Deciding on AI With Imperfect Information

Updated October 21, 2026. First published November 14, 2023.

Boards and chief executives in healthcare and life sciences rarely decide with complete information. The failure is not the missing data; it is using one decision method for four very different kinds of gaps.

In Brief

The Problem. High-stakes "go" or "no-go" decisions, from starting a research program to approving an artificial intelligence (AI) system for clinical care, reach the board with gaps in the evidence. Waiting for complete information is rarely possible, and delay has its own cost.

Why It Persists. Boards tend to treat every gap the same way, usually by asking for more data. More data helps with only one of the four kinds of gaps that volatility, uncertainty, complexity, and ambiguity create, and the four conditions compound one another.

What Boards Should Do. Name the kind of gap in front of the board, ask management which method it used to address that gap and what the method found, attach conditions to the approval, and leave a record that shows the board did so.

In late 2023 I wrote two articles on how healthcare and life sciences boards make decisions with imperfect information. One described the environment; the other proposed decision frameworks. This revision combines them, corrects them, and updates the evidence. The argument has not changed. Leaders cannot wait for perfect information that may never arrive, and the ability to act well with imperfect data is a core competency for boards and chief executives in this sector.

What has changed is the decision that tests the argument most. Whether to approve an AI system for clinical care, claims, research, or regulatory work now reaches the board with less complete evidence than almost any other decision it makes. The vendor validated the model on another organization's patients. The model may change after it goes live. The law differs by state and by country, and it is still moving.

My thesis is this: the board's task is not to eliminate imperfect information, which is impossible, but to recognize which kind of imperfection it faces and to match its scrutiny to that kind. The counterintuitive part is that the board's instinctive request, "bring us more data," is the right answer to only one of the four kinds of gaps. For the other three, more data delays the decision without improving it.

Why Waiting for Perfect Information Is Not an Option

Imperfect or incomplete information describes situations in which decision-makers do not have all the data needed to be fully informed, or in which the data they have are uncertain. It is the normal condition in life sciences because of emerging diseases, rapid technological change, and unpredictable market responses. When a board decides whether to fund a new research and development (R&D) program, it faces incomplete information about scientific feasibility, market potential, and resource requirements, all at once.

The primary risk is a poor decision built on skewed, incomplete, or misunderstood data. In healthcare and life sciences the stakes are high because these decisions affect patients, regulatory approvals, and large investments. The odds are also unforgiving. Across 2011 to 2020, only 7.9 percent of drug candidates entering Phase I reached approval, and only 57.8 percent of candidates in Phase III went on to a New Drug Application (NDA) or Biologics License Application (BLA) (Thomas et al., 2021). The estimated median capitalized R&D investment to bring a new medicine to market, including the cost of failures, was about USD 985 million (mean USD 1.3 billion) for products approved from 2009 to 2018 (Wouters et al., 2020).

Postponing a decision until the information is complete is seldom feasible. In my first version I wrote that delay can lead to patent cliffs. That was imprecise. A United States patent term ends 20 years from the filing date of the application (35 U.S.C. § 154(a)(2)), so the clock runs whether or not the company decides. Each month of delay before launch is a month less of protected sales. Delay also means missed opportunities in competitive markets, slow responses to public health needs, and resources tied up in programs that should have stopped. This matters most for time-sensitive decisions, such as regulatory submissions and commercial launches.

The practical skill, then, is to assess the information quickly: what is critical, what is uncertain, and what is unknown. That assessment tells the board where further data collection is worth the time, how reliable each source is, and how much uncertainty it can accept. In portfolio prioritization, for example, directors weigh market potential, risk profiles, and resource requirements that are each known to a different degree. Treating them as equally reliable is itself a decision error.

Four Conditions That Compound One Another

The acronym VUCA (volatility, uncertainty, complexity, and ambiguity) is often used loosely, as a single word for "difficult." Bennett and Lemoine (2014) showed that the four terms describe distinct situations that call for distinct responses. I adopt their definitions here, and I correct my 2023 version, which equated volatility with "known unknowns" and uncertainty with "unknown unknowns." Their definitions are more precise and more useful in the boardroom:

• Volatility is change that is frequent and sometimes unpredictable, but understandable; information is available.
• Uncertainty exists when cause and effect are understood, but it is unknown whether a given event will produce meaningful change.
• Complexity describes many interconnected parts, an elaborate network of information and procedures, with or without change.
• Ambiguity exists when cause and effect are unclear and there is no precedent; leaders do not yet know the rules of the game.

