With predictable costs and on-site data control, healthcare leaders no longer have to choose between innovation and compliance.

Across health systems, the debate over artificial intelligence has quietly moved on. The question is no longer “should we adopt AI.” It’s “how do we scale agentic AI safely and sustainably.” For many healthcare workflows, that means deciding which AI workloads should run close to sensitive data and expert users, and which should scale through shared enterprise or cloud environments. Agentic AI, meaning systems that can reason across tasks, orchestrate tools, and support next-step actions with appropriate human oversight, holds real promise for diagnostics, documentation, and operations. But moving from pilot to production is where many organizations stall.

The reasons are consistent. According to the Dell Modern Enterprise Readiness Study (June 2026) 1 , which surveyed 376 healthcare and life sciences leaders, almost two thirds (62%) have security or compliance concerns with employees using unapproved AI tools on company devices. At the same time, more than half (53%) rank establishing and scaling AI capabilities as a top priority for the next 24 months. That tension between ambition and reality defines the current moment.

Two barriers surface again and again: unclear cost and return on investment and worries over data privacy and HIPAA compliance. Both can be addressed with the right workload placement, governance model, and infrastructure strategy. Here’s what the data and the most recent published research show, and what leaders can do about it.

The Real Barrier Isn’t Adoption, It’s Scale

Most health systems have already run an AI pilot of some kind. The harder work is turning those experiments into secure, organization-wide deployments that hold up under clinical and financial scrutiny.

Leaders are setting a high bar before they commit. According to the Dell Modern Enterprise Readiness Study , 72% say they can’t move forward with new AI projects without a partner who can demonstrate clear business impact and a defined risk management plan. That discipline is healthy, but it also exposes a common pitfall: pilot sprawl. When isolated projects multiply without shared goals, data standards, or governance, they rarely reach production and rarely prove value.

A white paper published alongside the research, Building a Practical Architecture for Healthcare AI , puts it plainly: once AI begins to support real decisions or high-volume processes, architecture becomes inseparable from outcomes. The document argues that healthcare AI should be treated as a systems design challenge, not an overlay added on top of existing infrastructure.

The same white paper highlights a structural issue that compounds the scaling problem: healthcare data is distributed across clinical, operational, and imaging systems, often in unstructured or multimodal formats. Imaging data sits in one set of repositories. Clinical records sit in another. That fragmentation means the context required for AI is often present, but not accessible in a coordinated, repeatable way. Solving scale, therefore, starts with solving data.

The organizations pulling ahead aren’t the ones running the most pilots. They’re the ones scaling the right ones with a clear plan for cost, data, and compliance.

Barrier One: Cost and ROI Uncertainty

Public cloud made it easy to start with AI. It also makes costs hard to predict. Large model workloads common in clinical settings, including imaging analysis, natural language processing of clinical notes, and continuous inference, can drive cloud compute bills up quickly and unpredictably as usage grows.

That unpredictability is driving a significant shift. According to the eBook, Revolutionize Healthcare and Life Sciences with AI , 79% of enterprises have already moved AI workloads outside the public cloud, citing the need for predictable performance, stronger data governance, and more consistent cost structures. The same eBook notes that concerns about data sovereignty, latency, and operational control, combined with variable cloud pricing, are accelerating the move toward on-premises deployments.

A growing approach is deskside agentic AI: running models locally on high-performance workstations. The goal is not to move every workload on-premises, but to place each workload where it best fits: close to clinicians, researchers, imaging data, or operations teams when responsiveness, privacy, cost predictability, or expert review matter; and on shared enterprise or cloud infrastructure when broader scale is needed. Dell Pro Precision systems with NVIDIA RTX PRO GPUs, for example, are designed to support demanding healthcare workloads such as medical imaging and genomics at the desk-side level, according to When Every Second Counts: Five Ways to Scale Healthcare AI with Confidence . That architecture keeps costs steady and knowable.

The economics can be compelling. AI-agent assisted knowledge workers that adopt local processing can break even against leading frontier pricing in as little as three months 2 , after which the savings compound. For a CFO, that’s a return on investment timeline short enough to model with confidence.

Critically, getting the economics right requires right-sizing the infrastructure to match the workload. Revolutionize Healthcare and Life Sciences with AI offers a practical framework: factors such as model size, precision requirements, context length, latency targets, and input/output sequence length all shape what’s needed to perform reliably. Environments that are undersized for these needs slow clinical workflows and delay insights, while overbuilt systems add unnecessary cost. Neither outcome serves the business case for AI.

Key takeaway: Predictable, on-premises compute turns AI spending from a variable expense into a planned investment, provided the infrastructure is sized for the workload from the start.

Barrier Two: Data Privacy and HIPAA Compliance

Protecting patient health information (PHI) is the second major concern, and for good reason. The Dell Modern Enterprise Readiness Study found that 73% of healthcare and life sciences leaders are concerned about complying with government regulations when using their data for AI. Cloud-based models can add risk by moving sensitive data outside an organization’s direct control.

Preference is shifting accordingly. The same study found that 68% prefer to keep sensitive data and AI models on infrastructure they own or in a trusted local cloud. Local and edge-oriented deployment directly supports that preference. As Building a Practical Architecture for Healthcare AI notes, some workloads are best placed close to the data or close to the user, particularly when large files are expensive to move, latency matters, or sensitive data should remain under tighter institutional control.

