ANALYSISHEALTHCARE
AI in Healthcare: What Actually Helps Care Teams — and What It Costs
The AI that works in hospitals does not replace caregivers: it gives them time back — documentation, coding, flow, safety — provided it is run as an investment, from ready data to a decision on a fixed date.
Preparing PDF…

A sector declared at a standstill — yet already moving
In a healthcare or social care facility, the question is not whether AI will matter, but where it already helps, at what price, and under what conditions it respects care. The honest answer starts with an admission of delay: surveyed in late 2025, 45% of hospital staff report that no AI use has been adopted at their facility (Ifop / Acteurs Publics / UniHA). Half the field is therefore at a declared standstill — not from refusal, but for lack of a framework, ready data, and time.
This delay is not specific to healthcare. Across French companies with ten or more employees, AI remains a minority practice — 10% in 2024, versus 33% above 250 employees and 9% between 10 and 49 (INSEE) — and organization size matters more than sector. A thirty-bed facility starts with the same constraints as an industrial SMB: little in-house expertise, scattered data, scarce time. The difference lies elsewhere: in a hospital, the critical resource is not money, it is caregiver time — and that is precisely what the use cases that succeed start by giving back.
Where AI has entered, the hierarchy of uses is clear, and it is clinical first: medical imaging analysis leads (29%), followed by report drafting (23%), diagnostic support (15%), data analysis (13%) and administrative tasks (10%) (Ifop / Acteurs Publics / UniHA). In other words, maturity follows the best-tooled clinical use cases — not the tasks reputed to be "easy".
In hospitals, AI enters through the clinic — when it enters at all
The contrast with the broader momentum is striking. Among French SMB and mid-market executives, generative AI use went from 31% to 55% in one year, and 43% have defined an AI strategy (Bpifrance Le Lab). Healthcare is no exception to this movement — it simply approaches it with heavier data, regulatory and liability constraints than other sectors. One principle runs through this entire report and conditions the buy-in of care teams: AI assists; it does not replace clinical judgment.
The paradox: many trials, very little scale
Healthcare shares the paradox of every sector, in a sharper form. Everywhere, using AI has become ordinary — 88% of organizations use it in at least one function (McKinsey) — but barely 5% have genuinely deployed it at scale (Capgemini / Microsoft), and about 95% of generative AI pilots produce no measurable return (MIT Project NANDA). In hospitals, this mechanism has a familiar face: the brilliant imaging demonstrator that never gets integrated into the electronic patient record, the scribe experiment that stays confined to one department.
The cause is the same as elsewhere, aggravated by context: data never prepared, care processes never changed, measurement never set — and, specific to the sector, a clinical governance that must be built before any deployment. A pilot launched without a baseline number (data-entry hours, under-coded stays, no-show rates) can neither succeed nor fail: it can only last. The rest of this report reviews what, by contrast, already works in routine — with published results, in France and abroad.
In a hospital, the pilot that succeeds does not start with an algorithm: it starts with a number — data-entry hours, under-coded stays, poorly anticipated beds.
Five families already giving care time back
As in industry, value is not spread evenly: it concentrates in five use-case families that share three traits — a repeated task, a measurable current cost, a gain that translates directly into caregiver time, revenue or avoided events. All are documented below by real deployments, not laboratory experiments.
Clinical documentation is the most spectacular family, because it attacks the largest pool of lost time. Doctolib's consultation assistant saves the physicians who use it nearly an hour a day (Doctolib). Great Ormond Street Hospital evaluated the TORTUS ambient scribe across nine London NHS sites and more than 17,000 consultations, with +23.5% patient interaction time (GOSH / NHS); Oxford is piloting four scribe solutions in parallel to choose on evidence (Oxford University Hospitals NHS FT). In France, the Arles hospital is the first public facility to have integrated an ambient scribe directly into the electronic patient record (Nabla), followed by the Augustines clinic in Malestroit. And the logic extends to pharma: at Novo Nordisk, drafting a clinical study report went from fifteen weeks to ten minutes (Novo Nordisk / Anthropic).
