter on de-identified samples before pointing it at a real mailbox.

At the end of the month, compare the time per letter and the number of corrections you made. If corrections are frequent, tighten the template or the notes you give it, before adding more jobs.

If you want help choosing which job to start with and which of your existing tools can do it, that is what my AI implementation consultation covers. The rest works on your own.

Questions before you hand admin to an AI secretary

Is a virtual secretary the same as an AI secretary?

Not necessarily. A virtual secretary is often a human assistant or an answering service sold as a package. An AI secretary here means software drafting letters, sorting messages and filling forms for you to approve. The privacy questions differ, so ask each vendor who or what sees patient information.

Can I paste a patient's details into a free chatbot?

Treat that as a no. Consumer chat plans come with no agreement covering patient data, and on ChatGPT Free, Plus and Pro your chats can be used to train models unless you turn off the Improve the model for everyone setting. Use a tool where you have a signed agreement for patient information, or work from de-identified text only.

Do I need a signed agreement with the AI vendor?

If the tool will see identifiable patient information, you generally need a contract covering it: a business associate agreement where that is the local contract form, or a data processing agreement under the GDPR. Ask the vendor for it in writing and check that it covers the exact product and plan you use.

Will insurers accept AI-drafted forms?

Insurers care that the content is accurate and signed by you, not how the first draft was produced. Check the insurer's own rules on automated or assisted completion. You remain the person certifying the form, so read every field.

What should I do first if I only have an hour?

Start with referral letters. They follow a pattern, take five to ten minutes each by hand, and the draft sits in front of you before it goes anywhere, so mistakes are easy to catch. Leave inbox sorting until the letter workflow runs smoothly.

Further reads

• AI Chatbot for a Doctor's Website: Scope, Disclosure, Handover — A patient-facing job with its own scope and handover rules.
• AI Answering Service for a Doctor's Office: After-Hours Call Rules — Rules for calls that arrive when the surgery is closed.
• AI Dictation Software for Doctors: Dictation or Ambient Notes? — Choose between dictation and ambient capture for notes.
• AI Receptionist for Physical Therapy: Referral and Insurance Questions — How another clinic handles referral and insurance questions.
• AI Marketing for Medical Practices Without Making Health Claims — Marketing for a practice without making health claims.
• AI Claim Scrubber: What It Catches and What It Misses — Modifier, bundling and diagnosis edits are caught; eligibility, documentation and fresh payer rules often are not. Test one on 50 claims.
• AI Tools for Dietitians: Admin, Content and Follow-Up by Job — One AI tool per job for a private-practice dietitian: admin and notes, content, and between-session follow-up, with costs, privacy checks and what to skip.
• AI for Home Health Agencies: Faster Replies to Hospital Referrals — A referral intake checklist, ready-to-send missing-information requests and a response-time log, with AI under a signed business associate agreement.
• AI in Healthcare for Physicians: Personal Tool or Practice-Wide Tool? — A practice-owner decision guide for individual clinician AI use, staff workflows and the agreement boundary.
• How Can AI Help Physicians Finish Charting? A Four-Week Trial — A four-week AI scribe trial for a physician group: baseline minutes per note, two clinicians, a consent script, an error log and a go or no-go rule.
• How Can AI Help Physiotherapy Clinics Write Referrer Updates? — Turn a clinician's own bullet points into a clear referrer update, with templates for first, progress and discharge letters and a review step.
• Is ChatGPT HIPAA Compliant? Free, Business and Clinician Plans — Which ChatGPT plans can sign a BAA, which cannot, and the do-not-paste rules to give your front desk and clinicians.
• Is Claude HIPAA Compliant? Which Plans Offer a BAA — A plain plan-by-plan table of where Claude can sign a BAA, who has to click accept, and what stays outside the agreement.
• Is Copilot HIPAA Compliant? Microsoft 365 and Web Search Limits — Copilot inside a properly set-up Microsoft 365 tenant can sit under the BAA, but web search does not. Here is where the line falls.
• Prior Authorisation With AI: Drafting Letters and Tracking Status — A payer checklist, a redaction habit, a letter prompt and a status tracker speed up prior authorisation without handing patient data to a chatbot.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI business pricing page; OpenAI help-centre articles on business associate agreements, ChatGPT for Clinicians and data controls (all fetched 4 November 2026).