Algeria AgencyArtificial intelligence
AI in a private clinic: the appointment, never the diagnosis
An automated front desk can give out a slot. The moment it asks why you are coming, it is performing a medical triage nobody delegated to it.
Wednesday, between 8 and 10 in the morning, at a radiology practice in Algiers. The phone rings without a break, two people are waiting at the desk, and the receptionist is answering those present and those calling at the same time. A patient is calling for the third time, not for an appointment but to ask whether her results are ready.
That is the window where a practice loses the most time, and it is the one an artificial intelligence system is now offered for. The question is not whether it can hold a conversation — it can. It is what it is allowed to decide, and the answer is narrower than what is being sold.
This article sets the boundary: what an automated front desk may take on its own, what it must never ask, and the three refusals that are not negotiable. It does not repeat what a health practice is allowed to publish, nor who answers when the practice is closed: those two subjects are covered elsewhere and this article assumes they are settled.
The two-hour window, and what crowds into it
Count one morning, changing nothing, and note only the reason for each call. In most of the practices we have watched, four reasons account for almost all of it: making an appointment, moving or cancelling it, asking whether results are available, and asking a question only the practitioner can answer.
Those four reasons share neither a duration nor a value. The first and second are mechanical: find a free box and write a name in it. The third is a status check. The fourth is an act of care disguised as a quick question, and it is the one that spills over the other three.
The imbalance is there and it is structural: the mechanical calls arrive at the same time as the care calls, on the same line, and the person at the desk arbitrates live between a physical queue and a telephone one. No organisation makes that moment comfortable.
The useful conclusion before any purchase is that the possible gain is limited to the first two reasons. If your morning is mostly the fourth, an automated front desk will move your problem rather than reduce it, and it is better to know that before installing one.
An appointment is not a medical question, and the boundary is sharp
There is a precise line between booking a slot and assessing a need for care, and it runs exactly where you start asking why the person is coming. As long as the system asks for nothing but a name, a practitioner, a known type of procedure and a time preference, it is doing diary work.
The moment it asks for the detailed reason, it collects health information, and it uses it to choose the length of the slot, the practitioner or the order of play. That choice is an elementary medical decision, and nobody in the practice formally delegated it to software.
The boundary is not legal to begin with, it is practical: a system that allocates twenty minutes rather than ten because a word sounded serious will one day get it wrong the other way. And the error will not be visible, because too short an appointment looks like a practice running late, never like a fault in a tool.
The setting that holds is therefore to have the type of procedure chosen from a list written by the practice — consultation, review, a named technical procedure — and to let the duration follow from the type. The list belongs to the practitioner, it is short, and it contains no symptoms.
What the system may take on its own: the slot, and nothing else
The safe perimeter is four actions. Offering slots that are genuinely free in the practice diary; taking the person’s name and number; confirming in writing with the date, the time, the practitioner and the address; and recording a cancellation or a change the person asks for themselves.
Those four share one property: they are verifiable on the spot and reversible. A slot taken wrongly is moved, a cancellation is caught up with, a misspelled name is corrected at reception. None of them commits a judgement about anybody’s condition.
Two details make the difference between a usable tool and a source of arguments at the desk. The first is that the written confirmation has to carry the exact address and the named practitioner, because most practices with several practitioners receive patients convinced they have an appointment with somebody else. The second is that cancelling has to be as easy as booking, failing which people do not tell you and the slot is lost twice.
Note what this list does not contain: no answer to a question, no direction to a department, no estimate of waiting time. Those are the places where a system looks helpful and starts speaking on the practice’s behalf.
What it never asks: the detailed reason
The "briefly describe your problem" box is the most widespread defect in appointment forms, and it is indefensible for three cumulative reasons.
The first is that it collects health data in free text, which is the most sensitive category in the least structured field. The second is that this data travels: into the practice’s mailbox, often into a notification on a phone, sometimes to a supplier abroad if the extraction is automated. The third is that it serves nobody — the practitioner will take the history again from the start, as they must.
There is a narrow exception and it deserves naming: some technical procedures require preparation, fasting or a document, and the practice has to be able to say so. But that is handled through the type of procedure chosen from the list, not through a free description: the preparation message is attached to the procedure, not inferred from a sentence.
If you remove only one thing from your current form after this article, remove that field. It is the fastest change, it costs nothing, and it eliminates in one move the most exposed part of your data.
