Measuring the Ghost

AI Oversight From the Board Seat

If the return on an AI investment is easy to measure, the investment may not be ambitious enough. Boards need measures that fit the asset, and the discipline to hold management to them.

In Brief

The Problem. Payback periods and internal rates of return were built for predictable assets. Applied to AI, they steer boards toward small, certain savings and away from larger, uncertain gains in care and research.

Why It Persists. Administrative savings are easy to count; clinical and strategic value is not. "Strategic value" can also become an excuse for projects never held to account.

What Boards Should Do. Ask for sector-specific measures and a strategic scorecard alongside net present value, agree the targets and dates before approval, and review the same short report on every material AI investment.

How much is a good clinical decision worth? What is the net present value of a crisis that never happened? What is the value of an organization that patients trust because its AI is fair? These questions trouble chief financial officers. They are also the questions boards must ask about the return on investment from artificial intelligence (AI), and a clear answer is proving elusive.

The evidence so far points in two directions. In August 2025, researchers at the NANDA initiative of the Massachusetts Institute of Technology reported that about 95 percent of organizations in their study had gotten zero return from their AI efforts (Claburn, 2025). A year later, healthcare buyers surveyed by Bessemer Venture Partners and Bain and Company reported an average return of 3.5 times their spend (Guerra & Kraus, 2026). Both findings can be true. The difference lies in what was measured, and how.

My argument is that the measurement problem is a governance problem. The board does not need a better formula. It needs to agree, before it approves an investment, what will count as success and when it will look.

Why Standard Measures Mislead

Standard financial yardsticks, such as payback period and internal rate of return, were designed for predictable assets like equipment and buildings. AI is a probabilistic system whose value is often indirect, strategic, and slow to appear. Investing in AI is less like buying a vehicle and more like funding a research department. Its value is real, but no one can predict the specific discovery it will produce next year, or how that discovery might change the organization's model of care or business.

Ambient AI scribes show the problem. In March 2025, the Peterson Health Technology Institute assessed ambient AI scribes with eight health systems. It found solid evidence that the tools reduce clinician burnout; Mass General Brigham, for example, reported a 40 percent relative reduction in a six-week pilot. It also found that the scribes had not yet proven a financial return, and that health systems lacked standard measures to decide whether to scale them (Beavins, 2025). A board using payback alone would struggle to approve a tool whose main benefit is keeping clinicians in practice.

Directors know the information is thin. In a 2026 PwC survey of about 600 public company directors, 59 percent rated the information they receive on AI investments and returns as fair or poor; only 25 percent rated it good or excellent (Gaskin, 2026).

The Bias Toward the Certain

A narrow focus on direct, easily quantified returns can lead to poor strategic choices. A board might approve an AI scheduling assistant because it can calculate the exact savings in administrative cost. The same board might decline a more ambitious model that predicts readmission risk, because the savings, though large, are uncertain and difficult to model.

The stakes on the second project are not small. In fiscal year 2027, Medicare's Hospital Readmissions Reduction Program penalizes 2,334 hospitals, about 80 percent of those evaluated, with payment reductions of up to three percent (Pifer Parduhn, 2026). The value of a model that prevents readmissions is real; it is simply harder to put in a cell.

The 2026 Bessemer data show the tilt. Provider revenue cycle uses reported returns of 4.0 times spend, while provider clinical uses reported 2.9 times (Guerra & Kraus, 2026). The figures are self-reported, and administrative gains are easier to count. Repeated over several budget cycles, the portfolio fills with low-value projects that are easy to justify, while strategic initiatives that are harder to measure are starved of capital.

Measures That Fit the Sector

The first correction is to move beyond generic financial models to key performance indicators (KPIs) tied to the organization's core work. For a pharmaceutical company, the measure may be the acceleration of clinical trial timelines, such as the time to enroll a trial or the cost of a submission. For a hospital system, it might be a reduction in readmission rates, length of stay, or claim denials, or an improvement in bed utilization. Measures of this kind tie AI investment to outcomes the board already oversees. Many organizations are not yet equipped to do this: the Healthcare Financial Management Association (HFMA) reported in 2025 that most health systems lack the resources to identify where AI would deliver a return (Healthcare Financial Management Association, 2025). The board can ask management to build that capability.

