When AI Products Become an AI Portfolio: The AI Governance Shift Needed in Physician Enterprises

The early adoption phase of artificial intelligence (AI) has largely passed. Medical groups that once debated whether to invest in AI are now managing multiple tools that touch scheduling, clinical documentation, point-of-care decisions and revenue cycle.

It's easy for the excitement of efficiency gains and data-driven decisions to perpetuate a cycle of launch, measure and repeat. But even as setting baselines and measuring the impact of individual AI tools becomes standard practice, many medical groups are missing an equally important step: evaluating those tools as a portfolio. Health systems that measure AI's impact tool-by-tool may systematically understate its value, miss the returns that build between tools, and may risk letting one tool's failure overshadow the successes of others.

The governance gap doesn't wait for scale to appear. It can appear the moment two tools touch the same workflow. To help build a sustainable, cost-effective AI program that drives validated performance improvement, physician enterprises need a new governance operating model that takes a portfolio-wide approach from the start.

What an AI Governance Gap Looks Like

Consider three illustrative scenarios that show how governance gaps can emerge when organizations implement more than one AI-enabled tool. Each show how the absence of a portfolio view impacts an organization's ability to truly measure AI costs and performance.

Misattributed or Unattributed Value

A medical group deploys ambient AI scribes to reduce after-hours physician workload. Post-implementation, physicians report significantly less “pajama time.” Later, the revenue cycle team, running an AI-assisted coding pilot of its own, reports that coding accuracy has improved. What the revenue cycle team doesn’t see is that the ambient scribe is producing more complete documentation, giving the coding AI better inputs. The scribe vendor takes credit for the time savings. The coding AI vendor takes credit for the accuracy gain.

A governance process that evaluates the scribe and the coding tool in isolation will miss how one tool's output improves the other's performance. Only a portfolio-wide view of benchmarks and returns makes that gain visible.

Duplicative Costs

During a budget review, the chief financial officer (CFO) of the same health system asks for a cost-benefit analysis of every AI tool in use. The evaluation identifies several AI tools being used by multiple departments, with fees totaling more than $1 million annually. Only a handful have proven results. Each was approved individually based on its own isolated business case. No one is accountable for measuring the enterprise-wide cost.

Multiple licensing fees, data pipelines and dedicated FTEs end up buried in departmental budgets, invisible to anyone evaluating the total cost of the AI investment. The organization ends up carrying more costs than it realizes, missing the chance to share resources across departments.

Poor Sequencing

A health system CEO hears success stories from peers using AI-enabled tools in clinical workflows and asks the chief medical information officer (CMIO) to explore a pilot. For purposes of illustration, assume the organization selects a clinician-facing diagnostic support tool for emergency medicine physicians before it has fully tested its AI governance. Early in the pilot, the tool produces a recommendation that conflicts with the care team’s clinical judgment. Although the scenario is hypothetical and does not describe any specific vendor, product or implementation, it illustrates the governance risk. Without clear monitoring, escalation and pause criteria, a single concern in a higher-risk clinical workflow can trigger broader organizational hesitation around AI, including tools with lower-risk administrative or revenue cycle use cases that may be performing as expected.

Clinical AI isn’t off-limits early in a program’s maturity, but clinical, clinical-adjacent, and non-clinical tools carry different risk portfolios and demand different levels of governance intensity. A practical starting point is to limit the portfolio to one clinical AI tool onboarding at a time until the governance process for monitoring performance, and escalating concerns is tested and working.

Principles for Early AI Governance

AI governance doesn’t need to be built from scratch. Health systems already govern procurement, vendor evaluation, security and compliance for enterprise technology. Much of that infrastructure extends naturally to AI.

Where AI governance diverges from IT governance is in a narrower set of dimensions than most frameworks suggest. Data governance changes when AI models train on, learn from and generate clinical content. Model performance can degrade in ways that traditional software doesn’t, which means ongoing monitoring isn’t optional in the way it is for a static application. Privacy and business associate agreement (BAA) considerations specific to large language models are new territory for most compliance teams. These are the dimensions worth building new governance muscle around.

Effective AI governance calls for more than documentation and approval processes. It requires ownership, benchmarking, measurement and a portfolio-wide view of cost and return. For organizations just beginning to build this discipline, four questions applied consistently to every deployed tool are a practical starting point.

To set the foundation for an AI governance model, health leaders should look at each tool and ask:

• What outcome was this tool supposed to produce, in measurable terms?
• Who owns that outcome, by name and role?
• What data are we collecting to evaluate whether it's being achieved?
• What happens if the tool isn't working, and who makes the call?

These four questions do double duty. Applied to a new tool before it's approved, they force a business case that names an outcome and an owner. Applied to a tool already in production, they become a checkpoint: Is the tool still working and if not, who decides what to do next?

