A Worked AI Implementation Plan for a Small Medical Practice

A workable AI plan for a small medical practice runs about 12 weeks: two weeks measuring where time goes and setting rules, two weeks of low-risk admin wins, a scribe trial and rollout, and a phone fix that starts with reminders before any AI answers calls. It deliberately leaves clinical advice and triage out.

What follows is one illustrative practice walked through from start to finish, with the numbers, the costs and the mistakes. The practice is invented, but every tool, price and step is real. The shape matters more than the details: measure first, put governance in place before any patient data moves, trial with one clinician before four, and fix the process before automating it.

Follow me on Instagram@sagnikteaches

The practice before the plan

The practice is a private family-medicine practice with three doctors, a practice nurse, a practice manager and three receptionists, handling about 1,100 appointments a month. It already uses a clinical records system, a cloud phone system and Microsoft 365 Business Standard for email and documents. Nobody on the team has set up an AI tool before.

Connect on LinkedInSagnik Bhattacharya

The practice manager spends a week logging where time goes. The problems she finds:

Subscribe on YouTube@codingliquids

The partners agree three goals for 12 weeks: cut after-hours note writing, cut abandoned calls, and speed up letters, without any patient data going anywhere it shouldn't.

Weeks 1 and 2: rules before tools

No tools are bought in the first fortnight. Instead:

• Two named owners. The practice manager leads the plan (four hours a week protected); one doctor is clinical lead for anything that touches patients.
• A one-page AI policy. Approved tools only; no patient information in any tool that hasn't passed the vendor checklist; every AI output about a patient reviewed by a clinician before use; staff personal AI accounts not to be used for practice work. The template in how to write an AI usage policy gets them most of the way.
• A data-protection impact assessment started for the scribe, since it will process consultation audio. The tutorial on whether you need a DPIA before using AI tools explains the steps; the practice's data-protection adviser reviews the draft in week four.
• An email to the practice's medical indemnity provider describing the planned use of an ambient scribe and asking whether any conditions apply. The reply, received in week three, asks that clinicians review and sign every note, which the policy already requires.

Time spent: about 12 hours of the practice manager's time and 3 hours of the clinical lead's.

Weeks 3 and 4: admin wins with no patient data

While vendors are being checked, the team starts using AI where no patient information is involved. Microsoft 365 business plans include Copilot Chat at no extra cost, so there's nothing to buy. The rule for this phase: no patient names, records or messages, only practice documents.

The first job is the patient information on the website and in leaflets, which is written in dense clinical language. A typical prompt and its result:

Prompt: Rewrite this leaflet on preparing for a fasting blood test in
plain English for patients with a reading age of about 11. Keep every
instruction. Use short sentences and a numbered list. Do not add any
medical advice that isn't in the original.

[original leaflet text]

Illustrative output (extract):
1. Do not eat anything for 10 to 12 hours before your test.
2. You can drink water. Do not drink tea, coffee or juice.
3. Take your usual medicines unless your doctor has told you not to.
4. Bring a snack to eat after your test.

The clinical lead checks every rewritten leaflet against the original. In one, the AI had changed "10 to 12 hours" to "overnight", which is vaguer and, for an early-evening meal, wrong. The fix was the instruction to keep every number exactly as written, added to the prompt for all later leaflets. Over two weeks the team rewrites 14 leaflets and the website's 30 most-asked questions, and drafts a staff rota template and two practice policies. Time saved isn't huge, but the team learns how to prompt and check before any patient data is involved.

The phase had one near miss, and it was instructive. In week four a receptionist, pleased with the leaflet results, pasted a patient's long, upset portal message into Copilot Chat to get help drafting a calm reply. She removed the name first, but the message still carried a date of birth in the signature and a description of a recent diagnosis. She raised it herself at the Friday check-in. The practice manager treated it as a policy question rather than a disciplinary one: the message was deleted from the chat, the event was noted in the policy log, and the rule was made more concrete. "No patient information" became "no text a patient wrote, even with the name removed", because a patient's own words identify them in ways staff don't expect.

Weeks 3 to 6: choosing and trialling a scribe

The practice manager sends the medical practice vendor questions to three scribe vendors in week three and scores the written answers in week four. Two pass. The one with the clearest data-processing agreement and configurable audio retention goes to trial.

Her scoring sheet, filled in from the three vendors' written replies (vendors anonymised here):

Vendor C wasn't failed for a bad answer but for no answer. "Stored securely" doesn't say for how long, and a vendor that won't put retention in writing before you sign is unlikely to be clearer after. Vendor B won on the retention setting and the export route, which matter more over three years than a slightly nicer interface.

