Can AI Write After-Visit Summaries? A Safe Method for Small Practices

A small practice can use AI for after-visit summaries safely by drafting from a fixed template, using either de-identified details or a tool covered by a signed business associate agreement (BAA), having a clinician read and approve every summary before it goes out, and checking the reading level of the final text.

Most of the risk sits in two places: what you paste in, and what the model adds that nobody told it. The method below closes both. The clinician supplies the facts, the template controls the shape, and the AI only turns clinical shorthand into plain sentences.

Follow me on Instagram@sagnikteaches

What an after-visit summary has to do, and what AI must never do

An after-visit summary is the page the patient takes home. It says what was found, what was decided, what to do next, and when to come back or call. Patients forget much of what they hear in the room, so this page often does more work than the note the clinician writes for the record.

Connect on LinkedInSagnik Bhattacharya

AI is good at one part of that job: turning "BP 148/92, start amlodipine 5 mg od, recheck 4/52, home readings twice daily" into sentences a patient can follow. It is poor at another part: deciding what the summary should contain. If you let it decide, it will fill gaps with plausible advice, such as a warning sign nobody mentioned or a follow-up interval nobody set. Those additions are the dangerous ones, because they read well.

Subscribe on YouTube@sagnikteaches

So the working rule for the whole method is that the AI may rephrase, never add. Every fact in the summary has to be traceable to the clinician's note. Anything the model volunteers is deleted at review.

Three routes to drafting, and when each one fits

You can reach the same output by three routes. They differ in how much patient data leaves your hands and how much setup they need.

Start with the first route unless you already have a BAA in place. It needs no contract with a vendor, it forces the habit of stripping identifiers, and it is the fallback when someone forgets which account they are logged into.

On the second route, check the actual wording of the agreement before trusting it. OpenAI's help centre describes a BAA for its API, accepted in the API platform by an organisation admin, with self-serve enrolment depending on an established history of API use. For ChatGPT itself, OpenAI arranges a BAA through its sales team for sales-managed Enterprise and Edu accounts, offers none for ChatGPT Business, and gives eligible verified clinicians a separate in-product BAA flow in ChatGPT for Clinicians, a free individual workspace whose eligibility rules are worth checking for your country and profession. The same help page says that accepting a BAA does not, by itself, make your use compliant. Anthropic offers a BAA for its first-party API and Enterprise plans, and states that an Enterprise owner must switch on the health-privacy setting and accept the agreement before coverage applies. Anthropic's article also lists features that sit outside the agreement, such as connectors to third-party services, and says some of them are off by default once the setting is on. In both cases the agreement covers one organisation and specific services, so a staff member using a personal consumer account is not covered by your contract. Check the current pages before you rely on either.

What de-identified really means in a clinical note

Removing the patient's name is the easy part. The identifiers that slip through are the ones inside free text. Before pasting anything into a general chatbot, strip or generalise these:

• Names of the patient, relatives, employers, schools and other clinicians.
• Dates of birth, exact visit dates and exact dates of surgery or injury (use "three weeks ago").
• Phone numbers, addresses, email addresses, record or account numbers, insurance numbers.
• Rare details that point to one person: an unusual occupation, a named sports team, a distinctive injury story.
• Device serial numbers and photographs.

Here is a before and after, illustrative only.

BEFORE (do not paste):
Mrs [surname], DOB 14/03/1961, seen 3 Nov. Tennis club captain. Fell on
steps at work, left wrist. X-ray no fracture. Sprain. Wrist splint 2/52.
Review if no better. Daughter [name] to drive her home.

AFTER (safe to paste):
Adult patient, 60s. Fell at work, left wrist pain. X-ray: no fracture.
Diagnosis: sprain. Wrist splint for 2 weeks. Review if not improving.
Family member driving patient home.

The second version still contains everything the summary needs. The patient's name goes on the page afterwards, in your own system, never in the prompt. A good habit is to keep a short "never paste" card by each workstation with the list above on it.

