The AI medical administrative assistant: lighten the paperwork, not the care
Letters, files, formalities, coding: administration eats up precious time in a care practice, at the patients' expense. An AI agent can absorb a large part of it — on two absolute conditions: protecting some of the most sensitive health data there is, and never encroaching on the medical act. Hosted in France on dedicated, isolated infrastructure, with the HDS requirement settled with you at scoping, it gives time back to the secretarial team. The professional keeps the decision.
Updated on
Administrative file: the entitlement certificate and the treatment form are there; the proof of address requested at the last visit is missing.
Letter ready to read over, nothing is sent without your approval.
⛓ Sourced · the patient's administrative file + the practice's templates (hosting in France)
The coding is still to be approved by a professional — I am simply flagging a possible supplement that was overlooked on a similar procedure last month.
✎ Action · proposal to approve — the professional decides
A Blue Lemon Agent medical administrative assistant takes on the administrative tasks of a practice or a clinic — letters, patients' administrative files, formalities, help with coding — and answers administrative questions from your internal documentation. Hosted on dedicated, isolated infrastructure in France, it protects highly sensitive data, with the deployment objective of processing and access operated within the European Union and an architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. It performs no medical act, no diagnosis, no clinical triage: everything that touches on care stays with the professionals. Live in two to three weeks, the HDS requirement settled at scoping. Your teams write to it from Microsoft Teams, Slack or their email, and your patients reach it on WhatsApp Business, your website chat or email — with no account to create and nothing to install. These connections are included in every plan, at no extra cost, within the number of connections your level includes.
Reference points describing our offer, not results measured at a client. The scale of the gain is confirmed by a pilot on your own scope.
Why medical secretarial teams are overwhelmed — and why AI worries them
In a care practice, administration eats up precious time: letters, files, formalities, coding. But the data involved is health data, among the most protected there is — and the line with the medical act is never crossed.
! The issue
The secretarial team is caught between an administrative load that keeps growing and patients who expect to be attended to. Yet most consumer AI tools amount to entrusting health data — among the most sensitive under the GDPR, covered by medical confidentiality — to a service hosted outside Europe and subject to the Cloud Act.
✓ Our answer
AI is only of interest to a care practice if it is sovereign and confidential by design, and if it stays strictly administrative. Hosting in France, the HDS requirement settled at scoping, systematic human oversight, a scope locked down with you: time given back to the secretarial team is never paid for in lost confidentiality, and the agent never touches care. The aim is not to replace the care professional, but to give them back time for their patients.
Health data: HDS hosting required & the medical line locked down
Health data is among the most protected there is. Here is how the architecture of our agents secures it — and why the agent never crosses the line into the medical act.
Health data: the HDS requirement
Health data is entrusted to hosting located in France, under French law. In France, hosting health data on behalf of others is subject to HDS certification of the host: that requirement is settled with you at scoping, before any go-live.
Reduced extraterritorial exposure
Deployment objective: processing and access within the European Union; architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
Medical confidentiality respected
Encryption in transit and at rest, role-based access, strong authentication and logging of access.
Administrative, never medical
No diagnosis, no clinical triage, no clinical opinion: a tool that supported medical decisions would count as a medical device, and that is not what this is.
A locked-down scope
The field of action, strictly administrative, is defined with you and locked down: the agent works on the paperwork, not on the care.
AI Act: governed deployment
The agent is strictly in support; no coding is approved automatically; traceability and human oversight from end to end.
What depends on the architecture chosen These points are not general guarantees: they are settled deployment by deployment, in the quotation.
- The applicable location is that of the architecture set out in the quotation and verified before commissioning.
- The applicable isolation depends on the deployment mode set out in the quotation; no dedicated isolation is presumed.
- The encryption mechanisms in transit and at rest, their components and key management are those documented for the architecture chosen.
- Roles and permissions are configured and accepted for the identities and systems actually connected.
- The events logged, their content, their retention period and who may access them are defined for the deployment chosen.