In life sciences these conditions do not merely coexist; they amplify one another. A regulatory change (volatility) raises questions about a trial's outcome (uncertainty), which ripple through partners and suppliers (complexity), and force choices for which no rule exists (ambiguity). Even minor incidents can set off chain reactions. The examples below, most of them drawn from my 2023 articles and several now matched to events since, show how each condition appears in practice.

Volatility: Frequent Change That Can Be Understood

Breakthrough disruptions. A single approval can reset a treatment area. On December 8, 2023, the U.S. Food and Drug Administration (FDA) approved Casgevy, the first FDA-approved therapy to use CRISPR/Cas9 gene editing, and Lyfgenia, a second gene therapy, for sickle cell disease (FDA, 2023). When an approval like that arrives, competitors may need to reconsider years of work on less effective approaches, investors shift priorities, and the regulatory path for similar products changes.

Evolving regulatory frameworks. Changes in FDA guidance, payer reimbursement policy, or manufacturing rules can send companies scrambling. A more demanding trial protocol can delay approval, and a shift in what counts as sufficient evidence can require costly additional studies. AI regulation shows the pattern clearly. Colorado repealed and replaced its 2024 AI Act on May 14, 2026; the new law takes effect on January 1, 2027, replaces annual impact assessments with notice and disclosure duties, and exempts HIPAA-covered entities from many obligations, but not when they use AI for employment decisions (Darch et al., 2026). In Europe, Regulation (EU) 2026/1744 entered into force on July 27, 2026, and moved the main high-risk AI obligations to December 2, 2027, and, for AI embedded in regulated products such as medical devices, to August 2, 2028 (Hunton Andrews Kurth, 2026).

Domino effects. Volatility creates chain reactions. A regulatory shift in one jurisdiction can change clinical trial design, data collection, and budgets worldwide. My 2023 example was a European Union decision to tighten privacy standards for patient health information, forcing multinational companies to rethink how they collect and use trial data. The same pattern now applies to AI: a European rule on high-risk AI reaches any company that sells into Europe, and a vendor's model trained on data from several countries carries those countries' rules with it.

Supply chain shocks. A geopolitical conflict, a new tariff scheme, or a localized disaster can stop production. A quality failure at a single plant can do the same. In November 2022, FDA inspectors found data integrity problems and weak contamination controls at an Intas Pharmaceuticals plant in Gujarat, India, a key chemotherapy supplier; FDA then issued an import alert and allowed only certain cancer medicines in short supply to continue to the United States (Silverman, 2023). One plant affected patients far beyond the company that owned it.

Market dynamics. A competitor's lead candidate fails its Phase III trial. The market opportunity for rivals grows, but investors and regulators may also look harder at every product that shares the same mechanism of action. Good news and bad news arrive in the same event.

The response Bennett and Lemoine recommend for volatility is agility: build slack, such as extra inventory, a second supplier, or hedged costs, so the organization can respond when change arrives. It costs money in advance. The board's role is to decide how much of that cost is worth carrying.

Uncertainty: Known Causes, Unknown Outcomes

The stakes of clinical trials. A company must decide whether to advance a promising candidate into an expensive Phase III trial. Even with positive early data, the question remains whether it will work in a large population and whether unforeseen side effects will emerge. Delay burns cash without progress. Proceeding risks a multimillion-dollar failure and a setback to the whole program. The success rates cited above show that both outcomes are common.

Reimbursement. A company developing a therapy for a chronic disease cannot know whether payers will cover it broadly or restrict it. That uncertainty shapes market entry, pricing, and even the design of the trials that will produce the evidence payers review.

Long-term safety. Rare but serious adverse events may surface only after years of use. In April 2024, FDA required a boxed warning on all six approved BCMA-directed and CD19-directed autologous CAR T cell therapies for the risk of T cell malignancies, and stated that patients and trial participants should be monitored for secondary cancers for life (FDA, 2024). Companies face a choice between rigorous post-marketing surveillance, which costs money and may find problems, and a lighter approach that risks greater liability and reputational harm later.

The global factor. Differences in regulation across countries, and the ethical questions that arise when trials run in low-income settings, add further uncertainty. They call for scenario planning that looks beyond a single market.

An AI example. Uncertainty is the defining gap in most AI approvals. When Michigan Medicine externally validated a widely implemented proprietary sepsis prediction model across 38,455 hospitalizations, it found an area under the curve of 0.63, compared with the 0.76 to 0.83 reported in the developer's internal documentation, and the model missed 67 percent of sepsis cases at the alert threshold studied (Wong et al., 2021). Cause and effect were understood. The outcome in that population was not.