That second application matters as much as the first. An older AI tool can erode over time as workflows, patient populations or data quality shift underneath it. The ambient scribe that justified its cost two years ago, deployed against a different patient mix and a different EHR configuration, may not justify its contract renewal today. Reviewing existing tools with the same rigor as new ones turns governance from a checklist into an operating discipline.

Applying a Cost Center Model

In many organizations, each department carries its AI tool costs in its own budget. That structure creates blind spots. Creating a unified P&L allows a CFO to understand which costs, resources and results can be shared versus department-specific.

An AI program P&L organizes around two sides:

On the AI Program Costs Side

• Tool licensing and subscription fees.
• Shared infrastructure such as data pipelines, integration middleware and monitoring tools that serve multiple AI applications.
• Shared talent such as a clinical AI product owner, data engineers and/or an analytics team supporting the full AI portfolio.
• AI rollout training hours.
• Experimentation capacity inclusive of market surveillance time, pilot infrastructure and transition costs.

These are often built from scratch for each new tool when no one is looking across the portfolio; consolidating them into the program P&L makes the shared investment visible, increases consistency of fundamental capabilities and prevents duplication.

On the AI Program Returns Side

This should include all operational, clinical or financial metrics associated with each AI tool in the active portfolio. Using the scenarios above, metrics might include:

• Revenue uplift from productivity recapture (e.g., wRVU gains from ambient scribes freeing documentation time that converts to patient volume).
• Coding accuracy improvements.
• Denial reduction and prior authorization savings.
• Recovered appointment slots from scheduling AI, gained through reduced no-shows and improved fill rates.
• Interaction effects between tools.

What Changes When You Add It Up

Adding up costs and returns at the program level does three things no individual tool scorecard can: it surfaces shared costs that were hidden in departmental budgets, corrects returns that were misattributed to the wrong tool and produces a net position for the AI program as a whole. This shift is what a CFO sees when AI spend moves from a set of scattered line items to one program P&L.

The table below shows what changes.

Recommended Ownership and Operating Cadence

The AI program P&L should be owned or co-owned by the CFO's office and whoever holds enterprise AI accountability. Titles may vary across organizations, but in many cases, the ideal co-owner is a chief health AI officer, CMIO, chief transformation officer or chief information officer with a cross-functional mandate. Department leaders who own individual tools contribute performance data, but the portfolio-level view exists at the level where all that data converges and should be owned by someone responsible for that convergence.

In terms of cadence, a quarterly review suffices for most organizations. This aligns with how most health systems already run financial performance reviews, making the model easier to embed into existing operations. The annual budget cycle is where the portfolio view has the most leverage: Instead of each department submitting its own AI business case independently, the program P&L informs a single, portfolio-level investment decision.

How Premier Helps Build AI Programs Rooted in Verified Evidence

Health leaders don't need to be convinced that AI governance is important. Most are already applying the same operational rigor used for other enterprise technology to AI tools, including managing procurement, vendor evaluation, security and compliance.

Physician enterprise leaders increasingly need evidence:

• How are other health systems using AI?
• Is it producing returns?
• How are leaders managing AI resources, costs and measurement at scale?

AI costs are expanding rapidly across licenses, subscriptions, tokens, infrastructure and talent. Some organizations are making portfolio decisions based on vendor claims or peer anecdotes, because they don't have access to a cross-system evidence base showing which categories of AI tools produce measurable results.

Premier helps physician enterprises build critical evidence through two connected offerings.

By joining the AI Value & Governance Collaborative, members can see real performance data and use cases from other health systems. Similar AI implementations get benchmarked against peers, allowing health systems to replace vendor assumptions with real, cross-system evidence of what's working. Together, members contribute to a growing body of evidence around where AI is producing value and how much.

Collaborative members also get:

• A peer cohort with regular programming, peer benchmarking and an annual national meeting.
• An expert-validated governance maturity assessment and scored governance profile for a board-level audience.
• Benchmarking data from Premier's proprietary data set, supplemented by an organization's own data for Performance Insights Value Optimization Tool (PIVOT) members.
• Governance frameworks, templates and AI tool evaluation support.
• Quarterly advisory sessions and monthly office hours with Premier's AI advisory team.
• A 12-month reassessment to measure whether governance maturity improved.

Premier's AI Realization Advisory Practice works alongside the collaborative, helping member organizations put these frameworks into practice: from applying the four questions for every tool in the portfolio, to building the program P&L, to standing up the operating cadence that keeps the whole system accountable.

If your AI program is rapidly expanding, it may be time to consider a more sophisticated governance model. Find out how Premier can help your physician enterprise build the portfolio discipline that makes AI investment sustainable.