How this course works
Six sessions, 45 minutes each, one artifact: a decision memo template that fills in section by section - default action and criteria tonight, calibrated estimates in session 2, base rates and updates in session 3, promised uncertainty in session 4, a value-of-information verdict in session 5, and the communication plan in session 6. The running case is Mango Lane's decision desk: the same fashion-app universe as the ensemble and deep learning courses, now facing the decisions those models feed.
★ What you walk out with tonight. The one-sentence discipline definition from the person who named it. The default-first habit that inoculates a meeting against confirmation bias. And the six-word question that turns any opinion-holder into a testable forecaster.
LiveThe discipline, in its founder's words3 min▶
Cassie Kozyrkov, who built the field at Google and trained tens of thousands there, defines decision intelligence as "a new academic discipline concerned with all aspects of selecting between options" - one that "brings together the best of applied data science, social science, and managerial science into a unified field that helps people use data to improve their lives, their businesses, and the world around them."
Note what the definition centers: selecting between options. Not models, not dashboards, not pipelines - those are means. She calls it "a vital science for the AI era, covering the skills needed to lead AI projects responsibly and design objectives, metrics, and safety-nets for automation at scale." The whole course lives inside that sentence.
Mango Lane has eleven dashboards, three models, and a weekly meeting where the churn number is admired and nothing changes. Every tool upstream of a decision that never gets made is a cost center wearing a lab coat. The memo this course builds exists to make the decision - not the analysis - the unit of work.
Self-studyThe vocabulary for the whole arc3 min▶
- Default action - what you will do if no further information arrives. Every real decision has one, usually unstated. Stating it is tonight's first move.
- Decision criteria - the pre-committed rule: "we launch if the estimate clears X". Set before data, or the data gets fitted to the wish.
- Calibration - whether your stated confidence matches your hit rate. Session 2 measures yours, live, and it will sting a little.
- Base rate - how often things like this happen, before your case's details. Session 3's anchor.
- Value of information - the most a study, test, or dataset is worth to THIS decision. Computable, and session 5 computes it live.
Commit to the default before you look
7 min liveKozyrkov's Harvard Business Review advice for great decision makers opens with five words that reorder everything downstream: "Commit to your default decision up front." The reason is not procedure. It is what looking at data does to a mind that has not committed.
LiveWhy the order matters: the confirmation-bias mechanism4 min▶
Once a mind holds a preferred conclusion, data becomes a buffet: reach for what agrees, explain away what does not. Nobody experiences themselves doing it - that is what makes it bias rather than lying. Kozyrkov's essays frame the fix as an ordering constraint, not a virtue: if the criteria exist before the data arrives, the data can genuinely overturn the default. If they come after, the criteria were reverse-engineered from the wish.
The default itself is not a guess at the answer. It is the honest statement of what happens on no-new-information: keep the current voucher policy, do not launch, wait a quarter. Stating it out loud does two things - it exposes who actually holds the decision, and it converts the meeting's question from "what does the data say" to the answerable "does the data clear OUR pre-stated bar".
The six-word tool. Kozyrkov's signature question for any opinion-holder: what would it take to change your mind? A person with an answer is a decision-maker with criteria; a person without one is decoration around a foregone conclusion - and knowing which you are talking to changes what an analysis is worth building.
Self-studyThe role split: whose job is which question4 min▶
Kozyrkov's framework for data teams, paraphrased from her Harvard Business Review essay on great analysts: the roles answer different questions and mixing them up burns both.
| Role | Question | Speed vs rigor |
|---|---|---|
| Analyst | What is in the data? What is worth a closer look? | Fast exploration - conclusions stay INSIDE the data looked at |
| Statistician | Can we responsibly conclude this beyond the data? | Slow rigor - the ones who keep you from fooling yourself |
| ML engineer | Can we automate this at scale? | Production discipline |
| Decision-maker | What are the options, the default, the criteria? | The scarce skill the other three exist to serve |
The failure mode her framework names: organizations hire the first three lavishly and assume the fourth happens by itself. It does not - which is why this course exists and why its audience chip says leaders AND practitioners.
The decision memo: one page that holds the whole course
6 min liveEverything the six sessions teach lands in one artifact: a memo short enough to actually get read, structured enough that skipping a section is visible. Here is the skeleton; tonight fills the first two rows, and each session hands you the next.
LiveWhy a memo and not a dashboard3 min▶
A dashboard answers "what is happening"; a memo answers "what are we doing about it, and what would change that". The dashboard has no default action, no criteria, no owner - which is precisely why eleven of them can coexist with zero decisions at Mango Lane. The memo's six rows are each a question someone must own, and an empty row is a visible confession rather than a silent gap.
It is also deliberately one page. A decision that cannot state its default, bar, estimates and verdict in a page is not a complex decision - it is an unmade one hiding in length.