The risk of risk aversion. Uncertainty can also breed excessive caution, and caution has costs that do not appear on the agenda. Unconventional ideas, such as a novel delivery mechanism, an unusual partnership structure, or a therapy for a small patient subpopulation, are shelved early as "too risky." Companies become followers, waiting for a competitor or a regulator to resolve the uncertainty, and then scramble to catch up. A culture that punishes every failed bet teaches researchers to stop proposing bold ones, and the company's ability to attract talent suffers over time.

Bennett and Lemoine's response to uncertainty is information: collect it, interpret it, and share it, including through new partners and data sources. This is the one condition in which the board's request for more data is exactly right.

Complexity: Many Interacting Parts

The human body. Biological systems contain countless feedback loops, interconnected pathways, and individual variations. A drug designed for one mechanism can affect blood pressure or metabolism. Finding the root cause of a problem in a clinical trial can resemble untangling a knot with no clear beginning or end.

The global supply chain. Raw materials, components, manufacturing, packaging, and distribution span many suppliers and countries. A contamination problem at one plant, a shipping delay caused by new trade rules, or a shortage of specialized packaging can disrupt production and delay launch.

Many stakeholders. Drug development involves researchers, clinicians, regulators, payers, patient advocacy groups, and others, each with its own priorities and definition of success. A change requested by one group, such as an additional safety endpoint required by a regulator, carries cost, timeline, and commercial consequences for all the others.

Unexpected bottlenecks. Decisions made in good faith can have unintended effects. A company selects a new contract manufacturer to reduce cost; a subtle difference in its quality control, missed in the initial audit, produces a failed batch months later and invites regulatory scrutiny of the whole facility. An academic collaboration loses momentum when a research fellow leaves and knowledge of the experimental setup leaves too. A change to a country's tax rules alters the economics of a trial running there and forces a budget revision or relocation.

An AI example. An AI feature inside an electronic health record may draw on data that several subcontractors process, some across borders. One gap compounds another. An unclear data flow becomes an unclear contract term, and an unclear contract term becomes an unclear line of accountability.

The response to complexity is structure: specialists, mapping, and an organization that mirrors the complexity it faces (Bennett & Lemoine, 2014). For AI, that means an inventory of systems, the data each one touches, and the vendor terms that govern it.

Ambiguity: No Precedent to Follow

Gene editing. The approval of the first CRISPR-based therapy in 2023 (FDA, 2023) settled one question and opened others. Should germline editing, which would affect future generations, ever be permitted? How will patients gain equitable access to therapies that are likely to be very expensive? Regulation lags the science, so companies in this field need strong internal ethical frameworks of their own.

AI and bias. AI can analyze very large datasets in drug discovery and patient care, but bias in the training data travels into the output. Obermeyer et al. (2019) found that a widely used population health algorithm predicted health care costs rather than illness; because less was spent on Black patients' care, it underestimated their need, and correcting the disparity would have raised the share of Black patients identified for additional help from 17.7 percent to 46.5 percent. Companies must decide how to make AI fair when the data it depends on may not be.

Unknown long-term consequences. A new cell therapy may offer a life-changing benefit to patients with a terminal illness and still carry risks that appear years later, as the CAR T cell boxed warning shows (FDA, 2024). Should access be restricted, denying a chance to patients in great need, or made widely available, accepting the risk of a later crisis? No rule answers that question in advance.

A clash of values. Ambiguity also arises when stakeholders interpret ethical obligations differently. The World Medical Association's Declaration of Helsinki, revised in October 2024, requires that a new intervention be tested against the best proven intervention and permits placebo when no proven intervention exists, with "extreme care" to avoid abuse of that option (World Medical Association, 2024). In regions where even the existing standard of care is largely inaccessible, sponsors must balance scientific rigor with their obligations to underserved populations. When a breakthrough drug for a rare disease is approved at a price few can pay, is it a success? Advances in stem cell research or predictive genetic testing can collide with public sentiment, religious belief, or fear of misuse, even when the innovation appears beneficial.

An AI example. Generative AI that drafts patient messages or summarizes clinical evidence raises questions that no rule or prior case yet answers.

The response to ambiguity is experimentation: a bounded trial, with a narrow use, a defined population, a short duration, and clear criteria for stopping (Bennett & Lemoine, 2014). Gathering more information does little when the rules themselves are not yet known.

