AI Solutions for Medical Practices: Build, Buy or Hire Help?

A medical practice should buy ready-made software for anything that touches patient records, build its own AI workflows only for tasks with no patient data, and hire an implementation consultant when it has more than two candidate projects, no one to own them, or a large spend at stake. The rule is data sensitivity first, size second.

That rule needs unpacking because "AI" in a practice covers jobs with very different risks. Drafting a newsletter and writing consultation notes both use language models, but one involves no patient information and the other involves the most sensitive information you hold. Treating them as one decision is how practices end up with expensive tools for easy jobs and improvised ones for hard jobs.

Follow me on Instagram@sagnikteaches

Sort every job by what data it touches, before you pick a route

List the jobs you would like help with. For each, decide which of four data classes applies. This takes thirty minutes with a practice manager and a clinician and saves weeks later.

Connect on LinkedInSagnik Bhattacharya

Notice that "build" only appears at the top. That is deliberate. A homemade automation that sends a recall list through a general chatbot may work well and still breach your own confidentiality duties. For the mechanics, what to ask before buying any AI tool for a medical practice goes through the buying questions in detail.

Subscribe on YouTube@sagnikteaches

The three routes compared on cost, time and who carries the risk

The figures below are illustrative ranges for a small practice and are not vendor quotes. Replace them with real quotes before you decide. The structure is the point: what you pay, what you spend in time, and who is responsible if it goes wrong.

What buying actually costs: scribe prices as a worked example

AI scribes are the commonest clinical purchase, so they make a good benchmark. These are USD list prices from the vendors' own pages in November 2026; check before you buy.

• Heidi: a free plan with unlimited ambient documentation; the Clinician plan lists at $110 per user a month billed annually.
• Jane: AI Scribe Free gives five notes a month per practitioner, and unlimited notes are $15 a month per opted-in practitioner, on top of a Jane plan.
• Freed: $39 a month on Starter for up to 40 notes, $79 on Core for unlimited notes, as displayed (confirm whether that is monthly or annual billing), with a 7-day free trial.
• Carepatron: a free plan that includes an AI scribe and note taker; Plus is listed at $31 per user a month.

Illustration: a four-doctor practice with each doctor seeing about 20 patients a day. If all four use Jane's scribe at $15, the add-on is $60 a month. On Heidi's Clinician plan at $110 each it is $440 a month, or $5,280 a year. If each doctor saves 40 minutes a day (a number to measure in your own trial, not assume), that is about 13 hours a month each, 52 hours across the practice. At any realistic clinician hourly value the licence is small next to the saving; what is at stake is data handling, integration and whether the notes are good enough to sign. That is a buy decision, not a build decision, and the shortlist depends on your records system. The tutorial on choosing a scribe by record system covers that part.

When building is the right answer, and how to do it safely

Building means using a general assistant, spreadsheets and simple automations to do a job yourselves. It is the cheapest route and often the best for non-clinical text work. Illustrative examples where it fits:

1. Recall letter template. Draft a friendly letter with placeholders, "[first name]", "[appointment type]", and have staff merge details inside your booking system. The assistant never sees patient data.
2. Patient information sheets. Draft plain-language sheets on common procedures from your own clinical wording, then a clinician reviews and signs off each one.
3. Policy documents. First drafts of infection-control, complaints and data-protection policies that a partner then edits against your regulator's requirements.
4. Staff rota summaries. A weekly summary built from shift initials and dates, no names of patients involved.
5. Job adverts and onboarding checklists. Pure text work.

A safe build has four rules. One person owns each workflow, by name. The prompt is saved in a shared document, not in someone's chat history. No patient identifiers go in, ever, unless the tool has a signed agreement and your policy allows it. And each output is reviewed by a person before it leaves the practice. Here is a prompt template that follows those rules.

You are drafting a recall letter for a medical practice. Use a warm, plain tone, under 120 words.
Placeholders to keep exactly as written: [first name], [appointment type], [practice phone number].
Do not mention any condition, test result or medication.
End with how to book and how to opt out of reminders.
Output the letter only.

When buying is right, and the checks that decide which product

Buy when the job involves clinical content, when a mature product exists, and when you can verify its data handling. The purchase checks fall into five groups, which the practice manager and the lead clinician should each sign off.

• Data. A signed agreement, stated storage location, retention and deletion settings, and a clear statement on whether your data trains models.
• Fit. Works with your records system, ideally writing back into it, and with the consultation types you run.
• Safety. The vendor states whether the product counts as a regulated device where you practise, and there is a clinician-review step before anything is filed.
• Cost. Price per clinician, caps, annual lock-in, price rises, and what happens at the cap.
• Exit. You can export notes and delete your data when you leave.

For front-desk tools, the picture is similar but the checks shift towards call handling, hand-off to a human and booking accuracy. What each type of AI receptionist does sets out the types, and business agreement checks by plan covers the paperwork side for practices that need an agreement of that kind.

