Last updated: October 2026
Ask nurses what keeps them from their patients and many will give the same answer: the computer. Over a 12-hour shift, a large share of a hospital nurse’s time goes into charting vital signs, assessments, medications and care plans in the electronic health record. At the same time, nurses are usually the first to notice that a patient «just doesn’t look right», often hours before the numbers change.
Most of the artificial intelligence now entering nursing targets one of those two realities. Some tools try to give time back by drafting documentation from bedside conversations. Others try to turn nurses’ observations into earlier warnings of deterioration. This article looks at both, along with fall and pressure-injury prediction, virtual nursing and staffing, and separates what has been tested rigorously from what is still at the pilot stage.
Where AI enters a nurse’s shift, and how solid the evidence is
«AI in nursing» covers very different tools. The table groups them by the task they support and by the strongest type of evidence currently behind them.
| Task | What the tool does | Strongest evidence so far |
|---|---|---|
| Spotting deterioration | Analyses vital signs and nursing documentation patterns to estimate the risk that a patient is getting worse | Cluster-randomized controlled trial (CONCERN) |
| Documentation | Listens to bedside conversations and drafts flowsheet entries, assessments or notes for the nurse to review | Early pilots and pre-post studies, many reported by health systems or vendors |
| Falls and pressure injuries | Predicts which patients are at higher risk, sometimes combined with sensors or video monitoring | Mostly model-development and single-site implementation studies |
| Virtual nursing | Lets remote nurses support admissions, discharges and education by video, often with AI-assisted monitoring | Operational reports; limited controlled evidence |
| Staffing and scheduling | Forecasts patient demand and acuity to plan staffing | Operational reports; outcomes rarely studied independently |
Evidence level matters because a tool that saves time in a pilot unit may behave differently across a whole hospital, and a model that predicts risk well on past data does not automatically improve patient outcomes when deployed.
The documentation problem AI is trying to solve
Documentation is not optional busywork. It is how care is communicated between shifts, how medication errors are prevented and how quality is audited. But the volume has grown with each new requirement, and studies using electronic health record timestamps and direct observation consistently find that it takes a substantial part of every shift. Estimates vary with the setting and the method used, as the examples below show.
Share of a 12-hour nursing shift spent documenting: selected estimates
Sources: range of 19–40% summarized in Applied Clinical Informatics (2022), which also reports 162 minutes (22.5%) per 12-hour shift at a large academic health system in the southeastern United States before a redesign; 31.11% of shift time in flowsheets from EHR vendor data at NYU Langone Health, JMIR Nursing (2025). Figures come from different hospitals and measurement methods and are not directly comparable.
Earlier efforts to reduce this burden, such as redesigning flowsheets or adding documentation shortcuts, have saved minutes rather than hours. That is why hospitals are now looking at a different approach: letting software capture information as care happens.
Ambient documentation: listening at the bedside
Ambient AI tools first spread among physicians, recording a consultation and drafting a clinical note. Adapting them for nurses is harder. Nursing documentation happens throughout the shift, in short bursts, and much of it goes into structured flowsheet fields rather than free-text notes. Newer nursing versions use speech recognition and large language models to turn a conversation, for example during a pain assessment, into proposed values for the right fields, which the nurse then checks and accepts.
Early results from large U.S. health systems are encouraging but should be read for what they are:
- Mercy reported a 22% reduction in flowsheet documentation time per shift across three pilot units using Microsoft’s nursing ambient tool, and a fall in overtime. These figures come from the health system and a vendor case study.
- Advocate Health reported that the delay between a nurse’s assessment and its appearance in the record fell from 52 to 17 minutes over 13 weeks, according to its nursing innovation leadership.
The most detailed study so far: German long-term care
A pre-post time-motion study published in the Journal of Medical Internet Research in April 2026 evaluated a mobile AI speech assistant (voize) in 14 long-term care facilities in Germany. Trained observers followed 52 registered nurses through full morning shifts before and about eight weeks after implementation, recording roughly 770 hours of work. Observed documentation time fell from an average of about 54 minutes to about 39 minutes per shift; the adjusted reduction was 15 minutes (95% CI 8.2–21.8), or about 28% relative to baseline. Nurses also reported fewer interruptions and higher satisfaction with the documentation system, while overall workplace satisfaction did not change.