The Decisions Boards Actually Make

These conditions matter most at the "go" or "no-go" points where the board commits capital or reputation. In my 2023 article I listed four. The board examples in that article also referred to two more, so I list six here, and add the one that now comes up most often:

• R&D program initiation: whether to invest in a new program, based on scientific feasibility, market potential, strategic fit, and available resources.
• Clinical trial progression: whether to advance a candidate from preclinical work into clinical trials, or from one phase to the next, based on safety and efficacy data, regulatory requirements, and trial design.
• Regulatory submission: whether to file an NDA or BLA, based on trial results, the safety profile, and efficacy data.
• Commercial launch and market entry: whether to launch, based on demand, competition, pricing and reimbursement, and manufacturing readiness.
• Portfolio prioritization: which programs or initiatives receive scarce resources, and which stop.
• Acquisition, licensing, or partnership: whether to acquire or partner with another company or license its asset.
• AI deployment: whether to approve an AI system for clinical care, coverage decisions, research, or regulatory work.

Each decision combines all four conditions in different proportions. The board's first task is to say which condition dominates, because that determines which method will help.

Match the Method to the Gap

Matching the decision approach to the nature of the gap is the core of this article. Directors do not run these analyses themselves. The board keeps its nose in and its fingers out: it asks which method management used and what the method found. The table summarizes the match; the sections that follow describe each method and give a board example.

Gap	Response	Methods management can use	The board's question
Volatility	Agility and slack	Scenario planning; war gaming; red team and blue team review	Which changes are coming, from what date, and what is our response to each?
Uncertainty	Information	Local validation; the Delphi method; horizon scanning; wild card scenarios	What evidence shows this will work here, and what would change our view?
Complexity	Structure	Systems mapping; agent-based modeling; multi-criteria decision analysis	What does this decision touch, and who is accountable for each part?
Ambiguity	Experimentation	Bounded pilots; devil's advocate; dialectical inquiry; assumption surfacing	What would tell us we were wrong, and how quickly would we know?

Volatility: Scenario Planning, War Gaming, and Red Teams

Scenario planning explores several plausible futures by identifying the main drivers of change and their implications, so that leaders can prepare strategies that adapt as new data arrive. Pierre Wack, who led the business environment division of Royal Dutch Shell's planning department, developed the Shell approach to scenario planning with Edward Newland (Wack, 1985). War gaming simulates competitors' moves to test whether a strategy holds under different competitive responses. A red team and blue team review assigns one group to challenge another group's plan and assumptions in a structured debate.

Board example (R&D program initiation). A pharmaceutical company's board considers funding a program for a rare disease. Early data are promising, but market size, the competitive field, and the regulatory path may all shift. Management presents three scenarios with a contingency plan for each. The board's job is to test whether the scenarios are genuinely different, and whether the plan names the signal that would move the company from one scenario to another.

AI example. For an AI deployment, the volatile facts are the rules. The board asks which rules apply, from what date, and who tracks changes. A scheduled regulatory review, owned by a named executive, is the response.

Uncertainty: Information, Expert Panels, and Horizon Scanning

The Delphi method, which RAND developed in the 1950s, asks a panel of experts to answer questions anonymously, shows them a summary of the group's answers, and repeats the process in rounds to move toward consensus (RAND Corporation, n.d.). It is useful when data on future events or long-term outcomes are lacking. Horizon scanning systematically gathers and analyzes weak signals of emerging trends, risks, and opportunities. Wild card scenario planning extends ordinary scenario planning to low-probability, high-impact events.

Board example (a strategic investment decision). A medical device company's board evaluates a new technology that could change how a class of surgeries is performed. In 2023 I labeled this a commercial launch decision; it is not. It is a decision about where to invest ahead of an uncertain shift. Management convenes a Delphi panel of surgeons, engineers, and payers to estimate adoption, and builds wild card scenarios ranging from incremental improvement to rapid displacement of the current approach. The board asks which early signals would confirm each path.

AI example. The board asks what local evidence shows that the system works for this organization's patients or data. The sepsis model example above shows why a vendor's reported performance is not that evidence. Local validation, a period of running the model without acting on it, and a staged rollout are the response.

