AI should be managed as an enterprise investment portfolio—not as a collection of technology projects. The advantage will accrue to organizations that choose well, redesign workflows, measure honestly, and scale selectively.
Executive Summary
Healthcare organizations are moving rapidly from AI curiosity to active deployment, but visible activity is not the same as enterprise value. In 2024, 71% of U.S. hospitals reported using predictive AI integrated with the electronic health record, and 31.5% reported generative AI integrated with the EHR. Yet the harder work begins after adoption: selecting material problems, redesigning workflows, validating performance, converting productivity into measurable benefit, and deciding which pilots merit scale.
This article presents a six-dimension investment framework, a three-part portfolio architecture, and a stage-gate process for healthcare leadership teams. The central argument is straightforward: AI should be governed as an enterprise investment portfolio. A smaller set of scalable initiatives tied to strategic priorities will usually create more durable value than a larger collection of disconnected pilots.
The Distance Between AI Activity and AI Value
Healthcare's AI conversation has changed. The question is no longer whether organizations will experiment with artificial intelligence, but whether those experiments will improve care, expand capacity, reduce friction, strengthen economics, or create a capability that matters strategically.
The activity is visible: vendor demonstrations, innovation committees, ambient documentation pilots, predictive models, coding tools, patient-message drafting, scheduling applications, and a growing inventory of AI embedded within enterprise software. National evidence confirms that adoption is no longer marginal. In 2024, 71% of hospitals reported predictive AI integrated with their EHR, up from 66% in 2023. Nearly one-third reported using generative AI integrated with the EHR, with another quarter planning adoption within a year.
Those figures demonstrate momentum, not value. An organization can deploy multiple tools and still fail to improve a material outcome. It can save minutes in a workflow without converting that time into capacity, access, retention, or financial benefit. It can complete a successful pilot that cannot be integrated, governed, funded, or adopted at scale. It can also create new clinical, privacy, cybersecurity, reputational, and operating risks faster than it creates institutional capability.
The central executive challenge is therefore one of selection and execution: which opportunities deserve scarce capital and leadership attention, what evidence is sufficient to proceed, and how should the organization distinguish a promising demonstration from a scalable enterprise investment?
AI Is an Enterprise Strategy Question
AI is often introduced through technology, digital, analytics, or innovation functions. Those teams are essential, but the consequences of adoption extend well beyond them. AI decisions affect clinical practice, patient experience, workforce design, data infrastructure, operating processes, growth, capital allocation, compliance, and institutional trust. They are enterprise decisions delivered through technology.
That distinction changes leadership accountability. The board should understand material strategic and patient risks without approving individual applications. Executive leadership should set investment priorities, risk appetite, and performance expectations. Clinical leaders should define acceptable use and validate consequences for care. Operations must own workflow redesign and adoption. Finance and strategy must test whether claimed value is measurable and realizable. Technology and data leaders must assess integration, reliability, security, and lifecycle requirements. Legal, compliance, privacy, and risk functions should shape controls proportionate to the use case.
Federal policy is also moving toward lifecycle transparency and monitoring. The HTI-1 final rule established transparency requirements for predictive algorithms within certified health IT, and the FDA's 2025 draft guidance for AI-enabled device software emphasized total-product-lifecycle risk management, including postmarket performance monitoring. Even when a specific tool falls outside those requirements, the direction is instructive: evaluation cannot end at procurement or launch.
A Six-Dimension AI Investment Framework
Every proposed use case should pass through a consistent enterprise screen. The purpose is not to create a perfect score or force unlike investments into one ROI formula. It is to expose assumptions, make tradeoffs explicit, and prevent enthusiasm in one dimension from obscuring weakness in another.
| Dimension | What it tests | Diagnostic questions | Credible evidence |
|---|---|---|---|
| 1. Strategic alignment | Material enterprise priority, clinical need, operating constraint, or strategic capability | What priority does this advance? Why now? What happens if we do nothing? | Named enterprise objective; accountable sponsor; explicit problem statement |
| 2. Clinical, patient & workforce impact | Quality, safety, access, experience, capacity, burden, and new risks | Who benefits? Could any population be harmed? How does human judgment remain involved? | Baseline and target outcomes; user validation; subgroup testing; safety plan |
| 3. Financial & operational value | Revenue, cost, throughput, capacity, cycle time, risk avoidance, or capital efficiency | How does saved time become realized value? What costs sit beyond the license? | Fully loaded cost; benefit owner; scenarios; measurement method |
| 4. Implementation readiness | Data, integration, workflow, leadership, talent, training, and change capacity | Are data fit for purpose? Who owns the redesigned workflow? What dependencies could delay scale? | Readiness assessment; resource plan; integration design; adoption plan |
| 5. Risk & governance | Clinical, regulatory, privacy, security, bias, reliability, transparency, and reputation | What is the consequence of error? How will drift or misuse be detected? What escalation is required? | Risk tier; validation standard; monitoring thresholds; response protocol |
| 6. Scalability & advantage | Extension across locations, service lines, populations, and strategic capabilities | Can it scale economically? Is it differentiating or increasingly commoditized? | Repeatable operating model; unit economics at scale; enterprise roadmap |
A common warning sign is an initiative with an eloquent technology case but no operational owner. Another is a projected productivity benefit that appears in a business case without a mechanism to convert time saved into additional appointments, shorter turnaround, reduced overtime, avoided hiring, improved retention, or another observable result. Technology does not independently redesign an operating model.