A Strategic Scorecard

More important than refining old measures is the ability to assess strategic value. Does the AI investment create a durable advantage? Does it raise barriers for competitors? Does it build a proprietary data asset that grows more valuable over time? A board pack should contain not only a net present value calculation but also a strategic scorecard that rates each investment on criteria such as these:

• Building a proprietary data asset: whether the investment creates data the organization owns and can use again;
• Strengthening the competitive position: whether it creates an advantage that rivals cannot easily copy; and
• Improving decision speed: whether it shortens the time from information to a sound decision.

The Objection: "Strategic Value" Becomes an Excuse

The strongest objection to a scorecard is that it gives weak projects a place to hide. A project that never pays back can always be described as building capability, and skeptical finance committees are right to worry about that.

The answer is not to abandon the scorecard but to discipline it. Before approval, the board should require three things:

• the measures, financial and strategic, that will show whether the investment is working;
• the date by which each measure should move, and by how much; and
• the committee that will receive the results and decide whether to continue.

A strategic claim with a target and a date can be tested. One without them is a hope.

What the Report Should Show

Once the measures are agreed, a short report for each material AI investment, delivered to the responsible committee on a fixed schedule, is usually enough. It should show:

• the measure agreed at approval and its starting value;
• the current value and the target, with the date each should be reached;
• the cost to date, including monitoring and vendor fees, not only the cost to build; and
• management's recommendation to continue, adjust, or stop.

A board that sees the same four lines each quarter can compare one investment with another and see when a project has stopped moving.

Accountability Is the Point

Measurement is a governance tool, not only a finance exercise. When the board agrees on the measures before it approves an investment, and records them in the minutes, it creates the basis for later oversight. In my years as a chief executive, the measures that mattered most were the ones agreed before the money was spent. Measures chosen afterward tend to flatter the decision.

Organizations bound to a conventional spreadsheet optimize for the projects that are easiest to defend. For some investments, the most important returns are the hardest to count. That is a reason to define in advance how the board will judge them, not a reason to stop asking.

For the Board

• Which sector-specific measures will show whether each material AI investment is working?
• Are we favoring easily measured savings over uncertain but larger gains, such as avoided readmissions, and is that a deliberate choice?
• For each investment, what did we agree to measure, by when, and who receives the results?
• Does our board pack include a strategic scorecard alongside net present value, and does every strategic claim carry a target and a date?
• How good is the information we receive on AI returns, and what would make it better?
• What would cause us to stop funding an AI initiative?

Find the Finance Committee question set in the Avancer AI Governance Navigator.

References

Beavins, E. (2025, March 25). Early evaluation of AI scribes finds decreased burnout but limited financial ROI. Fierce Healthcare. https://www.fiercehealthcare.com/ai-and-machine-learning/early-evaluation-ai-scribes-finds-decreased-burnout-limited-financial-roi

Claburn, T. (2025, August 18). GenAI FOMO has spurred businesses to light nearly $40 billion on fire. The Register. https://www.theregister.com/2025/08/18/generative_ai_zero_return_95_percent

Gaskin, J. L. (2026, September 23). Boards say they need more from management on AI. Corporate Compliance Insights. https://www.corporatecomplianceinsights.com/boards-say-they-need-more-from-management-ai/

Guerra, S., & Kraus, S. (2026, September 30). The 2026 healthcare AI ROI scorecard. Bessemer Venture Partners. https://www.bvp.com/atlas/the-2026-healthcare-ai-roi-scorecard

Healthcare Financial Management Association. (2025, August 12). Health system adoption of AI outpaces internal governance and strategy [Press release]. https://www.hfma.org/health-system-adoption-of-ai-outpaces-internal-governance-and-strategy/

Pifer Parduhn, R. (2026, October 8). Medicare penalizes more hospitals for readmissions. Healthcare Dive. https://www.healthcaredive.com/news/hospital-readmission-reduction-program-penalties-rise-2027-cms/832473/