Complexity: Systems Mapping, Modeling, and Multiple Criteria

Systems mapping represents the main elements, relationships, and feedback loops of a system, so that leaders can see how a decision may cascade. Agent-based modeling simulates the behavior of many interacting participants to test what outcomes different strategies produce. Multi-criteria decision analysis is a group of structured techniques for decisions that involve several competing criteria; an International Society for Pharmacoeconomics and Outcomes Research (ISPOR) task force has published good-practice guidance on its use in health care (Thokala et al., 2016).

Board example (portfolio prioritization). A health system's board wants to improve the health of its community, which depends on social, economic, and environmental factors that interact. Management maps the drivers and feedback loops, models the likely effect of several interventions, and ranks them against the mission, feasibility, and expected impact. The board asks whether the criteria and their weights reflect the strategy the board approved.

AI example. The board asks what the system touches and who is accountable for each part. The AI inventory, a data-flow map, and a review of the vendor contract are the response.

Ambiguity: Structured Challenge and Bounded Experiments

In 2023 I wrote that ambiguity is present in all categories. Some ambiguity is present in every decision, but Bennett and Lemoine are right to treat it as its own condition, with experimentation as its own response. Three structured-challenge methods help a board see what it does not know. Devil's advocacy assigns a person or team to challenge the prevailing view. Dialectical inquiry develops a counter-plan built on opposing assumptions and debates the two. Researchers have compared both methods with consensus decision-making in controlled studies of strategic decisions (Schweiger et al., 1986). Assumption surfacing lists the assumptions behind a strategy explicitly so that each can be tested.

Board example (acquisition). A biotechnology company's board considers acquiring a startup with a novel gene therapy platform. In 2023 I labeled this a strategic partnership decision; it is an acquisition. The upside is large, but long-term efficacy and safety are unclear, and so is cultural fit. The board asks a director or adviser to argue against the acquisition, asks management for the counter-case, and asks for a written list of the assumptions that must hold for the price to make sense.

AI example. For a generative AI use with no precedent, the board approves a bounded pilot and asks what would tell it the pilot had failed, and how quickly it would know.

Leading With Imperfect Information

Frameworks are necessary but not sufficient. In my view, decisions made under uncertainty rarely go wrong for lack of a framework alone. They go wrong when the culture around the decision does not let bad news travel, or when no one goes back to check what the decision produced. Five leadership practices matter.

Accept uncertainty as part of planning. Imperfect information does not leave leaders powerless. Acknowledging that not every variable can be controlled shifts the effort toward decisions that hold up across several possible futures. Scenario planning, option valuation, and continuous monitoring all support this.

Decide in increments, and watch the total. A series of smaller decisions lets the organization adapt as information arrives. However, a series of small decisions can still add up to a commitment no one approved. Each increment should return to the decision-maker with the evidence the previous step was meant to produce.

Use technology and expertise, and govern both. Predictive analytics, simulation, real-world data, and AI can find patterns in large datasets and simulate scenarios. Expert panels and advisory boards can fill gaps the data cannot. When AI supports the evidence itself, governance follows. In January 2026, FDA and the European Medicines Agency (EMA) issued ten joint guiding principles for good AI practice in drug development, including a risk-based approach to validation and oversight scaled to the context of use (Al-Faruque, 2026).

Build a culture that learns. Organizations should update their understanding as information arrives, learn systematically from past decisions, and encourage calculated risk-taking. Leaders shape that culture by showing comfort with ambiguity. One practical tool is the pre-mortem: before a project starts, the team imagines that it has already failed and writes down why, which lets people with reservations raise them early (Klein, 2007).

Engage stakeholders and communicate the basis for the decision. On a clinical trial progression decision, hearing from clinical experts, patient advocates, regulatory advisers, and payer representatives gives a fuller view of risks and opportunities, uncovers blind spots, and builds trust. When a decision is made under uncertainty, leaders should say what information it rests on, what remains uncertain, and why they chose as they did.

Ethics and Values Under Uncertainty

Because these decisions affect patients, ethical considerations belong at the front of the analysis, not at the end. Beauchamp and Childress's four principles of biomedical ethics, respect for autonomy, non-maleficence, beneficence, and justice, give leaders a shared vocabulary for moral dilemmas (Trujillo, 2021). In clinical trial design, leaders must balance a therapy's potential benefits against risks to participants, with appropriate safeguards and fully informed consent. A clear set of organizational values, such as a commitment to patients, serves as a compass when the rules are silent. The board should expect ethical considerations to be discussed openly and recorded with the decision.

