In healthcare, artificial intelligence is moving fast on imaging and on data, and far more slowly on the daily life of a practice. The gap does not come from technology. It comes from the order in which projects are opened.
Tandem supports around thirty companies on AI applied to operations, including four healthcare assignments. This article gives the grid we use on assignment. It ranks eleven projects by value, feasibility and regulatory risk. The result often surprises management, and that is what makes it useful.
Where does AI in French healthcare actually stand?
AI is already in production in most public hospitals. A July 2025 survey by the French hospital federation measured that 65 % of public healthcare facilities were already using AI technologies in production, across 110 participating facilities. The figure is quoted by the French health authority and the data protection regulator in their guide on the proper use of AI systems in care settings, published as a working document in February 2026.
The state has set its frame. The national strategy on AI and health data was launched in July 2025, with a second chapter dedicated to AI in November 2025. It carries a training plan for 500,000 professionals over five years, funded with roughly 119 million euros, covering medical, paramedical and social staff.
- What is moving inside facilities. Imaging, clinical decision support and automated medical reports. Those are the three priorities stated by the national strategy.
- What is stuck inside practices. The phone, incomplete files and invoicing. None of these is a clinical AI topic, and that is precisely why they move more slowly.
Two announcements from 2026 show two distinct directions. The Elsan group renewed its AI imaging partnership in May 2026. The group reports 217 private hospitals and health centres. It states that more than 100,000 patients benefited from these solutions in 2025. Undetected fractures reportedly fell from 18 % to under 3 % (Elsan press release, May 2026). The group specifies that a solution is only deployed after validation by practitioners.
Ramsay Santé took the other direction, the data one. In July 2026 the group announced the launch of its first health data warehouse, declared compliant with the regulator's framework in May 2026. Around twenty medicine, surgery and obstetrics facilities are due to join in the first quarter of 2027. A first research project covers aortic prostheses, with more than 70,000 patients across eight surgical centres.
The first time AI frees up is not medical time
In community medicine, the first reservoir is the front desk. The French statistics directorate for health measured that 84 % of self-employed general practitioners reported having secretarial support in early 2022, but only 51 % an on-site secretary (DREES, Études et Résultats no. 1245). One general practitioner in six handles their own front desk. In solo practice the equipped share falls to 64 %, and only 24 % have an on-site secretary.
Online appointment booking was used by only 36 % of self-employed general practitioners at the same date. It rises to 48 % among those under 50 and drops to 26 % among those aged 60 and over. In other words, part of the call volume that saturates practices comes from a channel that was never opened.
On missed appointments, caution. The figure of 27 million missed consultations a year circulates everywhere, including in parliamentary debate. It comes from a 2022 survey by one regional union of health professionals, and franceinfo documented that no consolidated study establishes it. The range quoted by the academy of medicine and the medical council runs from 6 to 10 % of insured patients. Measure your own rate before sizing anything.
Augmentation or automation, two projects and two timelines
Every use case belongs to one or the other, and confusing them is the most frequent cause of overrun. In augmentation, AI prepares and the professional decides. The practitioner or the secretary stays in charge of the file, nothing leaves without validation. In automation, the flow runs and the team only handles exceptions.
| Augmentation | Automation | |
|---|---|---|
| Time to service | 2 to 6 weeks | 3 to 9 months depending on system access |
| Nature of the gain | Time per task, quality, consistency | Unit cost of the flow, processing time |
| Ease of measurement | Low, time saved is self-reported | High, volume and rework rate sit in the logs |
| Regulatory exposure | Low, the human remains the author of the decision | High, traceability and patient information |
| What makes it fail | Adoption, an unused tool produces nothing | Reference data, dirty data blocks the chain |
There is a level before automation, and it is heavily underused. I filmed a whole video on the fact that AI agents are too often the first reflex, when a configured assistant or a simple automation is enough in most cases. The four levels, from assistant to autonomous agent, are detailed in our article on levels of automation and in this video on the overuse of agents. In healthcare, skipping those steps is expensive, because each extra level adds regulatory exposure.
Which AI projects should a health centre open first?
Three scores from 1 to 5 are enough to decide. Value, which adds freed capacity and gains beyond time. Feasibility, which measures data access and whether rules are written down. Regulatory risk, which runs from internal data with no patient impact to a decision affecting care. The index is value times feasibility, divided by risk. The full logic is described in our AI audit method.
The result is unpopular, and it holds. The most requested project, the appointment-booking voice agent, comes tenth out of eleven. Assisted medical report writing, which frees the most clinical time, only comes seventh, because it touches the recording of the patient's voice. And the project nobody talks about, putting the document corpus back in order, comes first.
The document corpus comes before everything else
In a practice or a health centre, useful knowledge is scattered. Care protocols, procedure nomenclatures, coding rules, coverage conditions per insurer, template letters, site-specific internal instructions. A practitioner or a secretary spends a considerable amount of time looking for information that already exists somewhere.
It is the best-scoring use case in the grid, for three reasons. It touches no patient data, so its regulatory risk is minimal. It goes live in a few weeks. And it conditions half the projects that follow, because an agent answering the phone needs that same document foundation.
The technical building block is called RAG, retrieval augmented generation. In December 2025 I published a one-pager comparing classic RAG and agentic RAG, and my advice has not changed. Focus on data pre-processing and on re-ranking the retrieved passages. Those two steps are enough in the vast majority of use cases. The agentic version, where the model chooses which sources to query, mainly brings native governance, with traceability and respect for access rights.
I filmed the full build of an agent like this, without writing code. Our guide to RAG in business walks through the method step by step.