Which route fits a solo, small, mid-size or multi-site practice

When a consultation pays off, with a worked example

A consultation earns its fee when it prevents a wrong or premature purchase, finds a job you did not know was worth automating, or gets you from "we should" to a working plan. Illustration: a six-clinician practice has four ideas on the whiteboard: a scribe, an AI receptionist, a recall campaign and a billing checker. The practice manager has time for one project a quarter.

A structured conversation sorts them. The recall campaign sits in data class 2 and can be done inside the booking system the practice already pays for, so it costs almost nothing. The billing checker needs coding accuracy checks and an audit trail, which is a larger piece of work. The receptionist raises hand-off and call recording questions. The scribe has the biggest clinical time saving and the clearest buying process, so it goes first, with a two-week trial of two products on the same twelve consultations, and a pass mark set in advance. The output is a one-page plan: order of projects, owners, what to ask each vendor, what to measure. If the practice had bought the receptionist first, it could have spent three months on call-flow arguments before touching the biggest time sink.

A consultation does not do the clinical safety work for you, and it cannot replace the vendor's evidence or your regulator's rules. For a practice-specific view of when outside help is worth it, see whether small medical practices need an AI consultant. If you want to talk it through, that is what my AI implementation consultation covers; use the contact page.

The hidden cost of building: maintenance hours

A built workflow has no invoice, so its cost hides in staff time. Illustration: a practice manager spends four hours building a set of reminder and recall templates, then about 30 minutes a month fixing wording, updating opening hours and checking that outputs still follow the rules. Across a year that is roughly ten hours, which is cheap. Now imagine the same person building a multi-step automation that reads an inbox, classifies messages and drafts replies. Setup takes 20 hours, and each time the email provider or the assistant changes behaviour, someone spends two hours finding out why. If that person leaves, the workflow becomes a mystery. Before you build, ask three questions: who maintains it, where is it documented, and what happens when that person is away for a fortnight? If any answer is "nobody", buy instead.

A build example walked through: the weekly rota summary

A fifteen-person practice sends a weekly staffing summary every Friday. The manager used to write it by hand in about 40 minutes. The data class is 1 because the input contains only initials, shift times and room numbers. The build takes three steps. First, a shared spreadsheet holds the rota with a fixed layout. Second, the manager pastes the week's rows into the assistant with a saved instruction. Third, she reads the result and sends it.

Turn this rota into a one-page weekly summary for staff.
Sections: Who is in each day (initials only), Rooms in use, Gaps to fill, Changes since last week.
Use only the rows pasted below. If a day has no clinician listed, write "GAP: no clinician listed". Do not guess names.
Rota rows:
[paste]

Illustrative output: "Monday: AB, CD (rooms 1 and 2). Tuesday: AB (room 1). GAP: no nurse listed after 14:00. Wednesday: CD, EF (rooms 2 and 3)." What she fixes: the assistant has correctly flagged the Tuesday gap, but wrote "no nurse" where the rota only shows an empty slot, so she reads it against the rota and corrects the wording. The job now takes about 10 minutes, saving roughly 30 minutes a week, or 26 hours a year. No patient data is involved and no vendor contract is needed. This is what a good build looks like: narrow, text-only, reviewed and documented.

A scoring sheet to decide the route for each job

If a job does not fall clearly into one route, score it. Rate each line 0 to 2 and add up. Illustrative sheet for the job "draft replies to patient emails".

Reading it: a low score on the first and third lines rules out building, a high score on the second points to buying, and the booking-system line says check what you already pay for before buying anything new. The total of 4 out of 10 here points to "buy, after checking the existing tool". A job scoring 8 or more usually suits a build; a job with a zero on the first line should never be a build.

How to brief a consultant so the fee buys a decision

If you do decide to hire help, scope it. A short written brief gets a better quote and a better outcome than a general conversation. Illustrative brief:

Practice: [number] clinicians, [number] admin staff, [records system].
Goal: choose and pilot one AI tool in the next quarter.
Candidate jobs: scribe, call handling, recall messages, billing checks.
Constraints: no patient data to leave [region] without a signed agreement; clinician review of anything clinical; no more than [X] staff hours a week.
Deliverables wanted: ranked list of jobs with reasons; two vendors per top job with questions to ask; a pilot plan with pass marks; a handover document so we run it ourselves.
Not wanted: building software, long-term support contracts.

Check the quote against that list. A good proposal names the deliverables, the number of calls or sessions, the price, and what you will own at the end. A weak one promises "transformation" without saying what you will hold at the end.

Questions to put to a vendor, a builder and a consultant

Each route needs a different question set, and the answers differ in kind.

• To a vendor: Where is our data stored, and is it used to train models? Will you sign an agreement before the trial? How do I export my data and delete the account? What does it cost at twice our volume?
• To the colleague who wants to build: What data will it touch? Where will the prompt and workflow be documented? Who maintains it? What is the manual fallback if it fails?
• To a consultant: What will you deliver and what will I own afterwards? Do you take commissions or referral fees from vendors? How do you handle a recommendation that is "do nothing yet"? Can I speak to the person who will actually do the work?