Metrics, Feedback, and Data Governance

Decisions under uncertainty improve only if their outcomes are measured. Organizations should set clear objectives and metrics, create feedback loops, and refine their decision processes over time. After a launch, leaders should monitor uptake, patient outcomes, and financial performance, and feed what they learn into the next decision. Small, iterative pilots reduce the risk of large failures. Directors report that this is where AI oversight is weakest: in PwC's 2026 survey of about 600 public company directors, 66 percent rated the metrics linking AI outcomes, risks, and business performance as fair or poor, and only 18 percent rated them good or excellent (Gaskin, 2026). Data governance, with clear policies for the collection, storage, sharing, and analysis of sensitive health information, is the foundation for all of it.

Approve With Conditions, Not Only With Estimates

Most AI proposals reach the board with a savings estimate. Fewer arrive with conditions. Three practices make an approval under imperfect information defensible, and they apply equally to a research program or an acquisition.

First, stage the approval. Define the scope, the population, the duration, and the date the decision returns to the board. Second, attach conditions: a named accountable executive, the metrics management will monitor, how incidents reach the board, and the criteria for stopping. For AI, vendor terms belong on that list: no training on the organization's data, notice before the model changes, audit rights, and indemnification. Third, ask whether management ran a pre-mortem, and what it found.

Model change deserves particular attention. For AI-enabled medical devices, FDA finalized guidance on predetermined change control plans in December 2024, which allows manufacturers to make modifications described in advance without a new submission for each significant change (Organ et al., 2025). A board approving such a device, or buying one, should know what changes the plan permits.

Why the Record Matters

Under Delaware law, directors must make a good-faith effort to ensure that the company has a system for reporting compliance risk to the board. The duty traces to In re Caremark International Inc. Derivative Litigation, 698 A.2d 959 (Del. Ch. 1996). In Marchand v. Barnhill, 212 A.3d 805 (Del. 2019), the Delaware Supreme Court applied that duty to a compliance risk that was "mission critical" to the business (Marchand v. Barnhill, 2019). In In re McDonald's Corp. Stockholder Derivative Litigation, 289 A.3d 343 (Del. Ch. 2023), the Court of Chancery held that corporate officers owe a comparable duty of oversight (Chapman et al., 2023).

The duty has reached life sciences directly. On September 2, 2025, the Court of Chancery declined to dismiss a claim that directors of the pharmaceutical company Teligent utterly failed to establish board-level reporting on FDA compliance; the complaint alleged that the company lacked a board committee for FDA compliance for a prolonged period and that management decided whether and when to escalate regulatory issues (Jenner & Block, 2025). These are allegations at the pleading stage, not findings, but the lesson for boards is clear.

Whether a court would treat a particular AI deployment as mission critical depends on the facts. For a health system whose AI shapes clinical care, or a drug developer whose AI shapes regulatory evidence, the prudent assumption is that it could. On January 7, 2025, FDA issued draft guidance proposing a risk-based framework for establishing the credibility of AI models used to support regulatory decisions about drugs and biological products; as of October 21, 2026, it remains a draft (FDA, 2025). A board that approves a submission strategy, or the in-licensing of an asset whose evidence came from AI, should ask a reviewer's questions first.

The law itself remains volatile. A December 11, 2025, executive order created a Department of Justice (DOJ) AI Litigation Task Force to challenge state AI laws and named Colorado's AI Act (Paul Hastings, 2025). In April 2026, DOJ moved to intervene in xAI's lawsuit challenging the 2024 Colorado law (U.S. Department of Justice, 2026), weeks before the legislature replaced that law. A board cannot wait for these questions to settle.

Two external frameworks give management a structure and the board a benchmark. The National Institute of Standards and Technology (NIST) AI Risk Management Framework organizes the work into four functions: govern, map, measure, and manage (NIST, 2023). The Joint Commission launched its voluntary Responsible Use of AI in Healthcare certification on June 1, 2026; it certifies organizations rather than tools, across governance, data management, risk and bias reduction, safety monitoring and validation, and transparency and training (Guerra, 2026). The standard of oversight is procedural, so the evidence is the board's own record: the inventory it received, the method it asked about, the conditions it attached, and the reports it reviewed.

The Strongest Objection

The strongest objection is that structured methods slow decisions and lend false precision to guesses. Experienced directors, the argument runs, already sense which risks matter; scenario sets, Delphi panels, and decision matrices add cost and time, and a weighted score can make a judgment look like a measurement. Some directors add that these are management's tools, and that a board asking about them is drifting toward management.