Urgency is not a checkbox
Somebody describing chest pain to a form has not filled in a form, they have asked the first available interlocutor for help. The technical temptation is to recognise those cases and class them as high priority. That is exactly what must not be built.
A system that classifies urgency gets it wrong in both directions, and both errors are serious. Downgrading a real emergency is the case nobody recovers from. Wrongly promoting fills the practice’s day with slots taken for nothing, which ends in the mechanism being ignored.
The correct design is the simplest: at the first wording suggesting distress, the system stops managing a diary. It shows the emergency number and the practice’s direct number, it says clearly not to wait for a written reply, and it alerts a person. It asks no further question, estimates nothing, and records nothing beyond the fact of having handed over.
It is a rule a practice can check in a minute with any supplier, and we advise checking it before any sales demonstration: type an emergency sentence and see whether the tool tries to handle it.
The day-before reminder, and what it really changes
The automatic reminder is the most profitable function in this perimeter, and it is also the one sold with figures that mean nothing. We will publish none: we have not measured a no-show rate before and after across a panel of Algerian practices, and a statistic imported from another health system would describe other habits.
What we can describe is the mechanism. A reminder sent at the end of the previous day gives the person time to cancel in the morning, which makes the slot reallocatable. A reminder sent an hour before arrives too late to help the practice: it reduces forgetting, it does not recover the place.
The reminder’s content obeys the same rule as everything else: the date, the time, the practitioner, the address, and the preparation if the procedure needs one. Never the reason — a message naming the examination appears on a locked screen, potentially in front of somebody else.
The figure you actually want, you can produce yourself in four weeks: count your unattended slots before, add the reminder, count again. That is your rate, on your own patients, and it is worth more than ours.
What the practitioner should see before the patient
The request often comes back in an attractive form: an automatic summary of the person arriving, produced from what they wrote and from their history. That is the point where a diary tool becomes a clinical tool, and where we stop.
A summary produced by a model omits, reorders and smooths. On a medical file those three operations are exactly what loses the information that mattered, and the practitioner reading it does not know what disappeared. It is not a question of model quality but of the nature of the task: summarising is deciding what to leave out.
What is useful and free of risk is a list, not a text: previous appointments with their dates and procedure types, attached documents, what the person supplied themselves. Ordered facts, without interpretation, that the practitioner reads in ten seconds and completes themselves.
The difference between the two fits in one sentence, and it is worth putting to any supplier: does your tool display what exists, or does it manufacture a short version of what exists?
Results, prescriptions, certificates: three refusals
The third call reason in section 1 is the request for results, and it is the function everybody wants to automate first because it repeats. The system can say a document is available and is to be collected, or send it through the channel the practice has chosen and secured. It does not comment on it, does not summarise it, and answers no question about it.
Prescription renewal is the second refusal, and it is sharper still. A renewal is a medical act, it takes account of progress and tolerance, and a system that carries a line forward because it appears in the file takes a decision that law and professional duty reserve to the practitioner.
The third is the certificate, in every form. A certificate commits somebody’s signature to an observed fact, and nothing in an automated diary observes anything.
Those three refusals have something in common with the earlier ones: they are not about what the machine can do, but about what has to stay attached to an identifiable and accountable person. A supplier offering to lift them for a supplement is selling you the risk, not the function.
Health data has a regime of its own
Information relating to health is sensitive data within the meaning of law 18-07 of 10 June 2018, amended and completed by law 25-11 of 24 July 2025. Processing sensitive data is framed more tightly than an ordinary customer file, and installing an automated front desk creates one.
Three practical consequences for a practice. Appointing a data protection officer is assessed on criteria rather than on headcount, and processing sensitive data is one of those criteria: who can hold that role and who cannot is a question to settle before installation, not after.
Next, the processing has to appear in a register, and that register is filled in on a real case in one morning. Finally, if the tool relies on a service hosted outside Algeria, the messages go with it, and the transfer regime and the architectures that avoid it are looked at before the supplier is chosen.
This paragraph describes the state of a regulation at the date of publication and does not replace a lawyer’s advice. It is here because a practice that puts a form online on Monday began processing health data on Monday, whatever its size.
Language, time and name: three places it breaks
An automated front desk writes messages, and those messages contain exactly the three things that display badly. The person’s name, the appointment time, and the date. In Arabic those three are the ones that turn around or reassemble wrongly if the display has not been checked.
The reply language has to follow the language of the request, not the software’s. A patient who writes in Arabic and receives a confirmation in French will not always be able to say whether the stated time is right, and that is one more call to the desk — precisely the one you were avoiding.