Building the template the AI has to fill

A template is the single most useful safeguard, because it limits what the model can invent. Every summary uses the same headings, in the same order, with fixed phrases for the parts that must not vary. A workable layout for most small practices has seven parts:

1. Why you came in, one sentence in the patient's own terms.
2. What we found, two to four short sentences.
3. What we decided, including any medicine name, strength and instructions exactly as the clinician wrote them.
4. What to do at home, as a short list.
5. When to get help sooner, drawn from your practice's approved safety-net wording, not invented by the model.
6. Your next appointment or test, with the interval the clinician set.
7. How to reach us, fixed text added by your system, never generated.

Parts 5 and 7 should be pre-written paragraphs per condition type, approved once by the clinicians, and pasted in by staff. The model drafts parts 1 to 4 and 6 only. This split is what keeps the dangerous content, the "go to emergency care if" wording, under human authorship.

A prompt that rephrases and refuses to add

The prompt below is for the de-identified route. Save it as a pinned message, a Project or whatever your chosen tool offers for reusable instructions.

You turn a clinician's shorthand notes into a plain-language summary for a
patient. Use only the facts in the notes. Do not add advice, warning signs,
doses, timings or reassurance that are not in the notes. If something needed
for a section is missing, write [CLINICIAN TO COMPLETE] instead of guessing.

Write at about a grade 6 to 8 reading level. Short sentences. Explain any
medical term in brackets the first time. Use "you". Keep medicine names,
strengths and intervals exactly as written.

Use these headings: Why you came in / What we found / What we decided /
What to do at home / Your next appointment.
Do not write a "when to get help" section; staff will add it.

NOTES:
[paste de-identified notes]

The line about [CLINICIAN TO COMPLETE] matters. Without it, a model asked for a follow-up interval it was not given will pick one. With it, you get a visible hole that the reviewer cannot miss.

Illustrative output from the wrist-sprain notes above:

Why you came in
You hurt your left wrist in a fall at work.

What we found
Your X-ray did not show a broken bone. The injury is a sprain (a stretched
or slightly torn ligament, the tough band that holds a joint together).

What we decided
You will wear a wrist splint for 2 weeks.

What to do at home
[CLINICIAN TO COMPLETE]

Your next appointment
Come back if the wrist is not getting better.

What you would fix: the home-care section is correctly flagged as empty, so the clinician adds "rest it, raise it when sitting, take the splint off for washing only if told". "Come back if not getting better" is too vague, so the clinician adds a number of days. Neither fix needed the model to guess, which is exactly the point.

The main example: a three-clinician physiotherapy clinic

Say a physiotherapy clinic has three clinicians who each see 14 patients a day, four days a week. Today every clinician writes a hand-typed summary for roughly half of them, about 7 summaries a day each, and it takes around 6 minutes per summary. That is 7 x 6 = 42 minutes per clinician per day, or 126 minutes across the team, much of it done after the last patient.

The clinic moves to the de-identified route. Each clinician types or dictates three lines of bullets into their notes anyway. The new flow looks like this:

1. The clinician strips identifiers from the bullet points while writing them, using the "never paste" card.
2. They paste the bullets and the saved prompt into the chatbot and get a draft in about 15 seconds.
3. They compare the draft with their note, fix gaps, and paste in the approved "when to get help" paragraph for that condition.
4. They copy the result into the practice system, add the patient's name, and send.

Review and copying take about 2 minutes. The saving is then 7 x (6 - 2) = 28 minutes per clinician per day, or 84 minutes across the team, at the cost of a standard chatbot subscription. These figures are illustrative; time your own team for a week before and after, because dictation habits change the numbers a lot. If the saving turns out under 15 minutes a day per clinician, the practice should stay with its existing templates and not add another tool.

The same clinic can reuse the output for exercise sheets. The method for that is in creating home exercise programmes with AI for physio patients.

Three more summaries in different settings

The template holds across settings; the content changes. These short walk-throughs show what the model receives and what the reviewer should look for.