The study is unusually transparent about its limits. It had no control group, so other changes over the same period cannot be ruled out. The first author is employed by the company that developed the tool, and the observers were also company employees, although they had no role in developing it and were trained by the university research team. Nurses’ own ratings of documentation completeness and quality tended to fall after implementation; the decline was not statistically significant after correction for multiple testing, but the authors flag it as a possible trade-off between speed and thoroughness. Documentation accuracy was not independently audited.
CONCERN: an early warning system built on nurses’ instincts
Most early warning scores rely on vital signs and lab results. The CONCERN system, developed by researchers led by Columbia University with colleagues at Brigham and Women’s Hospital, takes a different approach. It analyses how nurses document, not only what they record. When nurses grow worried about a patient, they tend to check on them more often, add comments or record observations at unusual times, such as in the middle of the night. The model treats these surveillance patterns as a signal and generates an hourly risk score for the care team.
What sets CONCERN apart is the quality of the evidence. It was tested in a one-year pragmatic cluster-randomized controlled trial across two U.S. health systems. Seventy-four clinical units, rather than individual patients, were randomly assigned to use the system (37 units) or usual care (37 units), covering 60,893 adult hospital encounters. Results were published in Nature Medicine in April 2025, and the trial reported no adverse events.
| Outcome | Result in the CONCERN trial (intervention vs. usual care) |
|---|---|
| In-hospital mortality | 35.6% lower instantaneous risk of death (adjusted hazard ratio 0.64; 95% CI 0.53–0.78) |
| Length of stay | Shorter in the original analysis (adjusted incidence rate ratio 0.91, an 11.2% reduction). An author correction published in April 2026 revised the reported results; the corrected article is the reference for the current estimate |
| Sepsis | 7.5% lower instantaneous risk (adjusted HR 0.93; 95% CI 0.86–0.99) |
| Unanticipated ICU transfers | 24.9% higher instantaneous risk (adjusted HR 1.25; 95% CI 1.09–1.43) |
These are relative, adjusted estimates. A 35.6% lower instantaneous risk does not mean 35.6 percentage points fewer deaths; the absolute difference depends on how often patients died in the comparison units. The results describe what happened during this trial and should not be assumed to apply identically to every hospital or patient population.
Two features of the results are worth noting. The increase in ICU transfers is not a failure: it suggests the system prompted teams to escalate care earlier, which is what an early warning tool is meant to do. And because the model depends on documentation that nurses already produce, it did not add charting work. The researchers have noted, however, that differences in documentation habits between units and hospitals could affect how well it performs elsewhere, and they are studying how to implement it fairly in new settings.
Other ways AI is entering nursing care
Fall and pressure-injury prediction models are among the most studied nursing applications, because both problems are common, harmful and partly preventable. Many hospitals already use simple risk scales; machine learning models aim to make those predictions more precise by drawing on more of the record. The challenge is the same as with any alert: a model that flags too many patients adds work without changing care, and evidence that these tools reduce falls or injuries, rather than simply predicting them, remains limited.
Virtual nursing, in which an experienced nurse supports bedside staff remotely by video, is expanding rapidly, partly in response to staff shortages. AI is increasingly being added to these programmes for monitoring and documentation, but most published information so far comes from the hospitals running them.
What does not change: judgement, accountability and ethics
The American Nurses Association set out its view in a 2022 position statement on the ethical use of AI in nursing practice. Its central principle is that AI should support and enhance nursing’s core values, and that tools which appear to undermine them should be avoided or used only in ways that protect those values. The statement calls on nurses to make sure technology does not compromise human interaction with patients, to insist on high-quality data and validated testing, to watch for bias and to take part in shaping policy and accountability.
In practical terms, that means a nurse who accepts an AI-drafted flowsheet entry is still responsible for its accuracy, and a risk score is still a prompt for assessment rather than a substitute for it.
In the European Union, the rules depend on what a tool does. Software intended to support clinical decisions, such as deterioration prediction, may be subject to the Medical Device Regulation (MDR) and, depending on its classification, to the high-risk provisions of the EU AI Act. Documentation tools that process recorded conversations must also comply with the GDPR’s rules on health data. Requirements and implementation dates depend on the applicable legal framework, so hospitals should check each tool’s regulatory status and intended use.