Weeks five and six: one doctor, the most enthusiastic, trials it on every consenting patient. Receptionists hand patients a short card at check-in:

Your doctor may use a secure AI note-taking tool today. It listens to
the consultation and drafts notes, which your doctor checks and signs.
The recording is deleted after 7 days and is not used to train AI.
You can say no at any time and it won't affect your care.

Of 214 patients seen in the trial, 9 decline. The doctor's after-hours note time falls from about 35 minutes to about 10 on most days, and referral letters drop from 12 minutes to about 5, because the scribe drafts the letter from the consultation. The trial also finds the first real error type, described below.

The partners decide on rollout from a one-page scorecard the clinical lead fills in at the end of week six:

SCRIBE TRIAL SCORECARD (weeks 5-6, one doctor)
Consultations seen:              214
Patients who declined:           9 (4%)
Drafts signed with no edits:     61 of 205 (30%)
Drafts needing minor edits:      127 (62%)
Drafts needing major rewrite:    17 (8%), mostly phone consults
                                 and consultations in a language
                                 other than the template's
Clinically significant errors:   2 (both sound-alike drug names,
                                 caught at review)
After-hours note time:           35 min -> about 10 min most days
Doctor's verdict:                keep; not for phone consultations
Rollout condition:               medication section added to
                                 every template before week 7

The line that changed the rollout was the major-rewrite row. Seventeen drafts needed rewriting, and most were phone consultations where the audio was poor. Rather than blame the tool, the practice simply left phone consultations out of scope, and the three clinicians who followed started with face-to-face appointments only.

Weeks 7 to 10: scribe rollout, with a review routine

The scribe goes to the other two doctors and the practice nurse in week seven, with templates adjusted for the nurse's chronic-disease reviews. The rollout comes with three rules, taken from what the trial found:

1. Check medication names and doses in every draft against what was said. In the trial, the scribe twice wrote a sound-alike drug name, once hydroxyzine where the doctor had said hydralazine. Both were caught at review. The template now puts medications in their own section so they're checked on their own.
2. Delete any examination finding you didn't actually perform. Drafts occasionally "complete" an examination section.
3. Confirm every piece of advice given to the patient is in the plan section.

Once a month, the clinical lead reviews five signed notes per clinician against the transcript. Each check takes about four minutes and is logged in one line, such as "Nurse, 3 Nov, diabetes review: HbA1c value correct; foot check recorded as done, transcript confirms; advice on hypo symptoms missing from plan, added". Twenty lines a month is enough to see whether one clinician has started signing drafts without reading them, which usually shows up as a run of notes with no edits at all. The practice chose a plan at about $79 per clinician a month (Freed Core's published price is one example; the practice's chosen vendor quoted similarly), so four seats cost about $316 a month. More on scribes and patient trust is in whether an AI scribe is worth it for a small clinic.

Weeks 7 to 12: fixing the phones in the right order

The tempting move was an AI receptionist from day one. The call data said otherwise: about 30% of calls were patients checking or moving appointments. So in weeks seven to nine the practice fixes reminders first, using its existing phone and records systems: a text two days before each appointment with a link to confirm, cancel or rebook. Abandoned Monday calls fall from 18% to 13% before any AI answers a call.

The wording of the text mattered as much as sending it. The old reminder and the new one:

Before: Reminder: you have an appointment on Tue 14 Oct at 10:20.
        Please call the surgery if you need to cancel.

After:  Your appointment is Tue 14 Oct, 10:20, with the practice
        nurse. Confirm, cancel or move it here: [link]
        Please don't reply to this text with health questions;
        call us or use the patient portal.

The old version told every patient who needed a change to ring, which is how the Monday queue formed. The new one gives the change a route that isn't the phone, says who the appointment is with (a common reason for "when is my appointment?" calls), and heads off symptom replies to an unmonitored number. The simple sum the practice ran: of about 2,400 calls a month, roughly 720 were checking or moving appointments. If the link handles even a third of those, that's around 240 fewer calls a month, most of them at the busiest times.

In weeks ten to twelve, an AI phone assistant goes live for overflow only: calls not answered within five rings, the lunch hour and after hours. It takes messages and books routine nurse appointments; it does not triage symptoms. Any mention of chest pain, breathing difficulty, bleeding, or thoughts of self-harm gets fixed emergency wording and an immediate transfer or emergency advice, tested in advance with scripted calls.

The scripted calls earned their keep. The practice manager and a receptionist ran 25 test calls before go-live, including mixed ones where the urgent part is buried. One script went: "Hi, I need to move my Thursday blood test. Also my chest has felt tight since this morning, but it's probably nothing." In the first configuration the assistant moved the blood test, then offered a routine appointment for the chest tightness. That is exactly the failure the rule is meant to stop, and it happened because the booking instruction ran first. After the setup was changed so the emergency check runs on every call before any booking or message-taking, the same script got the emergency wording straight away. The practice kept all 25 scripts and reruns them after any change to the assistant's instructions.