A dental hygiene visit

Input bullets: "Scale and polish. Bleeding on probing, lower molars. Plaque score high. Advised interdental brushes size 0.6. Recall 3 months." The summary should say gums bled when probed, explain that this is a sign of inflammation, name the brush size, and give the three-month recall. The reviewer watches for the model inventing a diagnosis such as gum disease that the dentist never recorded. For treatment-cost conversations, see explaining dental treatment plans to patients with AI.

A blood pressure review in a general practice

Input bullets: "Home readings average 148/90. Start amlodipine 5 mg once daily. Check readings twice daily for one week before review in 4 weeks. Bloods booked." The reviewer checks that the strength and frequency are exactly as written, that "review in 4 weeks" is not turned into "next month", and that the model has not added side-effect lists, which belong to the pharmacist's leaflet.

A child's fever visit

Input bullets: "Viral illness. Fluids, rest. Paracetamol as per weight, dose given to parent verbally. Not eating much but drinking. Safety net given." The summary must not state a dose, because the dose is not in the notes in written form. The reviewer should flag the model if it writes one, and the clinician should either add the written dose or leave "your clinician told you the dose; ask us if you are unsure". Safety-net text comes from the approved paragraph for children, pasted in by staff.

The clinician review: eight checks before anything is sent

The review is the safety step, so make it short enough that people actually do it. A clinician with the note on one side and the draft on the other can run these in about two minutes:

1. Right patient: name, age band and visit date match.
2. Diagnosis or finding matches the note, with nothing added.
3. Every medicine name, strength, route and frequency matches character for character.
4. Every number (days, weeks, weights, readings) matches.
5. No [CLINICIAN TO COMPLETE] markers remain.
6. The "get help sooner" paragraph is the approved one for this condition.
7. Nothing reads as a promise ("this will clear up in a week") that the clinician did not make.
8. Tone is calm and plain; no jargon without a gloss.

Put this list in the template file, as a comment above the draft, so it travels with every summary. A reviewer who never sees the list drifts to skimming.

Checking the reading level without guesswork

Run the finished text, not the draft, through a readability check. Microsoft Word reports a Flesch-Kincaid grade level: in Word for Microsoft 365 it sits under Home > Editor > Document stats, and in older versions you tick "Show readability statistics" in the proofing options, then run a spelling and grammar check. Many free online readability checkers give a similar grade score. Many patient-education guidelines suggest aiming around grade 6 to 8, and lower is better for stressful news.

Three edits move the score fastest: split sentences over 20 words, swap Latin or Greek terms for common words ("kidney" for "renal"), and cut every sentence that explains what the patient already knows. Then do the check a machine cannot: hand the summary to a colleague who did not see the patient and ask them to say back what they would do at home, and when they would call. If they can't, the summary has failed, whatever its score.

Here is a before and after for one sentence, illustrative. Before: "Your radiograph demonstrated no osseous abnormality, consistent with a soft tissue injury." After: "Your X-ray did not show a broken bone. The pain is coming from the soft parts around the joint." The second is longer, but each sentence is short and every word is common.

Where this method goes wrong in practice

Four failures turn up repeatedly in setups like this, and each has a cheap fix.

• Identifiers creep back in. A clinician pastes the note from the record "just this once". Fix: separate the bullets from the note, so what gets pasted is always a fresh three-line summary typed for the purpose, and review a random sample for identifiers monthly.
• The personal account problem. A staff member uses a free account on a home phone. Fix: say in writing which accounts are allowed, and switch off training use on any consumer account (ChatGPT calls the setting "Improve the model for everyone"; Claude calls it "Model Improvement"), though nothing replaces the rule of no identifiers.
• Review fatigue. After a month of good drafts, reviewers skim. Fix: seed a deliberate error in one draft a fortnight, such as a wrong dose in a test case, and see who catches it. Keep it friendly; the aim is a habit, not a score.
• Copy-paste of the wrong summary. Two drafts sit side by side and the wrong one is sent. Fix: the name is added last, only after review, in the practice system, so the person who adds it is looking at the right record.