If you are a patient: what to expect and what to ask
You may notice AI in nursing care only indirectly, for example when a nurse uses a phone or badge to record an assessment, or when a rapid response team arrives because an alert was raised. Reasonable questions include:
- Is our conversation being recorded to help with documentation, and can I say no?
- Who can access the recording, and how long is it kept?
- Will the nurse review what the software writes before it goes into my record?
- Can I see or correct the information recorded about me?
- If an alert was raised about me, what did it mean and what happens next?
For nurse leaders evaluating a tool
The evidence above points to a few questions worth answering before scaling up: Has the tool been evaluated outside the vendor’s own reports, ideally with a comparison group? How will documentation accuracy be audited, not just documentation speed? Were bedside nurses involved in configuring it? And how will the organization know whether time saved is turned into patient care rather than absorbed by higher workloads?
Related reading on Helathora
Limitations of the current evidence
Nursing AI research is growing quickly but unevenly. One well-designed randomized trial supports the CONCERN early warning approach, but it was conducted in two U.S. health systems and needs replication elsewhere. Evidence for ambient documentation in nursing is mostly early, short-term and reported by the organizations and vendors involved, and few studies have yet examined documentation accuracy in detail. Fall and pressure-injury models are often validated only on data from the hospital that built them. Across all categories, outcomes that matter to nurses and patients, such as burnout, time at the bedside, patient experience and harm, are measured far less often than speed or predictive accuracy.
The Bottom Line
FAQ
Will AI replace nurses?
There is no evidence that it will. Current tools handle specific tasks such as drafting documentation or flagging risk. Assessment, clinical judgement, hands-on care, communication and accountability remain with the nurse, and nursing organizations stress that this should not change.
Is it legal to record my conversation with a nurse?
That depends on the country and the hospital’s policies. In the EU, recording and processing health conversations must comply with the GDPR. Hospitals using ambient tools generally inform patients, and you can ask whether recording is happening and how the data is handled.
Can AI-generated nursing notes contain errors?
Yes. Speech recognition and language models can mishear, omit or misplace information. That is why the nurse is expected to review and approve entries before they become part of the medical record.
What is the CONCERN early warning system?
It is a machine learning model that predicts patient deterioration from patterns in how nurses document care, such as how often they check on a patient. In a cluster-randomized trial in two U.S. health systems, patients on units using it had a lower instantaneous risk of in-hospital death than patients on units with usual care, along with more early transfers to intensive care.
Do nurses trust these tools?
Attitudes vary widely. Nurses tend to be more positive when they are involved in choosing and configuring tools, receive training and can see that the technology reduces work rather than adding alerts and checks.
Scientific Studies and Professional Sources
Peer-reviewed research
- Rossetti SC, Dykes PC, Knaplund C, et al. Real-time surveillance system for patient deterioration: a pragmatic cluster-randomized controlled trial. Nature Medicine, 2025; 31(6):1895–1902. PubMed abstract. Trial registration NCT03911687.
- Rossetti SC, et al. Author Correction: Real-time surveillance system for patient deterioration. Nature Medicine, 2026; 32:1556.
- Implementing best practices to redesign workflow and optimize nursing documentation in the electronic health record. Applied Clinical Informatics, 2022.
- Evaluating nurses’ perceptions of documentation in the electronic health record: multimethod analysis. JMIR Nursing, 2025.
- Schwabe K, Ferizaj D, Neumann S, Strube-Lahmann S, Lahmann N. Time savings through an AI speech assistant for nursing documentation: pre-post time-motion study in German long-term care. Journal of Medical Internet Research, 2026; 28:e86078.
Professional guidance and institutional information (accessed October 2026)
- American Nurses Association. The Ethical Use of Artificial Intelligence in Nursing Practice, position statement, 2022.
- Columbia University Irving Medical Center. Nurses and AI collaborate to save lives, 2025.
Health system and industry reports
- Becker’s Hospital Review. Why ambient AI is harder for nurses than physicians (Mercy and Advocate Health results).
This article is for general information only and does not constitute medical, nursing or legal advice. Results from individual studies and health systems may not reflect performance in other settings. Organization-reported figures have not necessarily been independently verified. Helathora is not affiliated with any of the companies, health systems or institutions mentioned.