The name raises a problem specific to the country: the same person exists under several Latin spellings, and a practice that lets a system write the name as it understood it creates two files for one patient. That is the subject of what voice systems cannot spell, and the answer is the same: match against the existing list, flag rather than correct.
Those three points are checked on your own phone in ten minutes, and they are fixed in the message template. The full list of places where machine-produced Arabic breaks is elsewhere, in an article devoted to display.
The check: thirty calls, three columns
Take thirty consecutive calls, on a normal morning, and note three things: the reason, the approximate duration, and whether the call could have been handled without medical judgement. One morning, one sheet, nobody extra at the desk.
The result reads one way only. If the share of calls handleable without judgement is above half, an automated front desk within the perimeter described here will give you time back. If it is clearly below, your queue is made of care rather than administration, and the tool will only move it.
Two checks complete the sheet and take five minutes each. Type an emergency sentence into the demonstration you are offered and see whether the tool tries to handle it. Then ask for a confirmation in Arabic and read the time and the date on your phone.
One practice in two, at this stage, discovers that its real difficulty is the third reason — results — and that the answer is not a bot but a place where documents are available without going through the phone. That sells less well and it is often the right answer.
What we do, and what we refuse
We connect booking to the existing diary, with the list of procedures written by the practice, the confirmation and the day-before reminder, one-step cancellation, and the emergency handover to a number and a person. We check the messages on a phone, in Arabic and in French, before the line opens.
We refuse the free-text reason field, urgency classification by the machine, the automatic file summary, and any answer bearing on results, a prescription or a certificate. Those refusals are not start-up stages: a practice lifting them later would change the nature of the tool and the responsibility that goes with it.
We announce no avoided no-show rate. We have measured none here, and a figure from another health system would be talking about appointment habits other than your patients’.
What you can do without us is immediate and free: remove the free-text reason field from your current form, and move your reminder to the evening before rather than an hour ahead. Those two take half a day and give a considerable share of the benefit described here.
Frequently asked questions
Can a patient book by simply writing what is wrong?
They will, whatever the interface. The question is what the system does with it: it should accept the message, not use it to sort, and leave it to the person receiving. What is avoided is a field explicitly asking for the reason, because that turns an occasional overflow into systematic collection.
Do we need medical software before automating appointments?
No, but you need one diary that counts. The most common failure is not technical: it is a practice where two diaries coexist, one at the desk and one online, and the system offers slots already taken.
Can the system call back patients who did not attend?
It can show the list to the practice, which is useful. It must not call or write on its own initiative to offer a new appointment: depending on the procedure, that message tells whoever reads it that the person is undergoing treatment, and somebody else may see it.
Does this replace the receptionist?
No, and a practice buying it for that reason will be disappointed. What is absorbed is the mechanical part of the first two call reasons. What remains — welcoming, arbitrating, reassuring, handling the day’s surprises — is the work, and it grows rather than shrinks when the practice fills its diary better.
Our colleagues use appointment platforms. Is it the same thing?
The perimeter is close, the question to ask them is different: where is the data hosted, who else has access, and what becomes of your list if you leave. A platform holding your patients and your diary holds your business.
What if the supplier offers automatic emergency triage?
That is the moment to refuse. No organisational benefit offsets the case where the system downgrades a real emergency, and that case is not hypothetical once the mechanism exists. Handing over to a number and a person costs less and cannot be wrong in that direction.
Where we come in
The morning’s sheet separates two queues: the one a diary settles, and the one where somebody has to listen before answering.
- The list of procedures is written with the practitioner; it fits on ten lines and names no symptom.
- When distress appears the exchange stops dead: a number displayed, an alert raised, not one more question.
- Every message template is opened on a real handset, in both languages, before the line goes live.
Software grading of serious cases, and automatic synthesis of a file, are not among the things we deliver, however firmly asked: the cost of their error lands on a patient.
Read next
Healthcare: what you cannot say, and what is left
Most ordinary marketing advice is prohibited or inadvisable here. What remains is more useful than it looks.Continuity: who answers when the practice is closed
A practice closes at six. The people who need it do not stop at six.The regional audit divisions: what an inspection asks for, in what order
An inspection begins with a list of documents, not with a visit. There are four, they are known, and nothing stops you producing them first.
Let us talk about your project
A free audit, no commitment: we look at your online presence and tell you what is holding it back.