Manage AI as Three Strategic Portfolios
A portfolio architecture helps leadership compare investments by strategic purpose while recognizing that different opportunities require different evidence, time horizons, and return expectations.
| Portfolio | Primary objective | Representative use cases | Typical horizon | Principal risks |
|---|---|---|---|---|
| Protect and strengthen the core | Reduce burden, leakage, cost, and risk; improve reliability | Documentation support; coding and denials; prior authorization; compliance; cybersecurity assistance | 0–24 months | Automation error, weak controls, benefits that remain theoretical |
| Improve care delivery | Improve quality, access, capacity, navigation, diagnosis, and experience | Clinical workflow support; imaging; capacity optimization; patient navigation; scheduling; monitoring | 1–3 years | Clinical harm, bias, adoption friction, workflow disruption |
| Create future advantage | Build differentiated capabilities and new models of growth | Precision medicine; research acceleration; personalized engagement; strategic intelligence; new care models | 2–5+ years | Technology uncertainty, long payback, dependency on scarce data and talent |
The first portfolio often contains the clearest near-term economics. The second may create value through access, capacity, quality, and workforce sustainability rather than direct expense reduction. The third may warrant strategic-option logic: modest investment today to develop data, partnerships, talent, or operating capabilities that could become important later. Leadership should resist evaluating all three through a single hurdle rate or a single definition of success.
Portfolio balance also matters. An organization that funds only back-office efficiency may improve near-term economics while failing to build future clinical or market advantage. One that concentrates on frontier applications may overextend implementation capacity and neglect more certain sources of value. The right mix reflects strategy, financial condition, risk appetite, and organizational maturity.
Build the Business Case Around Realized Value
A defensible AI business case begins with a defined problem, not a product. It establishes baseline performance, the target population, affected workflows, expected benefits, fully loaded costs, implementation dependencies, clinical and operational ownership, time to value, risk-adjusted scenarios, and the evidence required for scale.
The software license is rarely the full cost of adoption. The investment may also require data preparation, interfaces, validation, cybersecurity review, workflow redesign, legal analysis, training, change management, ongoing monitoring, model maintenance, vendor oversight, and retirement of redundant tools. These costs do not make the investment unattractive; excluding them makes the business case unreliable.
Benefits deserve equal discipline. A tool that reduces documentation time can create meaningful workforce value, but time saved is not automatically cash saved. The organization must decide how the benefit will be realized: more patient capacity, improved access, reduced after-hours work, lower vacancy pressure, better retention, less overtime, or improved clinician experience. Each pathway has a different owner, measurement method, and financial consequence.
The same principle applies to revenue-cycle and access applications. A forecast of fewer denials must be linked to denial categories, preventable baseline volume, net collection timing, staffing impacts, and sustained performance. A scheduling tool must translate utilization into completed visits, reduced leakage, or improved patient access—not simply more automated contacts.
Use Stage Gates to Move From Pilot to Scale
Pilots should be designed as decision instruments. Their purpose is to generate the evidence leadership needs to scale, modify, hold, or stop an initiative. Predefined gates protect the organization from both premature expansion and endless experimentation.
| Gate | Required evidence | Primary accountability | Decision |
|---|---|---|---|
| 1. Strategic screen | Material problem, sponsor, portfolio fit, preliminary risk tier | Strategy + accountable executive | Advance only if tied to a priority and plausible value |
| 2. Business case | Baseline, full cost, benefit pathway, alternatives, dependencies | Finance, strategy + operations | Fund validation only if assumptions are testable |
| 3. Validation | Technical fit, clinical safety, privacy/security, data suitability | Clinical + technology/risk leaders | Proceed only if minimum performance and controls are met |
| 4. Controlled pilot | Adoption, workflow impact, outcome signal, unintended effects | Operational owner | Scale, modify, hold, or stop against predefined criteria |
| 5. Enterprise deployment | Repeatable workflow, funding, support model, training, unit economics | Executive sponsor + operations | Release capital in phases as sites/users meet readiness criteria |
| 6. Lifecycle review | Outcome durability, drift, incidents, vendor performance, replacement options | Portfolio governance body | Continue, remediate, renegotiate, restrict, or retire |
Stopping an initiative that fails predefined criteria is not a failed innovation program. It is evidence that portfolio governance is functioning. The greater failure is allowing a weak pilot to persist because the organization has invested political capital in it—or allowing a successful pilot to stall because no one planned for integration, operating ownership, and scaled funding.
