Part III · Strategic Leadership in the AI Era

The Data-Driven Decision Process

The Data-Driven Decision Process applies scientific method to business decisions: frame a question, state competing hypotheses, identify what evidence would separate them, gather it, conclude, act, measure.

What it diagnoses

Analysis that cannot change anyone’s mind.

The Data-Driven Decision Process applies scientific method to business decisions: frame a question, state competing hypotheses, identify what evidence would separate them, gather it, conclude, act, measure. The stated objective in the book is the scientific method for business, and the word doing the work in that phrase is competing.

What it diagnoses is the sequence error. Most organizations start with data. They pull everything available, build the model, and present the analysis, and the analysis supports the position the requester already held, because the question was never specified tightly enough to have been capable of producing a different answer. That is not a data problem and more data does not fix it.

Three states this surfaces.

Data-rich and decision-poor. Substantial analytics investment, no observable change in how calls get made. HBR has covered the gap between stated data culture and actual practice for years, and the persistent finding is that the constraint is rarely the tooling.

Dueling datasets. Two teams, two sources, two opposite conclusions, and no agreed method for resolving it. This is not usually a measurement dispute. It is an unstated disagreement about what question is being answered.

Judgment abdicated to the model. HBR’s work on where data-driven decision-making goes wrong describes the overcorrection directly: managers deferring to the output and setting aside domain knowledge that the data does not contain. The research position, consistently, is that data combined with managerial expertise beats either alone.

How to run it

The book runs six stages.

1. 1

   Stage 1, frame the decision and articulate competing hypotheses (days 1 to 2). Write the decision as a question with at least two live answers, and write what each would predict. If you cannot state a hypothesis that would lose, you are not running an analysis, you are building a case.

2. 2

   Stage 2, identify required data and existing evidence (days 3 to 5). For each hypothesis, name the evidence that would separate it from the others, then inventory what you already hold. Most decisions are less data-starved than they feel, and finding that out early prevents a collection exercise nobody needed.

3. 3

   Stage 3, design the collection or the experiment (days 6 to 10). Decide in advance what result would change the decision. This is the discipline that makes the whole thing falsifiable, and it has to be written down before any data arrives.

4. 4

   Stage 4, collect and run (variable). Hold the design. Mid-stream changes to what is being measured are how a test becomes a story.

5. 5

   Stage 5, analyze and conclude (days 1 to 3 after collection). Report against the hypotheses you stated, including the ones that lost. A conclusion that does not mention the rejected alternatives is not showing its work.

6. 6

   Stage 6, decide, execute, measure (ongoing). Log the decision, the reasoning, and the expected outcome, then check it. Without stage 6 the organization learns nothing transferable and runs the same debate again next year.

How operators misuse it

They gather first and frame never. This is the dominant failure and it produces the most confident wrong answers, because the analysis is genuinely rigorous and aimed at a question nobody specified.

They treat precision as the goal. The required precision of a number is set by what the number will be held to. A directional estimate with its assumptions documented beats a precise figure nobody can reconstruct, and chasing the second one is how a two-day decision becomes a six-week project.

They skip the decision log. Recording the outcome is boring and it is the only stage that compounds. Without it, every conclusion has to be re-litigated from scratch because nobody can reconstruct why the last one was made.

What the pattern shows

The documented failure is not scarcity. It is a translation gap between analysis and decision, and the research keeps locating it in the same place.

Harvard Business Review’s 2024 work on where data-driven decision-making goes wrong describes organizations that have over-rotated: managers who defer to the model, discount what they know from operating the business, and produce decisions that are defensible on the analysis and wrong in the context. The failure mode is not insufficient rigor. It is rigor applied without the domain knowledge required to interpret it.

HBR’s earlier research on combining data with managerial expertise establishes the other half. The consistent finding across that work is that neither the model alone nor the experienced manager alone outperforms the combination. Which means the interesting design question is not whether to be data-driven. It is where in the process human judgment enters, and what it is allowed to override.

McKinsey’s work on expertise and judgment in a data-driven world lands in the same territory from the capability side, treating domain judgment as an input to be deliberately engineered into the analytical process rather than a bias to be removed from it.

And HBR’s reporting on organizations describing themselves as data-driven without the practice to match names the honest starting condition for most operators: the investment is real, the culture claim is aspirational, and the gap is in decision process rather than in tooling.

Read together, the pattern is consistent. Organizations that improve here do not buy better analytics. They specify the question earlier, name what would change their mind, and decide in advance where judgment is allowed to overrule the output.

Where it lives in the book

Framework 18, Chapter 6, page 348. The Scientific Method for Business. It sits between Human-in-the-Loop and the Trust Equation: Human-in-the-Loop establishes who holds the decision right, the Data-Driven Decision Process is the method that decision runs through, and the Trust Equation determines whether the organization will act on the result.

Questions operators ask

How is this different from just being analytical?

The competing hypotheses requirement. Being analytical describes rigor. This framework requires that a specified outcome could have lost, which is what makes the exercise capable of changing a decision.

What if the data is incomplete?

It always is. Stage 3 asks what result would change the decision, which is answerable under incomplete data. Waiting for completeness is a decision to decide late.

Two teams, two datasets, opposite answers. Where do I start?

Not with the datasets. Get both teams to write the question they think they are answering. The conflict is usually there, and it is invisible while everyone argues about numbers.

How long should this take?

The book’s stage timings run roughly two to three weeks for a substantial decision. If a decision is taking a quarter, the framing step was skipped and the collection has no stopping condition.

Does this replace managerial judgment?

The opposite. The research is consistent that data plus domain expertise beats either alone. The framework’s job is to put judgment in a named place in the sequence rather than leaving it to leak in wherever it can.

Sources

1. Michael Luca and Amy C. Edmondson, "Where Data-Driven Decision-Making Can Go Wrong," Harvard Business Review, 2024
2. Paolo Gaudiano, "The Best Approach to Decision Making Combines Data and Managers’ Expertise," Harvard Business Review, 2017
3. McKinsey & Company, "The role of expertise and judgment in a data-driven world"
4. Sudheesh Nair, "Is Your Business Masquerading as Data-Driven?" Harvard Business Review, 2020