What an AI assistant should be allowed to do in your practice

It is Thursday afternoon, and Dr Papadaki has just rung in sick for Friday. Eleven patients are booked with her. Someone has to find them all, see who else has room, move each one and let them know. On a good day that is an hour at the front desk, between phone calls.

This is exactly the kind of job an AI assistant is good at, and it is why so many practice owners are being told they need one. The harder question is not what it can do. It is what it should be allowed to do, sitting that close to your patient records.

It is good at the dull middle of the job

Most of what fills a front desk is finding, moving, recording and explaining. Find everyone booked with Dr Papadaki on Friday. Move Mr Georgiou to next Tuesday afternoon. Add a note that his referral letter arrived. Remind me how we record a credit.

None of this needs judgement about a patient's care. It needs the right record, the right slot and a dozen clicks you would rather not make. That is where an assistant earns its place. Diagnosis, treatment and anything a clinician signs are a different matter, and a good assistant does not pretend otherwise.

It should act as you, with your permissions

The first thing to ask about any assistant is whose keys it holds. The safe answer is yours. When the receptionist asks it to move an appointment, it should move it as the receptionist, through the same checks the screens use. If the receptionist's role cannot open the revenue report, neither can the assistant.

The risky shape is an assistant with an account of its own, set up once with wide access so it never gets stuck. It is convenient. It also means every question anyone types is answered with more than that person could see. And it makes the record less honest: when something changes, the log should name the person who asked.

It shows you the change before it makes it

Back to Friday. A well-behaved assistant finds the eleven appointments and reads them back to you. Then it proposes the moves: this patient to Monday morning with Dr Alexiou, that one to Tuesday afternoon. Nothing has changed yet. You spot the patient who can only come before work, fix that one, and confirm.

Reading can be instant. Writing should stop and wait. The pause costs you a second, and it is the moment you catch the wrong patient or the wrong week. Some changes deserve that pause every single time, even once you have told it to stop asking about small edits: deleting a record, sending something to a patient, and anything that goes out on your website.

It reads the record instead of guessing

Someone types: add a note to Maria's record that she prefers morning appointments. There are three Marias. A careful assistant searches, finds all three, and asks which one you mean. It does not pick the most recent and hope. Once it has the right record, it names the patient back to you, so you can see it understood.

The same goes for anything missing. Ask it to book a check-up without saying how long, and it should ask, or take the length from your own list of services. A language model is very good at producing something plausible. Near a patient record, plausible is not good enough. Read first, ask when unsure, and never fill a gap with a guess.

It never answers a patient clinically

Now the moment that matters most. A patient opens your website late at night and types into a chat window: are my blood test results normal?

No assistant should be answering that. It cannot see the patient, it cannot tell a worried question from an urgent one, and whatever it says reads as if your practice said it. A reply that sounds kind and is wrong is worse than no reply at all.

So our advice is plain. Use an assistant for your team, behind your sign-in, and keep it off anything a patient types into. Your website can take bookings and answer the questions you have written the answers to yourself. Clinical questions belong to a clinician, on the phone or in the room.

A few questions tell you most of what you need

Whoever supplies the assistant, ask these before you switch it on:

• Does it act as the person who asked, with that person's role, or with an account of its own?
• Does it show each change and wait for a yes before it saves anything?
• Which actions always ask, even when confirmations are switched off?
• Is every change it makes in the log, under the name of the person who asked?
• Where do our words and our records go to produce a reply, and is any of it used to train models?
• Can it ever talk to a patient?

These tell you how the tool behaves. They do not tell you what the law expects of you. Take the answers to whoever advises you on data protection, because they will know your obligations where you practise.

Where HyperCRM fits

We built our own assistant around these rules, so it is fair to say how. It sits in the corner of your dashboard, behind sign-in, and only your team talks to it. It runs as you, with your role, and it is never offered a tool your role cannot use. Changes it makes to patients, appointments and billing are logged under your name. Anything that changes a record stops on a card with the exact values and waits for you to confirm. You can switch that off for ordinary edits, but deletions, bookings, billing, changes to a patient's contact details, and anything that emails or texts a patient or goes public always ask.

It acts only when someone on your team asks. It does not write your reminders, and it does not score or predict anything. What you type, and the records it reads to answer, go to OpenAI to produce the reply, and none of it is used to train models. Each message draws on the same shared credit balance as your SMS reminders. The details are on the assistant page.

This is the assistant inside the app. HyperCRM's separate phone assistant, on the Complete plan and on request, does answer patients' calls: it says it is an AI, books into your free times, gives no medical advice and tells a caller with an emergency to call the emergency services.

If you would rather use an AI client of your own, the Complete plan includes an MCP server with 75 tools. It signs in as you and only ever reaches your own practice.

More on the blog.