The white paper describes imaging as one of the clearest examples of where on-premises architecture matters. Imaging workflows are data-intensive, time-sensitive, and closely integrated with clinician workflows. When AI runs closer to where imaging data already lives, organizations can reduce unnecessary data movement, keep more information within controlled environments, and simplify parts of the compliance and risk-management workflow.

Consider a practical scenario from When Every Second Counts : a radiology team wants to use AI to accelerate image review and surface critical findings sooner. With a local deskside or on-premises setup, the imaging data and the model stay inside the hospital network. Results reach the clinician faster, and no PHI travels to an external cloud. The same principle applies to summarizing clinical notes or supporting operational workflows. Sensitive information is processed where it lives.

Real-world results support the case. Northwestern Medicine achieved up to a 40% productivity improvement in radiology image review using an on-premises AI deployment. The white paper describes this as evidence that when infrastructure is aligned with data locality, user needs, and institutional control, production AI can deliver measurable workflow gains.

Every organization should validate its own compliance posture but architecturally, keeping data in place removes an entire category of exposure.

What Research Shows: Governance Can’t Be an Afterthought

Beyond cost and privacy, the research consistently points to a third enabler that determines whether AI scales or stalls: governance.

Governance in healthcare AI is not only a policy issue; it is an engineering issue. As AI becomes more embedded in workflows, organizations need visibility into how systems access data, how outputs are produced, and how activity can be audited or corrected over time. That visibility, the white paper argues, is what allows organizations to expand use while maintaining confidence in how systems behave.

AI agents can reduce delays across healthcare workflows and connect fragmented information, but they only work safely in healthcare when there are humans in the loop and appropriate guardrails. This is why Local AI is especially relevant for agentic workflows that touch sensitive data, specialized applications, and expert review: it helps keep models, tools, logs, and outputs closer to the teams responsible for validating them. To operate safely and in compliance with HIPAA, agentic systems need an architecture that supports event logging, auditability, and the ability to reverse or correct actions when needed.

What This Means for Decision Makers

The practical implications span the whole leadership table. For clinical leaders, faster local processing means quicker insights at the bedside and less administrative burden for care teams. For operations leaders, predictable infrastructure reduces the friction of scaling. For the CFO, a clear break-even timeline makes the business case defensible.

Real-world examples illustrate what’s possible when the architecture, governance, and use case are aligned. Fulgent Genetics achieved a 30% increase in pathology processing speed using an on-premises AI infrastructure, reducing test turnaround times from two to three days to under one day. The Guthrie Clinic deployed an on-premises AI solution in its call center to help teams track patient needs in real time, contributing to the organization’s ability to accept 85% of transfers from other hospitals and clinics. And researchers at the Hopp Children’s Cancer Center Heidelberg (KiTZ) reduced tumor analysis time by 60%, scaling diagnostics from days to hours using an on-premises AI infrastructure.

None of these outcomes required moving sensitive clinical data to a public cloud. In each case, the AI ran close to where the data already lived.

When Every Second Counts offers a five-step framework for healthcare organizations ready to move from experimentation to scale:

Start with a use case that has visible value. Focus first on a use case with clear clinical or operational relevance to build momentum and support for broader adoption. Imaging workflows, patient access, and clinician efficiency are strong starting points for many provider organizations.

Understand where your most important data lives. Healthcare data is often spread across clinical, operational, and imaging systems. Before scaling AI, organizations need to know which data sources they need most, how accessible they are, and whether the right deployment model is at the edge, on premises, in the cloud, or across all three.

Build governance in from the beginning. Trust can’t be an afterthought. As AI becomes more embedded in workflows, organizations need observability, security, and clear control over how systems access data and act on information.

Think platform, not point solution. A practical platform strategy should allow teams to begin with responsive local AI systems for development, validation, inference, visualization, or expert review, then extend successful workflows into governed enterprise infrastructure as adoption grows. Organizations that make the most progress build on a foundation that can support multiple use cases over time, rather than solving one problem at a time with isolated tools.

Scale what’s working for early adopters. Identify the teams and users already driving results, learn from their workflows, and use those lessons to expand success across the enterprise.

From a technical perspective, this means that healthcare IT teams need to: define workload placement before scaling, build a reusable infrastructure core, and prioritize observability and repeatable deployment patterns so growth doesn’t create unnecessary fragmentation.

Avoid the common traps: launching pilots without owners, ignoring where data resides, treating security as an add-on rather than a design principle, and underestimating the operational burden of sustained inference.

Turning Readiness into Action

Healthcare organizations don’t have to choose between innovation and security. With predictable costs, on-premises data control, and governance built in from the start, they can address both barriers at once and scale AI responsibly, starting now.

The organizations that succeed will pair the right infrastructure with disciplined governance and a clear-eyed view of return on investment. That combination turns AI from a promising experiment into a dependable capability that serves patients, clinicians, and communities.

For leaders ready to go deeper, the published resources referenced throughout this article offer practical frameworks, architecture guidance, and validated real-world data on cost and compliance outcomes. A focused pilot in imaging, clinical documentation, or operational workflow automation, grounded in workload placement, governance, and human oversight, is a sound next step.

1 Modern Enterprise Readiness Study

2 Scaling On-Premises Agentic Workloads from Desktop to Datacenter with the Dell AI Factory with NVIDIA