Coding and billing is the family with the most directly quantifiable return, because every badly coded stay is lost revenue. The Saint-Quentin hospital has run third-level AI coding control since 2016 — more than 400,000 euros recovered in 2019, about 1,600 euros per revalued stay (I4H). In the UK, the NHS's first automated clinical coding project, started in endoscopy at Kettering, frees the equivalent of one full-time post per year, worth more than 35,000 pounds (Kettering General Hospital / vendor). In the US, Geisinger deployed autonomous coding across the nine emergency departments of its network (Geisinger / Nym); in France, Annecy Genevois (Sancare) and Rambouillet (Collective Thinking) equip their medical information physicians — the human accepts or rejects each proposal, the AI prepares.
Flow and capacity is the family that speaks to management. The Valenciennes hospital anticipates emergency arrivals and bed needs up to five days ahead, with reliability reported near 90% (CALYPS); Amiens university hospital predicts admissions from the emergency department with a published model (AUROC 83.1%) (CHU Amiens-Picardie); Utrecht cuts missed appointments by 15 to 20% with an open-source model (UMC Utrecht). On scheduling, Brest university hospital halves the time spent on staff rosters (CHU de Brest / Hopia), Vivalto Santé automates operating-theater schedules, and the founding example comes from Hong Kong: a constraint-based engine has produced the rosters of more than 48,000 staff across 40+ public hospitals since 1997 (Hong Kong Hospital Authority). Scheduling AI is not a promise: it is a 25-year-old technology.
Safety and vigilance is the most social-care family — and the most human. At Poitiers university hospital, the Aladin+ connected lamp detects loss of verticality and notifies teams: about 70 falls detected in three months, with intervention between 30 seconds and 3 minutes, and an extension to five facilities of the regional hospital group (CHU de Poitiers / Domalys). Orikio's acoustic monitoring equips more than 17 DomusVi residences, Kaspard detects falls without cameras at Clariane, and more than 1,000 Belgian care-home rooms are equipped (DomusVi; Clariane; sector press). At system level, the NHS is launching an AI early-warning system on maternity data — stillbirth, neonatal death, brain injury — in near real time (NHS England), and Lausanne university hospital detects sepsis earlier to speed up care (CHUV). The common trait: AI continuously watches what reduced teams cannot watch, at night in particular.
Knowledge and back office, finally, follow the general movement of organizations. The Rouen hospital group deployed a trusted generative AI to 15,000 professionals, with data hosted 30 km from Rouen — a grant application goes from three weeks to two days, a procurement specification from fifteen days to one (GHT Rouen / Orange Business). AstraZeneca screens scientific literature seven times faster for its systematic reviews (AstraZeneca), and the European Medicines Agency uses AI to spot safety signals in the literature (EMA). The prerequisite for this family is authorised access to relevant, quality data, appropriate hosting and integration proportionate to the use case; where several cases must share that data, the health data warehouse pools the effort. The Lyon university hospitals, AP-HP (38 hospitals), Bordeaux-Limoges-Poitiers (EDS NOVA) and the hospital groups pooling their warehouses have made that pooling choice at their scale; a narrow administrative use can do without it.
Two speeds of return — and what determines them
Five families, one shared promise: give care time back — and prove it, end figure against baseline figure.
Proof at facility scale
The objection "this is for university hospitals and large groups" holds no better in healthcare than elsewhere. The French deployments that succeed share the same rules: a narrow scope, data already there, a baseline figure, a dated decision.
None of these facilities started with an AI program. Each started with a quantified problem — under-coded stays, time-consuming rosters, night-time falls — and chose the tool afterwards. The order of operations is the method.
Admin, flow, clinical: evidence that does not transfer
In a care facility, the economics of an AI project depend first on the nature of the use. Administration and documentation give time back and can secure revenue; flow organisation improves anticipation; clinical support requires its own validation, with the clinician deciding. These forms of evidence do not substitute for one another: a scribe that saves data-entry time does not, on its own, demonstrate a better outcome for the patient.
Three pathways, three appropriate validations
Administration and documentation
- Quantified data-entry time or stays to re-value
- EHR integration
- Clinician review; coding accepted or refused by the medical information physician
- Proof: time given back or revenue — not a clinical effect
Flow organisation
- Admissions and staffing history assembled
- Forecast checked against actual
- Adoption by managers
- Proof: anticipation measured on comparable periods
Clinical support
- Clinical question and protocol defined
- Qualified clinical data, regulatory framework
- Appropriate clinical validation; the physician decides
- Proof: the protocol’s clinical indicators
YPrtnrs Institute reading, qualitative: evidence from one pathway does not carry over to another.