Patients, consent and the paperwork around it

If a summary was drafted with help from AI, say so in your privacy notice and at the desk in plain words, and tell patients that a clinician checks every one. Wording options are in telling patients you use AI note-taking. For the wider checklist of confidentiality controls, including what to log and who can access what, use patient data and AI: a confidentiality checklist for small practices. Whether you need written consent, and what your regulator expects of AI-assisted records, depends on the health-privacy law that applies to your practice; ask your compliance adviser or professional body before you go live, not after.

Fixing a summary after it has gone out

Sooner or later a patient will phone to say the summary does not match what they were told. Decide the process before that call. The clinician pulls up the original note, corrects the summary by hand, and sends the corrected version with a short line saying what changed and why. Nobody retypes the patient's details into the chatbot to "fix" it. Log the incident in the same anonymous way as the audit log, noting which check in the review list should have caught it, and tighten that check.

A second case is an emailed summary sent to the wrong address. That is a privacy incident under your own breach procedure, not an AI problem, but the AI drafting route does not change who is responsible: the practice is, as it always was.

How to tell that it is working

Measure three things for the first eight weeks. First, the edit rate: for a sample of ten summaries a week, how many needed a clinical correction? A falling number means the template and prompt are settling. Second, the time per summary, recorded by the clinicians themselves for one day a fortnight. Third, patient questions: count calls and messages that begin "what did you mean by". A summary that does its job should reduce those.

Keep the log free of patient details: note the type of correction ("dose frequency changed", "follow-up interval missing"), not the person. If corrections cluster around one section, fix the template. If they cluster around one clinician's shorthand, fix the bullets they type in, because the model is only as clear as its input.

When the numbers settle and volume is high, the next step may be a scribe that writes the note itself. Compare that cost in AI scribe costs compared for small clinics before you commit. If you would like to map the route with someone, that is what my AI implementation consultation covers, and you can reach me here.

Questions practices ask before trying this

Can I paste a patient's name into ChatGPT or Claude to get a summary?

Not on a free or standard consumer account. Treat names, dates of birth, contact details and record numbers as off limits unless your plan comes with a signed business associate agreement that covers the product you are using. Safer: write the summary from the clinician's own bullet points with every identifier removed, then add the name back in your own system.

Does a BAA make AI-written summaries safe to send?

A BAA covers how the vendor handles the data. It does not check whether the summary is correct. The clinician still has to read every summary against the visit record before it reaches the patient, and your own team still decides which features, such as connectors or plug-ins, stay switched off.

What reading level should a patient summary aim for?

Many patient-education guidelines suggest roughly a grade 6 to 8 reading level. Run the finished text through the readability statistics in Word or a similar checker, then rewrite any sentence with unexplained jargon. Test it on a colleague who did not see the patient.

How often should we audit the summaries?

Weekly for the first month, then monthly. Pull ten at random, compare each with the clinician's note, and log what had to be corrected. If the same correction appears three times, change the template or the prompt rather than relying on reviewers to keep catching it.

Further reads

• Patient Reviews and AI Replies: Staying Within Confidentiality — Replying to patient reviews without breaching confidentiality.
• How to Roll Out an AI Scribe Without Losing Patient Trust — How to introduce AI to patients without losing their trust.
• AI Tools for Physiotherapy Clinics: What Each One Actually Does — Tools by job for a physiotherapy clinic.
• Do Small Medical Practices Need an AI Consultant? — Whether a small medical practice needs outside help.
• Questions to Ask Before Buying AI That Touches Client Data — Vendor questions for any tool that touches client data.
• Is AI Medical Translation Safe for a Practice? Rules and Workflow — AI can translate reminders and directions with review, but consent, clinical instructions and live consultations need a qualified interpreter or translator.
• 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 for the API, ChatGPT for Clinicians and data controls; Anthropic privacy centre articles on business associate agreements for commercial customers and on model training; Microsoft support page on Word readability statistics; a peer-reviewed 20-year analysis of patient-education readability citing sixth- and eighth-grade targets.