See the agent at work
4 real situations, taken from those that come up most often. Pick one: the exchange unfolds as it would in your organisation.
A scripted demonstration. These exchanges show how the agent behaves — its sources, its refusals, what it leaves to your teams. Nothing is sent from this page, no model is queried here, and the matters named are fictional. That is precisely what we promise your data.
The behaviours shown here — monitoring, automation rules, routing and reminders — are configured with you during deployment, from your tools, your rules and your thresholds.
The architecture points named in these exchanges — location, local execution, isolation, encryption, role-based access, logging — are not a guarantee attached to the demonstration: they are those of the architecture set out in your quotation, and verified before commissioning.
· A 45-minute slot freed up on Thursday through a cancellation, and four patients are on the waiting list for that type of appointment and have been for more than three weeks.
· Eleven patients have an appointment next week with an incomplete administrative file. Nine times it is the same document missing.
· Three letters from correspondents have arrived and are not attached to a file — the name is on them, the date of birth is not.
· The did-not-attend rate on 8 a.m. slots has doubled over six weeks. It is stable on every other slot. morning-watch_4-flags.pdf4 flags · no clinical data
⛓ Source · diary, waiting list, administrative files, incoming letters
What I propose: the four patients waiting, in the order they were listed, with the date of their request and the recorded administrative reason.
The order I give, and the one I do not compute: order of listing is an administrative criterion, checkable, and defensible to a patient who asks. A priority between patients is a medical act — even when it looks like a diary question. "Whoever needs it most" cannot be inferred from a diary, and I will never compute it.
If the clinician wants another order, they give it — that is their decision, entirely theirs, and the four calls reorder in a second.
What I provide for each call: the slot, the duration, and whether the patient has already declined an early slot, because that is diary information and it saves a pointless call. The wording is written for all four; all that is left is to dial.
What I never mention in those preparations: anything from the medical record. Reception does not need it to offer a date. thursday-slot_4-patients.pdf4 patients · order of listing · no triage
⛓ Source · waiting list, listing dates, history of declined slots
Routing follows the practice: a freed slot goes to reception within the hour — after half a day it no longer fills; incomplete files to reception, grouped by missing document and not one at a time; unattached letters to reception, with what is needed to attach them; the 8 a.m. trend to the clinician, because it is an organisational decision.
With a chase: 4 h on a freed slot, 7 days on the rest. Then a monthly summary: by administrative reason, never by patient and never by clinician.
And I do not stop at the flag: the calls and the letters are prepared. The freed slot comes back with the four patients waiting and the availability they have already given; the incomplete files are grouped by missing document, the message to the patient already drafted for each; the unattached letters carry the identifying element they lack, named. Reception dials and sends; it no longer searches.
One rule holds up all the rest, and it is absolute: no medical act — no advice, no clinical guidance, no urgency triage. I see only administration, and that is precisely what allows this arrangement to exist. Any expression suggesting a life-threatening emergency stops the exchange and refers to the emergency number, without my seeking to know more.
✎ Framework · health-data hosting — clinical access is not opened to the agent
The message displayed, as it stands: "I am not a medical service and I cannot assess your situation. If you have chest pain, call the emergency medical number immediately. It is free and available 24 hours a day."
What happens next: the exchange stops. I do not ask since when, or whether it radiates, or the age. Each of those questions is an act of triage, and each costs time.
Why not "just" ask two questions: because an agent that asks two questions gives the impression of assessing, and a patient who feels assessed waits for the answer. The seconds they wait are the only ones that matter.
What I do in parallel: I alert reception immediately, with the time and the patient's exact wording — with no interpretation. Somebody calls back.
The list of expressions triggering that stop is written down, consultable, and extendable by the practice. It holds 84 formulations, and it is deliberately broad: one stop too many costs a call back; a stop missed costs something else. 84-formulations_stop.pdf84 expressions · not one question asked
✎ Framework · 84 stop formulations — an open list, extendable by the practice
The questions I receive most often: "can I take this medicine with the other one?", "do I need to fast beforehand?", "is this normal after the procedure?".