Measure Enterprise Value, Not Activity
Pilot counts, users, adoption rates, generated outputs, and training hours are useful operating indicators. They are not, by themselves, evidence of value. Leadership needs a balanced scorecard that connects adoption to outcomes.
- Clinical and patient: quality, safety, access, experience, equity, diagnostic or treatment performance.
- Workforce: documentation burden, time after work, retention, engagement, capacity, overtime, vacancy pressure.
- Operational: throughput, cycle time, utilization, turnaround, error rates, rework, reliability.
- Financial: net revenue, collections, labor or vendor cost, avoided hiring, margin, capital efficiency, risk avoidance.
- Adoption and execution: appropriate use, workflow compliance, training, support burden, site-level readiness.
- Risk: performance by population, model drift, overrides, incidents, privacy/security events, escalation response.
- Strategic capability: scalability, reusable data and infrastructure, speed to launch future use cases, differentiation.
The scorecard should identify the baseline, target, data source, measurement frequency, accountable owner, and threshold for intervention. Measures should also reflect the use case's risk. A low-risk administrative drafting tool does not require the same clinical validation as an algorithm influencing diagnosis or treatment, but both require clarity about acceptable performance and human oversight.
The national data illustrate why this matters. Among hospitals using predictive AI in 2024, 82% reported evaluating accuracy, 74% bias, and 79% post-implementation evaluation or monitoring. Those are encouraging figures, but the federal analysis noted that fewer hospitals applied evaluation to all or most models. Enterprise oversight must reach the full portfolio, including AI acquired indirectly through existing platforms.
Governance Should Enable Disciplined Scaling
Poor governance can become a slow committee that reviews every minor decision. Effective governance does the opposite: it establishes clear risk tiers, decision rights, evidence standards, and escalation paths so appropriate initiatives move faster while consequential ones receive deeper review.
At the enterprise level, governance should include executive sponsorship; clinical, technology, data, finance, strategy, and operational leadership; legal, compliance, privacy, cybersecurity, and risk expertise; and meaningful frontline-user input. It should maintain an inventory of deployed and contracted AI, assign accountable owners, set approval thresholds, and conduct recurring portfolio review.
Individual implementation teams should remain responsible for workflow, adoption, training, support, and results. Enterprise governance sets policy and allocates capital; it should not substitute for operating management. This separation is particularly important as AI becomes embedded in products purchased through multiple departments and vendor contracts.
Governance must also continue after launch. Local performance can change as patient populations, workflows, coding conventions, underlying models, or vendor products evolve. Ongoing monitoring is therefore both a risk requirement and an investment discipline: leadership should know whether expected value is holding, deteriorating, or expanding.
Five Actions for the Next 90 Days
- Inventory the portfolio. Identify active pilots, embedded AI capabilities, contracts, owners, costs, populations affected, and renewal dates.
- Map investments to enterprise priorities. Classify each initiative by strategic portfolio and identify those with no clear objective or accountable sponsor.
- Adopt common investment and risk criteria. Use the six dimensions to screen new proposals and reassess material existing deployments.
- Define stage gates and outcome measures. Establish baseline performance, evidence thresholds, decision rights, and explicit scale or stop criteria.
- Create a recurring portfolio review. Reallocate capital and implementation capacity toward initiatives demonstrating the strongest combination of value, readiness, safety, and scalability.
The Institutional Capability Will Matter Most
Access to AI will become increasingly broad. Many tools will be incorporated into core platforms, and some capabilities that appear differentiated today will become standard. Sustainable advantage will therefore come less from possessing a particular tool and more from the institutional ability to identify consequential problems, evaluate alternatives, implement responsibly, redesign workflows, measure results, and scale what works.
Healthcare organizations will not create durable value by pursuing the greatest number of pilots or acquiring the newest technology. They will create it by making better choices and executing them with discipline. A smaller portfolio of scalable initiatives—connected to enterprise priorities and supported by accountable operating owners—can create more value than a larger collection of demonstrations.
The next phase of healthcare AI will be defined by this capability: choosing well, implementing effectively, measuring honestly, managing risk, and scaling selectively.
Sources and Evidence Notes
- ASTP/ONC, “Hospital Trends in the Use, Evaluation, and Governance of Predictive AI, 2023–2024,” September 2025.
- ASTP/ONC, “HTI-1 Final Rule,” published January 9, 2024.
- FDA, “FDA Issues Comprehensive Draft Guidance for Developers of Artificial Intelligence-Enabled Medical Devices,” January 6, 2025.
- JAMA Network Open, “Uptake of Generative AI Integrated With Electronic Health Records in US Hospitals,” 2025.
- American Hospital Association, “How to Build and Implement Your AI Health Care Action Plan,” January 14, 2025.
- American Hospital Association, “Avoiding Missteps: 3 Keys to Governing, Scaling and Deploying AI Responsibly,” July 21, 2026.
- NEJM Catalyst, “Ambient Artificial Intelligence Scribes to Alleviate the Burden of Clinical Documentation,” March 2024.