Each point has merit, and each is answered by the distinction this article draws. The board does not run the methods; it asks one question about each gap, and the cost of a question is small. Delay, by contrast, is costly, as the patent clock and the trial economics above show. Matching the method to the gap is what prevents waste: it stops the board from demanding more studies when the real gap is ambiguity, where only a bounded experiment will help. As for judgment, experience is valuable, but judgment that leaves no record does not satisfy an oversight duty that courts assess by what the board did and documented. The methods do not replace judgment. They make it visible.

The Point for Directors

The goal is not to eliminate imperfect information; in healthcare and life sciences that is impossible. The goal is to build the habits, methods, and culture that let an organization decide well when information is incomplete. For a board, that means naming the gap, matching its questions to it, attaching conditions, and recording what it decided and why. Imperfect information is not a reason to defer a decision. It is also not a reason to approve one on an estimate alone.

For the Board

• For this decision, which gaps are volatility, uncertainty, complexity, or ambiguity, and which one dominates?
• Which method did management use for the dominant gap, and what did it find?
• Are we asking for more data where the real gap is ambiguity, and would a bounded pilot answer the question faster?
• What local evidence shows that this AI system performs for our patients or our data, and who will monitor it after it goes live?
• What conditions are attached to this approval, including vendor terms on training data, model changes, audit rights, and indemnification, and when does the decision return to the board?
• What will the minutes show about how we reached this decision?

The board question sets for AI approvals are in the Avancer AI Governance Navigator.

References

Al-Faruque, F. (2026, January 14). EMA, FDA issue joint AI guiding principles for drug development. Regulatory Affairs Professionals Society. https://www.raps.org/resource/ema-fda-issue-joint-ai-guiding-principles-for-drug.html

Bennett, N., & Lemoine, G. J. (2014). What VUCA really means for you. Harvard Business Review, 92(1/2). https://hbr.org/2014/01/what-vuca-really-means-for-you

Chapman, J. J., Rothschild, J. L., & Jaroff, A. F. (2023, February 8). Delaware Court of Chancery extends oversight duties to corporate officers. McGuireWoods. https://www.mcguirewoods.com/client-resources/Alerts/2023/2/delaware-court-chancery-extends-oversight-duties-corporate-officers?p=2

Contents and term of patent; provisional rights, 35 U.S.C. § 154. Retrieved October 21, 2026, from https://www.law.cornell.edu/uscode/text/35/154

Darch, J. E., Brooks, P., Daudi, S., Hinckley, D., & Kaye, A. (2026, May 19). Colorado scales back AI law with targeted implications for health care. Ropes & Gray. https://www.ropesgray.com/en/insights/alerts/2026/05/colorado-scales-back-ai-law-with-targeted-implications-for-health-care

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, A. (2026, June 1). Joint Commission AI certification. HealthSystemCIO. https://healthsystemcio.com/2026/06/01/joint-commission-ai-certification/

Hunton Andrews Kurth. (2026, July 28). EU Digital Omnibus on AI enters into force. Privacy and Cybersecurity Law Blog. https://www.hunton.com/privacy-and-cybersecurity-law-blog/eu-digital-omnibus-on-ai-enters-into-force

In re Caremark International Inc. Derivative Litigation, 698 A.2d 959 (Del. Ch. 1996).

In re McDonald's Corporation Stockholder Derivative Litigation, 289 A.3d 343 (Del. Ch. 2023).

Jenner & Block. (2025, September). Delaware court denies dismissal of Caremark oversight claims in Teligent case, highlighting board and officer compliance obligations. https://www.jenner.com/en/news-insights/publications/delaware-court-denies-dismissal-of-caremark-oversight-claims-in-teligent-case-highlighting-board-and-officer-compliance-obligations

Klein, G. (2007). Performing a project premortem. Harvard Business Review, 85(9), 18–19. https://hbr.org/2007/09/performing-a-project-premortem

Marchand v. Barnhill, 212 A.3d 805 (Del. 2019). https://law.justia.com/cases/delaware/supreme-court/2019/533-2018.html

National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). U.S. Department of Commerce. https://www.nist.gov/itl/ai-risk-management-framework

Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447–453. https://pubmed.ncbi.nlm.nih.gov/31649194/

Organ, S. A., Beaver, N. A., & Rosen, D. L. (2025, January 6). AI-enabled device software functions: FDA's final guidance for predetermined change control plans. Foley & Lardner. https://www.foley.com/insights/publications/2025/01/fda-final-guidance-ai-device-software-predetermined-change-control-plan/