The second looks administrative and is not. "Fasting" depends on the examination, on the patient and sometimes on their treatment — and a generic answer right nine times out of ten is wrong the tenth, for the patient who should have been told otherwise.
What I do: I pass the question to reception, exactly as it was asked, and I tell the patient a call back is planned and within what time. I do not leave them without an answer: I tell them when they will have one.
What I can say without risk: opening hours, the address, access, documents to bring when they appear in the practice's own information sheet, and the usual call-back time.
The boundary is this: I repeat what the practice has written, I produce no health information. what-the-agent-says-and-does-not.pdf7 subjects handled · 5 always passed on
✎ Framework · no advice, no health information produced
What is missing: on 9 of the 11, it is the same certificate. On the other 2, two different documents.
Why the nine: I looked at when those patients booked. All nine came through the online form, where that document is requested after the confirmation screen — that is, after the point at which the patient considers themselves finished.
So it is not nine careless patients: it is a form asking a document of a patient who has already left.
What I propose, in this order: a message to the eleven naming the document and how to send it — not a generic "your file is incomplete", which forces the patient to call to find out what. Then, going forward, move the request before the confirmation screen.
What that represents: over three months, 67 incomplete files, 51 of them on this same document. The form costs roughly four hours of reception time a month in chasing. 11-files_9-one-cause.pdf51 of 67 · a position in the form
⛓ Source · 11 files, online form, 3-month history
What I record: the three letters carry a surname and a first name, and no date of birth. For each one, several patients of the practice share that name — two in one case, three in another.
Why I do not decide: attaching a correspondent's letter to the wrong file puts health information into another person's record. It is the gravest possible error here, and it is silent: nobody notices before the next consultation, sometimes never.
What I provide: for each letter, the same-name files, and what would settle it — the address on the letter, the date of the procedure mentioned, or the correspondent's name if they follow only one of the namesakes.
What I do so the check takes ten seconds: the letter comes up on screen with the same-name files side by side and the deciding element highlighted in each — address, date of the procedure, correspondent followed. Of the 3 letters, 2 are settled on the address at the head, the third on the correspondent's name. The attachment is made with one click, by the person with the file in front of them: a misattached letter cannot be fully repaired — even withdrawn, it has been read, and that is why that click is worth more than a resemblance of mine. 3-letters_namesakes.pdf3 letters · 2 to 3 namesakes each
⛓ Source · 3 letters, same-name files, correspondents followed
What coding support does across the month's 340 procedures: for each procedure as it is already recorded in the administrative report, I propose the code, the applicable modifiers and the amount, and I display the rule from your reference table that supports it, with its version date.
What the check found: 29 codings to review out of 340. 17 are procedure combinations where the billing rule requires a reduction that had not been applied — €412 overbilled for the month, which would come back as a recovery. 9 are the reverse: an applicable modifier not entered, €287 not billed. 3 concern a procedure whose coding depends on an element I do not have, because it is clinical: I pass those back as they are, uncoded, to the practitioner.
What coding support never does, and this is not caution: choosing the procedure. The procedure performed is clinical data, written by the practitioner; I start from what is written and I stop at the administrative code. Coding support that inferred the procedure from the reason for the visit would be performing a medical act, and nothing permits that.
The figure that does not flatter me: of the 29 items raised, 6 were justified on the practitioner's review — 21 % false alarms, all on procedure combinations where your reference table sets an exception I was reading badly. What I did with it: the three exceptions concerned are now read with their condition of application. Over the following month: 24 items raised, 1 justified.
What remains for you: approval, procedure by procedure or in bulk. Every coding arrives with its rule and its date — a few minutes of review instead of a search through the reference table.
⛓ Source · 340 procedures for the month, 29 codings to review, €412 over and €287 unbilled, 21 % then 4 % false alarms
What I record: the did-not-attend rate on the 8 a.m. slot went from 4% to 9% in six weeks. It is stable everywhere else, which rules out a general cause.