Each handled call produces a summary in the reception inbox:

Illustrative call summary:
12:41  Caller: patient (DOB confirmed)
Reason: repeat prescription query, says pharmacy hasn't received it
Action taken: message logged for prescriptions team
Promised: call back this afternoon
Flags: none

A general AI receptionist on overflow volumes costs roughly $130 to $150 a month on published plans such as Goodcall's Growth or Smith.ai's Pro. The setup steps followed setting up an AI phone agent for after-hours calls.

The practice's 90-day plan on one page

What it cost and what changed by day 90

At day 90, against the baseline week: doctors' after-hours note time is down to about 10 minutes on most days, referral letters take about 5 minutes instead of 12, and Monday abandoned calls are at 9%, with roughly half the improvement coming from reminders rather than AI. Nine scribe drafts in the month had a clinically significant error, all caught at review. The partners keep all three changes and add one: moving the scribe's letter templates into the records system so letters don't have to be copied across.

What the practice deliberately didn't do

• No symptom checker or AI triage. Deciding who needs to be seen urgently stays with clinicians and trained staff.
• No AI reading results or letters from hospitals to decide what action is needed.
• No patient-facing chatbot giving health information in the first 90 days. It may come later, with tight limits, once the practice is used to supervising AI.
• No patient messages pasted into general chat tools, even on the business plan, until a specific use was assessed and approved.

Five lessons from the 12 weeks

1. The cheapest fix came first. Better reminders did half the work on the phones for almost nothing.
2. The trial found the real risks. Sound-alike drug names only showed up with real patients, and the template change that followed was simple.
3. No-patient-data work built skills safely. Two weeks on leaflets taught everyone to write precise prompts and check outputs before the stakes were high.
4. Named owners kept it moving. The four protected hours a week were the difference between a plan and a good intention.
5. Measuring at the start made the review easy. Without the baseline week, "it feels better" would have been the only evidence.

If your practice's list of candidate jobs is longer than this one's, admin tasks a small clinic can hand to AI helps rank them before you decide what goes into weeks three and four.

Questions about adapting this plan

Can a practice run this plan without a practice manager?

It needs one named person with protected time, about three to five hours a week for twelve weeks. In a practice without a manager, that is usually a senior receptionist or a partner. Without a named owner the plan tends to stall at the vendor-checking stage, because nobody has the job of chasing written answers and booking the trial.

Why not start with a patient-facing chatbot?

Because it carries the most clinical risk and the least evidence of benefit for a small practice. A chatbot answering health questions needs careful limits, testing and supervision, and patients may treat its answers as advice. Fixing reminders, notes and phone overflow first gives measurable gains with lower risk, and teaches the practice how to supervise AI before anything talks to patients about their health.

What if the scribe trial fails?

Then the plan has done its job: you've learned cheaply. Check whether the failure was the tool, the templates or the setting, such as a noisy room or consultations with little spoken detail. Try a second vendor only if the first failed for reasons specific to it. Otherwise move the budget to letters or admin, where the gains may be easier.

How does this change for a practice twice the size?

The order stays the same; the pilots get longer and the governance gets more formal. Trial the scribe with two clinicians of different styles rather than one, run the phone pilot on one site or session first, and consider a group plan with admin controls once more than five or six clinicians are using the same tool.

Further reads

• How to Roll Out an AI Scribe Without Losing Patient Trust — The patient-trust side of a scribe rollout in more depth.
• Patient Data and AI: A Confidentiality Checklist for Small Practices — The confidentiality checks to run before week five.
• How to Review an AI Tool After 90 Days: Keep, Fix or Cancel — How to run the day-90 review: keep, fix or cancel.
• How to Roll Out an AI Policy So Staff Actually Follow It — How to roll out an AI policy so staff actually follow it.
• Do Small Medical Practices Need an AI Consultant? — Whether a practice like this needs outside help to run the plan.
• AI Scribe Costs Compared: What Small Clinics Pay in 2026 — Published scribe prices to compare against your quotes.
• What Is an AI Scribe and How Does It Work in a Consultation? — The consultation stage by stage: capture, transcript, speaker separation, note drafting and sign-off, with the errors each stage produces and how to spot them.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft 365 business plan inclusions (Copilot Chat) and scribe and AI receptionist list prices from vendor pages (Freed, Twofold, Goodcall), checked 27 September 2026. The practice, its figures and results are illustrative.