Mike Ryan: How to Use AI for Better Business Decisions

An executive opens an AI-generated report before a critical meeting. The document is polished, logically structured, and reassuringly specific. It compares competitors, estimates market growth, identifies risks, and recommends a clear course of action.

There is only one problem: nobody in the room knows which claims came from verified evidence, which were inferred from incomplete information, and which were simply generated because they sounded plausible.

This is the new danger facing companies that use AI for business decisions. The technology can compress days of research into minutes, but speed does not automatically improve the quality of the evidence being processed. A weak assumption can now travel through an organization faster, appear more authoritative, and influence more people than it ever could in a manually prepared report.

Mike Ryan encountered that problem after a career that included two decades at Goldman Sachs, where he ultimately led its global equity business, followed by managing investments for Harvard’s endowment. When generic AI tools repeatedly produced convincing but inaccurate information, he began developing a more controlled approach to AI-assisted analysis. His blunt assessment captures the challenge: “Despite its power, AI wouldn’t pass a first-round job interview with most firms because it’s not trustworthy.”

Can AI Be Trusted for Business Decisions?

AI should not be treated as trustworthy by default, but that does not make it unsuitable for serious work. It means its output must be governed by a process appropriate to the consequence of the decision.

A chatbot suggesting possible titles for a presentation creates limited risk. The same system calculating a company valuation, interpreting a legal obligation, assessing an acquisition target, or recommending a major capital investment operates in an entirely different risk category.

This distinction matters because business adoption is moving faster than many organizations’ ability to evaluate and govern AI. Stanford’s 2026 AI Index describes a widening gap between rapidly expanding AI capabilities and the frameworks needed to measure, understand, and manage them. McKinsey’s 2026 research similarly found that 74% of respondents regarded inaccuracy as a highly relevant AI risk.

The first practical step is therefore to classify AI work by consequence. Low-risk uses can tolerate experimentation and light review, while high-impact recommendations require approved sources, documented assumptions, validation, named decision owners, and a clear audit trail.

Where is your organization currently using the same level of control for a social media caption and a seven-figure investment decision?

Better AI Starts With Better Evidence

Many people try to improve AI output by rewriting the prompt. Prompting matters, but even an excellent question cannot rescue unreliable or irrelevant evidence.

Ryan describes AI as having “a big, big stomach, but a very small mouth.” It may have access to an enormous world of information, yet each task depends on what enters the model’s working context and how relevant that material is to the question.

For business leaders, the implication is straightforward: trustworthy AI begins before the prompt is written.

A useful source hierarchy starts with verified internal data, such as audited financial statements, approved operating metrics, contracts, customer research, and trusted spreadsheets. The next layer contains authoritative external material, including regulatory filings, primary research, official statistics, and recognized industry sources. Expert interpretation may then be added explicitly, while open-web discovery should remain a source of leads rather than unquestioned facts.

Before asking AI to analyze a decision, define which sources it may use, how current they must be, which documents take precedence when information conflicts, and what the system must do when evidence is missing.

That last condition is essential. A trustworthy workflow permits the answer, “There is not enough evidence to conclude.” An unreliable one rewards the system for filling every blank.

Do Not Confuse Fluency With Reliability

Human beings naturally associate clear language with clear thinking. Generative AI exploits that instinct unintentionally because it can present weak reasoning with strong grammar, confident formatting, and impressive speed.

Ryan warns that AI can provide “wrong answers that look beautiful and sound smart.”

The antidote is not simply asking the model to be more accurate. Leaders should require every meaningful output to distinguish among three categories:

Verified facts should identify their sources and dates. Assumptions should be visible, testable, and open to revision. Inferences should explain how the available evidence supports the conclusion.

Consider a company evaluating a new market. An AI system might report that demand is growing, customers are underserved, and the expansion should generate attractive returns. A decision-grade analysis would go further: Which data demonstrates the demand? Is the growth historical or forecast? What must be true about pricing, customer acquisition cost, hiring, regulation, and local competition? Which assumption has the greatest influence on the outcome?

What would happen in your organization if every AI recommendation had to display the evidence that could prove it wrong?

Why Human Oversight Is Not a Rubber Stamp

Adding a person to the end of an AI workflow does not automatically make the result reliable. If the reviewer is rushed, overly impressed by the output, or unclear about their responsibility, “human oversight” becomes ceremonial approval.

Ryan’s preferred model is “AI plus one”: the system processes information and performs analytical tasks, while one accountable person directs the work, supervises the evidence, interprets its meaning, and decides what action to take.