Four registers, four ways to scope and validate
| Use | Cost to scope | Relevant indicator | Dependency | Validation |
|---|---|---|---|---|
| Administration and documentation | EHR integration, review of reports, training | Data-entry time, patient interaction time, re-valued stays | Consultation and EHR integration, patient information | Clinician review; coding accepted by the medical information physician |
| Flow organisation | Assembling history, external data, adoption by managers | Forecast versus actual, no-shows, rostering time | Admissions and HR history | Managers and leadership, on comparable periods |
| Clinical support | Clinical validation, integration, continuous vigilance | Clinical indicators defined by the validation protocol | Qualified clinical data, applicable regulatory framework | Appropriate clinical validation; the physician decides |
| Data and security | Appropriate hosting, integration, access rights | Compliance, incidents, access traceability | Authorised access to relevant, quality data; a health data warehouse where pooling is needed | Data and security governance |
The cases cited stay within their register. Saint-Quentin hospital recovered more than €400,000 in 2019 with AI-based PMSI coding control (I4H): that is revenue, not a measure of care quality. Doctolib reports nearly one hour a day saved by physicians using its consultation assistant (Doctolib), and Great Ormond Street Hospital’s evaluation of the TORTUS scribe across nine NHS sites measured +23.5% patient interaction time (GOSH / NHS): time given back, not demonstrated clinical effectiveness. Valenciennes hospital reports reliability close to 90% for its five-day forecast (CALYPS); Amiens University Hospital’s admission model is published with an AUROC of 83.1% (CHU Amiens-Picardie) — forecasting performance, to be validated in each context.
For leadership the consequence is simple: cost what is not the licence (EHR integration, training — 6% of staff trained today (Ifop / Acteurs Publics / UniHA) —, clinical governance, operations), choose the indicator that matches the register of the use, and never draw a clinical conclusion from an administrative gain.
A facility project's journey — and its four breaking points
Hospital AI projects that stall look alike. Laid on top of each other, they reveal a typical journey and four breaking points — each avoidable by a decision taken before launch, and two of them are specific to the sector.
- Break #1 — the demonstrator with no path to care. The pilot is chosen for its showcase effect (often imaging), with no path to patient-record integration and no baseline figure: it joins the ≈ 95% with no measurable return (MIT NANDA). The counter: choose within the five documented families, where the current cost can be measured — data-entry hours, under-coded stays, no-shows, falls.
- Break #2 — unprepared data. Without organised data access, every use case pays the access cost all over again; a health data warehouse pools that cost when several cases justify it; that is the point of the Lyon, AP-HP and EDS NOVA programs. The counter: audit the data of the chosen case only, within fifteen days — and, if several cases justify it, put the warehouse on the multi-year plan, not on the pilot's budget.
- Break #3 — training and integration left out of the budget. The budget describes the licence, then discovers training (6% of staff trained), EHR integration, organisational change and operations. The countermeasure: require a total-cost estimate where these items are visible before signature, and an indicator that matches the register of the use — administrative, flow or clinical.
- Break #4 — trust without clinical governance. 52% of staff cite ethics and security as the leading barrier (Ifop / Acteurs Publics / UniHA) — and they are right to demand a framework: controlled data hosting (Rouen hosts its data 30 km away), systematic human validation (the medical information physician accepts or rejects every proposed code), and the inviolable principle: AI assists, clinical judgment decides. The counter: a named clinical governance BEFORE the pilot's first day, and a decision date — industrialize or stop — in the executive calendar.
In healthcare, trust is not an obstacle to deployment: it is the deployment itself. It is built by the framework, not by the promise.
What leadership must decide — and when
Neither a five-year plan nor wait-and-see: the trajectory that works holds in five decisions over ninety days, each with its indicator.
- Choose a family, not a technology. Among the five documented families, take the one where your current cost is most measurable. Indicator: a validated baseline figure (data-entry hours, euros of under-coded stays, no-show rate, night-time falls).
- Audit the data of the chosen case — not the whole facility. Does it exist, is it accessible, can the patient record be integrated? Indicator: a written verdict within fifteen days.