The consultant's answer to "do nothing yet" is the most revealing. Someone paid only to sell implementation has an incentive to find a project. Someone paid for a decision does not.

Three ways the wrong route fails

• The homemade scribe. A technically keen partner builds a pipeline that records consultations and sends audio to a general transcription service. It works in a demo, but there is no data agreement, no retention rule and no vendor to call when it breaks. Fix: buy a product designed for clinical use and spend the partner's time on the trial.
• Buying everything at once. A practice signs for a scribe, a chatbot and a reminder tool in one month. Staff are overwhelmed, nobody can tell which tool helped, and three data agreements need review. Fix: one tool per quarter, each with a named owner and a measure.
• Hiring help with no owner. A consultant delivers a plan and leaves, and nobody in the practice is responsible for carrying it out. Fix: name an internal owner before the engagement starts and agree a handover.

A first month that works for any route

1. Week 1. List every job you want AI to help with and classify each by data class. Stop and decide which two matter most.
2. Week 2. For the top job, check whether software you already pay for has the feature. Use it if it does; see the tutorial on a worked implementation plan for a small practice for the sequence.
3. Week 3. Shortlist two options. Ask each the data questions in writing. Drop any that will not answer.
4. Week 4. Run a trial with pass marks and a named owner. Decide to adopt, extend or stop, and record why.

If at the end of week 1 you have more than two good candidates, or the largest one touches clinical content and you have nobody to own it, that is the signal to bring in help. Otherwise, you can carry on with buying one tool and building the rest.

Questions practice owners ask about the three routes

Is building our own AI workflow ever sensible for a practice?

Yes, for tasks with no patient data, such as drafting website copy, recall-letter templates with placeholders, policy documents or staff rota summaries. Once patient information is involved, buying a tool with a signed data agreement is almost always safer than assembling your own.

What does a consultant do that software vendors do not?

A vendor sells and configures its own product. An independent consultant looks across your jobs, tells you which are worth automating, whether software you already pay for can do them, and what to ask each vendor. The value is the sequence and the decision, not extra software.

How many tools should a small practice buy in year one?

Usually one or two. A scribe or a front-desk tool, not both at once, then review after a quarter. Buying several at once makes it impossible to tell which one helped and multiplies your data agreements, training and monitoring.

What if a vendor says its tool is compliant?

Ask for the evidence: the signed agreement terms, where data is stored, retention settings and whether your data trains models. A badge on a website is not a contract. Your own regulator and insurer set the requirements, so check against them rather than the vendor's wording.

Further reads

• AI Medical Coding Software: Audit Trails and Upcoding Risk — Audit trails and upcoding risk in coding tools.
• Best AI Tool for Medical Billing: Scrubbing, Coding, Denials — Scrubbing, coding and denials: which tool for which job.
• Medical AI Chatbot for Providers: Clinical Search or Patient FAQ? — Clinical search or patient FAQ chatbot?
• What an AI Consultant Can't Do for You, and What You Must Own — What a consultant cannot do and you must own.
• AI Consultant or DIY: Which Suits a Small Professional Firm? — Consultant or DIY for a small professional firm.
• AI Scribe for GPs: Costs, Free Tiers and the NHS Pilot Limit — What GP-focused scribes cost, including free tiers.
• What Does an AI Implementation Consultant Actually Do? — What an implementation consultant actually delivers.
• Do I Need an AI Consultant? 8 Signs It's Time to Get Help — Eight signs it is time to get outside help.
• AI Medical Transcription Services vs Human Transcribers for Small Clinics — Per-minute quotes, turnaround, BAA and an error-check step, plus a five-dictation test that shows which service suits your clinic.
• AI Platform for Doctors: One Suite or Separate Scribe and Phone Tools? — A buying rule for small practices: when one suite wins, when separate scribe and phone tools win, and the integration test to run first.
• Medical Practice AI Budget: What to Fund First and What It Costs — A budget order for a small practice's first AI spend: a decision rule by job, cost bands per tool class from vendor pages, and payback inputs you can measure.
• AI Assistant for a Medical Practice: Build a Staff Policy Q&A — Turn your protocols into a staff Q&A assistant in an afternoon, with a 20-question test set and no patient data anywhere near it.
• Medical Laboratory SOP Template: Draft and Review SOPs With AI — A medical laboratory SOP template plus an AI workflow: draft from source documents, trace every step, review with the director and keep a version log.
• Medical Practice Credentialing: Track Payer Renewals With AI — Build a payer credentialing and renewal tracker with AI: columns, expiry alerts, a document checklist and questions to ask about delegated credentialing.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Heidi, Jane, Freed and Carepatron pricing pages (checked November 2026); all other figures in the examples are illustrative.