Paul Hastings. (2025, December 16). President Trump signs executive order challenging state AI laws. https://www.paulhastings.com/insights/client-alerts/president-trump-signs-executive-order-challenging-state-ai-laws

RAND Corporation. (n.d.). Delphi method. Retrieved October 21, 2026, from https://www.rand.org/topics/delphi-method.html

Schweiger, D. M., Sandberg, W. R., & Ragan, J. W. (1986). Group approaches for improving strategic decision making: A comparative analysis of dialectical inquiry, devil's advocacy, and consensus. Academy of Management Journal, 29(1), 51–71. https://doi.org/10.2307/255859

Silverman, E. (2023, June 6). FDA will allow troubled chemotherapy maker to resume distribution of cancer drugs in short supply. STAT. https://www.statnews.com/pharmalot/2023/06/06/fda-cancer-intas-cisplatin-cancer-chemotherapy-shortages/

Thokala, P., Devlin, N., Marsh, K., Baltussen, R., Boysen, M., Kalo, Z., Longrenn, T., Mussen, F., Peacock, S., Watkins, J., & Ijzerman, M. (2016). Multiple criteria decision analysis for health care decision making: An introduction (Report 1 of the ISPOR MCDA Emerging Good Practices Task Force). Value in Health, 19(1). https://www.ispor.org/publications/journals/value-in-health/abstract/Volume-19--Issue-1/Multiple-Criteria-Decision-Analysis-for-Health-Care-Decision-Making%E2%80%94An-Introduction--Report-1-of-the-ISPOR-MCDA-Emerging-Good-Practices-Task-Force

Thomas, D., Chancellor, D., Micklus, A., LaFever, S., Hay, M., Chaudhuri, S., Bowden, R., & Lo, A. W. (2021). Clinical development success rates and contributing factors 2011–2020. Biotechnology Innovation Organization, Informa Pharma Intelligence, and QLS Advisors. https://go.bio.org/rs/490-EHZ-999/images/ClinicalDevelopmentSuccessRates2011_2020.pdf

Trujillo, G. M., Jr. (2021, February 16). Principlism in biomedical ethics: Respect for autonomy, non-maleficence, beneficence, and justice. 1000-Word Philosophy. https://1000wordphilosophy.com/2021/02/16/principlism-in-biomedical-ethics/

U.S. Department of Justice. (2026, April 24). Justice Department intervenes in xAI lawsuit challenging Colorado's algorithmic discrimination law [Press release]. https://www.justice.gov/opa/pr/justice-department-intervenes-xai-lawsuit-challenging-colorados-algorithmic-discrimination

U.S. Food and Drug Administration. (2023, December 8). FDA approves first gene therapies to treat patients with sickle cell disease [Press release]. https://www.fda.gov/news-events/press-announcements/fda-approves-first-gene-therapies-treat-patients-sickle-cell-disease

U.S. Food and Drug Administration. (2024, April 18). FDA requires boxed warning for T cell malignancies following treatment with BCMA-directed or CD19-directed autologous chimeric antigen receptor (CAR) T cell immunotherapies. https://www.fda.gov/vaccines-blood-biologics/safety-availability-biologics/fda-requires-boxed-warning-t-cell-malignancies-following-treatment-bcma-directed-or-cd19-directed

U.S. Food and Drug Administration. (2025). Considerations for the use of artificial intelligence to support regulatory decision-making for drug and biological products (Draft guidance; Docket No. FDA-2024-D-4689). Retrieved October 21, 2026, from https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological

Wack, P. (1985). Scenarios: Uncharted waters ahead. Harvard Business Review, 63(5), 72–89. https://hbr.org/1985/09/scenarios-uncharted-waters-ahead

Wong, A., Otles, E., Donnelly, J. P., Krumm, A., McCullough, J., DeTroyer-Cooley, O., Pestrue, J., Phillips, M., Konye, J., Penoza, C., Ghous, M., & Singh, K. (2021). External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Internal Medicine, 181(8), 1065–1070. https://doi.org/10.1001/jamainternmed.2021.2626

World Medical Association. (2024, October 19). WMA Declaration of Helsinki: Ethical principles for medical research involving human participants. https://www.wma.net/policies-post/wma-declaration-of-helsinki/

Wouters, O. J., McKee, M., & Luyten, J. (2020). Estimated research and development investment needed to bring a new medicine to market, 2009–2018. JAMA, 323(9), 844–853. https://doi.org/10.1001/jama.2020.1166