- Cost at total cost and name the expected proof. Licence + integration + training + governance + operations; and state whether the use targets time, revenue, better anticipation or a clinical effect — the latter requiring appropriate clinical validation. Indicator: a budget where items outside the licence are visible and the proof indicator is written down.
- Name the clinical governance and set the decision date before starting. Industrialize or stop, at twelve or twenty-four weeks. Indicator: a named clinical lead and a date in the executive calendar.
- Instrument the measurement from day one. End figure against baseline figure, same scope. Indicator: a before/after dashboard active from week 1.
Doing nothing also has a cost, and in healthcare it is not counted only in euros: every data-entry hour not given back is an hour of care lost, every under-coded stay is lost revenue, every undetected fall is an avoidable risk. The facilities cited in this report had no more resources and no more data than the others — they had a method.
Method
This report draws on three materials. First, reference surveys and studies — the national hospital AI barometer (Ifop for Acteurs Publics and UniHA), INSEE, Bpifrance Le Lab, McKinsey, MIT, Capgemini, BCG — whose full references appear below. Second, the Institute's knowledge base: more than 200 verified cases and data points, including some forty specific to healthcare and social care, each tied to an original publication. Third, the Institute's own analysis, whose estimates are systematically flagged as such ("YPrtnrs Institute estimate") and never presented as measurements.
Resources
Boston Consulting Group. (2026). Scaling AI Requires New Processes, Not Just New Tools [cross-sector 10/20/70 effort-allocation principle; not sector-specific, not an accounting breakdown]. bcg.com.
Bpifrance Le Lab. (2025). L'IA dans les PME et ETI françaises : une révolution tranquille [survey of 1,209 leaders; updated late 2025]. Bpifrance.
CALYPS. (2025). [SANIIA at Valenciennes hospital: emergency and bed-need forecasting up to 5 days ahead].
Capgemini Research Institute, & Microsoft. (2025, October). [Study on scaling AI in organizations].
CHU Amiens-Picardie. (2024). [3P-U model predicting admissions from the emergency department, AUROC 83.1%].
Great Ormond Street Hospital / NHS. (2025). Ambient voice technologies: from early enthusiasm to informed adoption [TORTUS evaluation, 9 sites, 17,000+ consultations].
Ifop, for Acteurs Publics and UniHA. (2025). Baromètre national IA à l'hôpital [1,051 hospital staff, June 23 – September 26, 2025].
INSEE. (2025). L'économie et la société à l'ère du numérique — Intelligence artificielle dans les entreprises [TIC entreprises 2024 survey].
I4H. (2020). [Saint-Quentin hospital: third-level AI coding control — more than €400,000 in 2019].
Kettering General Hospital — University Hospitals of Northamptonshire NHS Group. (2025). [Automated clinical coding: the equivalent of one FTE per year, more than £35,000 — vendor figure].
McKinsey & Company. (2025, November). The State of AI [McKinsey Global Survey].
MIT Project NANDA. (2025, August). The GenAI Divide: State of AI in Business 2025. Massachusetts Institute of Technology.
NHS England, Department of Health and Social Care, & Care Quality Commission. (2025). [AI early-warning system on maternity data].
UMC Utrecht. (2024). [No-show prediction, −15 to −20% — open-source code].
Facility and vendor communications (published deployments): AP-HP (data warehouse, 38 hospitals); Arles hospital, & Nabla (scribe integrated into the EPR); AstraZeneca (literature review ×7); Bordeaux, Limoges and Poitiers university hospitals (EDS NOVA); Brest university hospital, & Hopia (rosters −50%); CHUV (sepsis detection); CHwapi, & MintT (3D sensors in geriatrics); Clariane, & Kaspard (camera-free detection); Doctolib (consultation assistant); DomusVi, & Orikio (acoustic monitoring, 17+ residences); EMA (literature screening); Geisinger, & Nym (autonomous coding, 9 emergency departments); Rouen hospital group, & Orange Business (trusted generative AI, 15,000 professionals); Hong Kong Hospital Authority (rosters for 48,000+ staff, AAAI/IAAI publication); Lyon university hospitals (data warehouse); Maubeuge hospital, & MiPih (emergency forecasting); Novo Nordisk (clinical study reports); Oxford University Hospitals NHS FT (4-scribe pilot); Poitiers university hospital, & Domalys (Aladin+, Vienne hospital group); Vivalto Santé, & Hopia (operating-theater schedules).