What I looked for: what changed six weeks ago. The automatic day-before reminder moved from 6 p.m. to 8 p.m. in a settings change on 26 June.
What that might explain: a reminder at 8 p.m. for an 8 a.m. appointment leaves twelve hours, one of them a night. A reminder at 6 p.m. leaves fourteen hours and an evening — the moment when one can still let you know, reorganise, or cancel and free the slot.
What I do not say: that it is the cause. Six weeks is short, and a calendar coincidence remains possible.
What I propose: go back to 6 p.m. for slots before 9 a.m. only, and measure for six weeks. If the rate returns to 4%, the question is settled; if not, one hypothesis has been eliminated. no-shows-8am_4-to-9-pc.pdf4 → 9% · setting change of 26/06 · stable elsewhere
⛓ Source · did-not-attend rate by slot, settings log, 6 weeks
What is kept: appointment requests and what became of them, the administrative documents missing, letters awaiting attachment, emergency stops with their time and the patient's wording, and the slots offered.
What is not kept, and cannot be: no medical record content, no reason for consultation, no ranking of patients by priority, no statistics per clinician.
Why emergency stops are kept: they are the only exchanges where the agent interrupts a patient. The practice must be able to reread what was said, word for word, and check the call back happened. Over three months, 17 stops, 17 call-backs within the hour.
What the monthly summary contains: the most frequent administrative reasons, the documents most often missing, and the emergency stops with what followed. Three indicators about the organisation, none about people. what-is-kept.pdf5 items kept · 4 out of reach
✎ Framework · retention periods to be set by the practice, within the health-data hosting rules
Your case is not here? That is exactly what a 15-minute conversation is for. Book the free audit →
The uses of AI for a medical secretarial team
Each use corresponds to an agent we deploy, strictly administrative and subject to your approval. Everything that touches on care stays with the professionals.
Administrative letters
Drafts appointment letters, certificates and administrative referral letters from your templates, ready to read over.
Patients' administrative files
Prepares and organises patients' administrative files; flags the missing documents before anything goes out.
Administrative formalities
Assists with putting together treatment forms and other forms, from the elements in the file.
Help with coding
Proposes and checks the administrative coding of procedures, subject to a professional's approval — never a medical decision.
Summarising a file
Summarises the administrative elements of a bulky file so the essentials can be found at a glance.
Administrative questions
Answers administrative questions from your internal documentation (procedures, templates, instructions).
Administrative support for patients
Answers patients' recurring requests: documents to provide, the status of a formality, opening hours.
Need to go further?
These agents handle a different business process, with their own owner and their own price. They are added to this one.
In 15 minutes we identify the most relevant agent — without oversizing the project.
How much time can a medical secretarial team recover?
By automating the drafting of letters, the keeping of administrative files and the formalities, a secretarial team gives time back to the care professionals and cuts down on oversights. The gain depends on your volume of procedures and administration.
The stages of your AI agent project
Audit & scoping
15 minutes to target the use case with the best return.
Quote or direct sign-up
A catalogue offer is bought online; a specific need gets a costed quote.
Design
We design the agent and its guardrails.
Integration & testing
We connect your tools to the agent, which is itself hosted in France.
Rollout
Going live and training your team.
Operation
Continuous supervision and improvement.
One package, one agent
A medical administrative assistant (letters, files, formalities, help with coding), installed and operated for you, with the HDS requirement settled with you at scoping.
Four guarantees that matter to a care practice
Your questions, our answers
Does the AI touch on the medical side or make diagnoses?
Is health data protected?
Does the agent help with the coding of procedures?
How does this differ from the voice agent and the healthcare page?
Is it suitable for a small practice?
How long does it take to deploy the agent?
Which tools can people use to talk to the agent?
Other agents for your care practice
Let's size up the potential in your centre
15 minutes to identify the administrative tasks that eat up the most time — hosted in France (HDS requirement settled at scoping), supervised, with no commitment and never touching care.