Research supports the need for deliberate workflow design. An MIT Sloan analysis found that human-AI combinations did not automatically outperform the strongest human or AI working independently on decision-making tasks. Collaboration produced value only under the right conditions; simply pairing a person with a model was not enough.

Effective oversight therefore requires distinct roles. An operator prepares the AI-assisted analysis, a decision owner accepts responsibility for the conclusion, and a challenger actively searches for missing evidence, weak assumptions, alternative explanations, and downside scenarios. In smaller companies, one person may perform more than one role, but the activities should remain separate.

The purpose of the human is not to make the AI feel safer. It is to add context, skepticism, ethics, creativity, and accountability that the system does not possess.

Use AI to Decide What Deserves Attention

One of AI’s most valuable contributions may occur before the main analysis begins.

Executives, investors, and founders rarely suffer from a lack of possible work. They face too many opportunities, warnings, reports, customer signals, requests, and unfinished decisions. The scarce resource is not information but attention.

AI can screen a broad field of possibilities and help identify which issues warrant deeper human investigation. A startup can rank prospective partnerships by strategic fit and evidence quality. A family business can compare expansion opportunities without committing senior leadership to a full analysis of every location. An investment team can filter hundreds of companies before experienced professionals examine the most promising or concerning cases.

The screening criteria should be defined in advance: potential value, downside exposure, reversibility, evidence quality, urgency, and the cost of acquiring more information. AI can then organize the field consistently while humans focus their time where judgment creates the greatest advantage.

This is a better use of automation than asking a model to make the final decision. Let AI reduce the search space; let people carry the responsibility.

Complex Decisions Need a Living Record

Important business decisions rarely arrive fully formed in one meeting. Customer evidence appears gradually, forecasts change, competitors react, financing conditions shift, and assumptions that once seemed reasonable become obsolete.

A one-off conversation with a general chatbot is poorly suited to that reality. Even when the initial output is useful, the organization may lose track of which evidence was used, what changed between versions, why a forecast moved, or who approved the final recommendation.

A decision-grade AI workflow should preserve the history of the analysis. Each version should record its sources, assumptions, calculations, uncertainties, scenarios, unresolved questions, and the reasons behind material changes.

For a major investment, leaders might maintain a base case, an upside case, and a downside case, while recording the signals that would cause the organization to move from one interpretation to another. A separate challenge team can then test the recommendation in much the same way that a patient may seek a second medical opinion before a consequential procedure.

This process does not eliminate uncertainty. It makes uncertainty visible and manageable.

A Five-Step Decision-Grade AI Workflow

Leaders do not need to build a sophisticated technology platform before improving the quality of AI-assisted decisions. They can begin with a disciplined five-step workflow.

1. Frame the decision. 

State the exact decision, its owner, deadline, consequences, and criteria for success. “Research the market” is too broad; “Determine whether we should enter this market within the next 12 months” creates a decision boundary.

2. Build the source pack. 

Assemble the internal records, primary research, trusted calculations, and authoritative external material that should govern the analysis. Label documents by date, ownership, and reliability.

3. Structure the output. 

Require AI to separate facts, assumptions, inferences, uncertainties, conflicting evidence, and unanswered questions. Ask it to cite every consequential claim.

4. Challenge the recommendation. 

Assign a person or team to construct the strongest opposing case, identify what would invalidate the conclusion, and test the most sensitive assumptions.

5. Record and monitor the decision. 

Preserve the final rationale, the evidence used, the person accountable, and the conditions that should trigger a review.

NIST’s Generative AI Risk Management Profile similarly emphasizes defined responsibilities, evaluation, documentation, risk-appropriate human oversight, and ongoing monitoring rather than treating trust as an inherent property of a model.

What Every Leader Should Do This Week

Choose one recurring decision that matters but is not irreversible: prioritizing sales opportunities, evaluating a new supplier, reviewing customer churn, ranking product initiatives, or screening possible hires.

Instead of asking AI for a recommendation immediately, give it a controlled set of sources and require five outputs: the available facts, the assumptions being made, the strongest recommendation, the strongest argument against it, and the missing evidence that would most improve confidence.

Then ask another person to challenge the result before action is taken.

The exercise will reveal more than whether the AI produced a good answer. It will show whether your organization has agreed sources of truth, clear decision criteria, visible assumptions, and genuine accountability.

Trustworthy AI is not created by making the machine sound more certain. It is created by making evidence, uncertainty, and human responsibility impossible to hide.

Apply that discipline to one decision this week, discuss the result with your leadership team, and listen to the full podcast conversation with Mike Ryan for a deeper examination of AI, investment analysis, and human judgment.