Why the patient voice agent is not the first project
The technology works, and I say that without reservation. A voice agent chains three blocks. Speech converted to text, a model that produces the answer, then speech synthesis. I detailed that voice-to-voice pipeline and the options available at each layer. Latency has come down to an acceptable level, the French voice is convincing, and the cost per call is measured in cents at the provider. A French phone number takes 24 to 48 hours with a company registration extract.
Our own assignments confirm it. At a medical equipment rental group, an inbound voice agent fully handled around 70 % of calls, with conditional transfer to a human and SMS follow-up. The same setup places more than 10,000 outbound calls a month for administrative follow-ups. The return is real, and it is measurable in the logs.
- What pushes it down the grid. Regulatory risk, not feasibility. The caller must know they are speaking to an AI, and a patient's reason for calling is health data as soon as it is tied to a name.
- The metric that counts. The share of requests closed without human intervention, corrected for the call-back rate. An agent handling 70 % of calls where 30 % call back the next day really handles 49 %.
- Steering runs on transcripts. Every call leaves a text, a duration and a category. An unexpected category that starts rising is the best available warning signal.
So it is not a project to rule out, it is a project to place after the foundation. Our complete guide to AI voice agents details the pipeline, the technology trade-offs and the production pitfalls.
What the regulatory frame actually allows in healthcare
Four gates are crossed in order, and a closed gate is not a technical obstacle. It is a project to remove from the current wave.
The most misunderstood point concerns liability. The guide from the French health authority and the data protection regulator is explicit. No specific liability regime has been established to date for health professionals using AI. The professional remains fully responsible even when the act is performed with the assistance of an AI system. They answer for how they use it, and for the degree of autonomy they grant the tool.
The guide also grades supervision. It is full, with human review of 100 % of outputs, for decisions with a high impact on the patient. It can be lighter and targeted on sensitive cases for non-clinical uses. And it recalls that a care organisation removing all human review would today be contrary to regulation.
| Deadline | What applies |
|---|---|
| 2 August 2026 | Article 50 of the EU AI Act. Transparency, the person must know they are speaking to an AI |
| 2 December 2027 | Obligations for Annex III high-risk systems, including healthcare, postponed by omnibus Regulation (EU) 2026/1744 |
| 2 August 2028 | AI embedded in a medical device, postponed from 2 August 2027 |
| Pending | Implementing texts for article L. 4001-3 of the French public health code on informing the patient |
What AI cannot do in a healthcare practice
Five limits, and none of them is a temporary difficulty. Knowing them avoids committing budget to a use case that cannot succeed.
- Decide a diagnosis or a coverage refusal on its own. The GDPR guarantees the patient the right not to be subject to a fully automated decision significantly affecting them. A system can prepare and recommend, a human must be able to step in.
- Repair inconsistent reference data. Three labels for the same procedure and two coding grids depending on the site, no model settles that. It is a data project, not an AI project.
- Run a flow whose rules are not written down. When the rule lives in the heads of two experienced people, the first task is to write it down. That value outlives the project.
- Absorb exception cases. In a real flow, a significant share of files falls outside the frame. An illegible document, an atypical situation, a patient in dispute with their insurer. Those cases carry most of the risk and will stay human.
- Run without anyone watching it. This is the most expensive limit because it is invisible at launch. An automated flow drifts. A format changes, a product evolves, a model is updated. Without weekly sample-based quality control, quality declines for months before anyone notices.
Four healthcare assignments, and what they taught us
Tandem publishes the sector, the volumes, the scope sold and the sequence followed. Not our clients' figures, which they have not authorised us to circulate, and which would be useless to you outside their volume context. The assignments below are therefore anonymised.
- A network of around twenty dental centres, 500 to 600 employees. Fifteen days of on-site audit alongside the teams, to map the patient journey end to end. Several use cases costed in three scenarios, prioritised by volume and expected return. Two of the most expensive friction points were solved by switching on features already present in the group's tools.
- A medical equipment rental business, a group of several companies. An inbound voice agent live 24 hours a day with SMS follow-up and conditional transfer, plus an outbound administrative follow-up campaign. Around 70 % of inbound calls fully handled by AI.
- A digital health software vendor with several thousand employees, operating in two countries. Three months of adoption support across five business teams, with governance adapted to each country's regulatory constraints. The platform already existed, the challenge was usage.
- A pharmaceutical distribution player on the African continent. Mapping of more than twenty use cases, sorted into four families by technical nature, from business rules to agentic systems, with a recommended technical foundation shareable across subsidiaries.
Do you need an AI agency to deploy AI in healthcare?
Not always, and the question is settled case by case. Buy for processes shared across the sector, such as online booking, appointment reminders or transcription. Build when the constraint is structural, notably as soon as health data circulates or certified hosting becomes mandatory. Our AI for healthcare page details the six families of use cases we equip.
A hybrid approach works well in healthcare. A bought base to move fast on the patient channel, plus an in-house rules module for the part touching coverage and compliance. That is also the setup we chose in insurance, where the regulatory constraint is of the same order.
On the prioritisation method, our AI audit starts from the real work before costing anything. And our article on AI in insurance applies the same grid to a neighbouring sector, which helps calibrate your own scores.
Which AI project should healthcare start with?
Start with the document corpus and with appointment reminders. Those are the top two lines of the ranking, and that is no accident. The corpus conditions the projects that follow and touches no patient data. Reminders have the best feasibility in the grid, because the processing rule is already written down.
Keep clinical decision support for later, and prepare it seriously. The 2 December 2027 deadline is coming, human supervision will remain required, and an undocumented system will be a liability. Before any arbitration, measure the volume, unit time and rework rate of your three heaviest flows. One week of sampling is enough.
One last rule, valid everywhere in healthcare. If your system cannot say that it is unsure, it is not ready to touch a patient file.
