AI in nursing: A framework for safe adoption: PwC

Kathryn Burg Plaza

Principal, PwC US

Email

Patrick Voight

Managing Director, PwC US

Email

Key takeaways:

AI entering nursing care is fragmented, often with little input from nurses.
Nursing leaders should use a framework to evaluate tools before they go live and while they are in use.
The framework assesses patient risks, clarifies where nurses should remain decisionmakers, and defines how success will be evaluated.

AI tools are entering nursing workflows and platforms, often with little visibility or input from nurses. AI may support documentation, patient communication, workflow, decision-making, shift scheduling, and so on. The promise is real, but so are the risks. Because nursing directly shapes patient safety, equity, trust, and quality, responsible AI adoption requires a formal framework to assess risk, establish governance, validate performance in real-world settings, and monitor results over time.

Early experience shows why oversight matters. Some ambient inpatient scribes have added to nurses’ workloads by requiring them to narrate care out loud. That’s caused some nurses to abandon their use. Nurses have also reported issues with AI-powered scheduling tools, including schedules that don’t account for staff preferences, workload distribution, fatigue, and contextual staffing needs.

And reports of nurses who said their roles were being replaced by AI-powered technology have raised concerns that efficiency gains could be pursued without sufficient consideration of the consequences for patients, nurses, and care delivery. AI capabilities are increasingly being embedded within the digital platforms nurses already use, making it important for organizations to understand where AI is operating and ensure that appropriate review occurs before it influences nursing work or patient care. These risks reinforce the need for nurses to participate not just in implementation but in decisions about how AI is selected, validated, monitored, and governed. Trust in AI builds over time and is sustained through continuous operations.

How to evaluate AI tools for nursing, step by step

The framework begins with a simple and critical question: What problem is this tool trying to solve? It ends with two equally important questions: Is the tool succeeding? Has it introduced new problems? In between, AI tools should be evaluated deliberately and in real-world care settings, with nurses actively involved in testing usability, workflow fit, safety, and unintended consequences before broader implementation.

Step 1: Start with the problem. What nursing or care problem is the AI intended to solve?
Step 2: Classify the use case by risk level.
- Low risk: Administrative or operational support, with limited influence on care or consequential decisions.
- Moderate risk: Documentation summarization, communication support.
- High risk: Predictive, triage, or clinical recommendation tools.
Step 3: Define the human-AI decision boundary—what AI may recommend, what a nurse must independently verify, when escalation is required, and what AI may never autonomously decide.
Step 4: Specify the evidence needed such as usability, fit-for-purpose performance, equity, workflow impact, and safety, and how the tool will be tested before deployment.
Step 5: Pilot the tool in real care settings before scale-up.
Step 6: Make a go, revise, or stop decision after the pilot by weighing technical performance alongside workflow burden, override rates, and appropriate reliance. Determine whether nurses recognize when to accept, question, override, or escalate an AI output.

Once a tool is in use, keep watching for unsafe outputs, bias, workflow disruption, privacy concerns, and other unintended consequences. Override rates, incident reports, user feedback, escalation frequency, and changes after vendor or model updates can signal emerging risks. A tool that performs well in testing may still create new risks over time, so periodic reevaluation matters.

A deeper dive into risk assessment

These questions can help nurse leaders spot risks before implementation and track them once a tool is in use.

Clinical impact: What are the consequences for patients if the output is wrong, incomplete, or delayed?
Fit with nursing workflow: Does the tool fit how nurses work at the bedside, in handoff, in documentation, and across shifts? Could it add burden, duplicate work, alert fatigue, or interruptions?
Bias and fairness: Does the tool perform consistently across patient groups? Could it worsen disparities by race, language, age, disability, or access?
Privacy and security: What data is being used and how is it stored, shared, and protected, including when a vendor supplies the capability?
Transparency and explainability: Can nurses understand what the tool is doing well enough to use it appropriately? Are its limitations visible?
Accountability and professional responsibility: Who is responsible when AI output is wrong or harmful? Who has the authority to override or escalate?
Staff readiness: Do nurses have the training and confidence to use the tool critically, including an understanding of its limitations, escalation expectations, and signs of poor performance?

Risk assessment should sit inside a larger governance structure

Risk assessment should be part of an interdisciplinary governance structure that includes nursing, informatics, quality, compliance, privacy, IT, and operational leadership. Governance should establish who may propose and approve AI use cases, what evidence is required, who monitors outcomes, and who can pause, revise, or retire a tool. Integrated into existing quality, safety, and compliance processes, this oversight should span the full AI life cycle and include incident reporting, audits, and review of unintended consequences, which also helps prevent informal use of AI.

Next steps: A seven-part path

The assessment steps above guide the review of a single tool. Drawing on established healthcare and AI-governance principles, organizations can structure their approach around seven parts.

Inventory existing, embedded, proposed, and informally used AI, and define the clinical, operational, or workforce problem each tool is intended to solve.

Identify the highest-value opportunities based on expected clinical and operational value, feasibility, strength of evidence, risk, workforce impact, equity, and strategic alignment.

Evaluate each use case for patient safety, clinical impact, workflow and human-factor risks, bias and equity, privacy and security, transparency, and accountability.

Assign risk tiers based on the potential consequences of error, degree of AI influence on care or decisions, ability to detect and correct errors, and scale of exposure. Use those tiers to determine the required level of evidence, validation, oversight, and monitoring.

Define governance structures, accountable owners, decision rights, approval processes, permitted and prohibited uses, escalation pathways, incident reporting, and authority to modify, pause, or stop use of the AI.

Test the AI locally in the intended population, workflow, unit, and clinical context before broader deployment. Establish predefined thresholds for safety, performance, equity, and workflow impact; prepare and educate the workforce; pilot under real-world conditions; and require an explicit go, revise, or stop decision before scale.

Continuously monitor safety, effectiveness, equity, workflow and workforce impact, appropriate user reliance, safety events, and material model or vendor changes. Reassess performance over time and revalidate, modify, restrict, or retire the tool when defined thresholds are not met.

Integrated thoughtfully, AI can improve patient care, nurses’ working lives, and nursing education. Nurses should help develop and adopt these tools, and they should remain the principal decisionmakers for the care they deliver. Governance that starts with nursing needs, matches oversight to risk, tests tools locally, and keeps watching after go-live gives organizations the best chance of choosing tools that help patients and the nurses who care for them.

Angela Mathew also contributed to this article.