The voucher question that runs through this course - "spend $12k on a retention campaign for the flagged segment?" - has sat in Mango Lane's backlog for two months, revisited weekly, never decided. Not for lack of data: the churn model works. For lack of a stated default and a bar. Tonight it gets both, and by session 5 the memo will say precisely what the much-requested "more analysis" is worth: less than it costs.
The judgment layer: what this course adds to your models
5 min liveMango Lane already has a churn model at 73 percent holdout and a routing playbook for its AI spend. What is left is the layer those courses point at but do not teach: the human judgment that turns a score into an action.
LiveWhy intuition alone fails here, measurably3 min▶
The judgment layer cannot be vibes, because human judgment has measured, replicated failure modes exactly where decisions live. Two the course will demonstrate on you, not at you:
- Overconfidence: across the classic calibration literature, untrained people asked for 90 percent confidence intervals capture the truth far less often - typical hit rates run 40 to 60 percent. Session 2's live lab measures yours; almost everyone's first number stings.
- Neglected base rates: the case's vivid details crowd out how often things like this actually happen. Session 3's forecast game makes the cost visible: a bot that only ever forecasts the base rate is embarrassingly hard to beat.
Both are trainable - that is the course's optimistic spine. A roughly 45-minute probabilistic-reasoning module measurably improved forecasters' accuracy for months in the largest forecasting tournament ever run. Judgment is a skill with a gym, and sessions 2 and 3 are the gym.
Write the memo's first two sections
★ 18 min · everyone buildsThe voucher decision, for real: should Mango Lane spend $12k on a retention campaign for the churn model's flagged segment? Tonight you write the default and the criteria - and feel how much of the meeting's usual fog those two sections burn off.
State the default action in one sentence. What happens if no new information arrives? For the voucher case: "we do not launch; the $12k stays in the retention budget." Notice the default is not neutral - not-launching is also a choice with consequences, and writing it down makes that visible.
Name the decision-maker. One name, not a committee. The person whose mind the criteria must be able to change. If you cannot name them, you have found out why the decision has been circling for two months.
Write the criteria as a flip-condition. "We launch IF the campaign's expected net gain exceeds $4k under our own estimates." Numbers you do not have yet - that is fine and correct: the criteria's job tonight is to define what sessions 2 through 5 must estimate, not to contain the answer.
Ask the six-word question of yourself. What would it take to change YOUR mind about the launch? Write the honest answer under the criteria. If your answer is "nothing realistic", the memo has just saved everyone the cost of the analysis - decoration cancelled.
Now repeat all four moves on a decision from your own week. Real one, live one. Default, owner, flip-condition, mind-changer. Keep it - every remaining session adds its section to this memo, and session 6 delivers the finished one.
What to bring to session 2. Your memo with sections 1 and 2 filled. Session 2 fills section 3 - and before it lets you estimate anything for the memo, it measures how much your "90 percent sure" is actually worth, live, on ten questions with verified answers.
Before session 2
30 min- Audit one recent decision meeting against tonight's order: was a default stated? Criteria before the data came out? Write two sentences on what actually happened.
- Collect three "what would change your mind" answers - ask colleagues holding opinions about live decisions. Score each answer: criteria, or decoration?
- Find the unstated default in one stalled decision you know. Stalling IS choosing the default without admitting it - write down what is being chosen by the stall.
- Read Kozyrkov's HBR piece "The First Thing Great Decision Makers Do" - ten minutes, and the default-first move sticks permanently.
- Optional: start her free Making Friends with Machine Learning course (the first hour) - the same decisions-first framing applied to the whole ML lifecycle.
Three questions
Clear all three and session 1 gets stamped in your passport.
1 · Why must the default action and criteria be committed BEFORE looking at the data?
The ordering constraint is the discipline's core move: pre-committed criteria let data genuinely overturn the default; post-hoc criteria get reverse-engineered from the wish, invisibly, by normal human minds. Kozyrkov's five-word version: commit to your default decision up front.
2 · A stakeholder insists the retention campaign is essential. You ask what it would take to change their mind, and they say "honestly, nothing - the data team should just confirm it." What have you learned?
The six-word question is a detector: an answer names criteria; "nothing" names a foregone conclusion. Analysis that cannot change an outcome has zero decision value - session 5 makes that literal with a formula, and it prices this analysis at exactly $0.
3 · In Kozyrkov's role framework, what makes the decision-maker "the scarce skill"?
Her framework (paraphrased from the HBR analyst essay): analysts explore fast, statisticians conclude carefully, engineers automate - and all three serve a decision question someone must frame well. The framing skill gets no headcount line, which is why data stacks outnumber good decisions.
What this session teaches, and where it came from
Every fact on this page, and its verification tier, is recorded in the course's source map.
Session 1 cheat sheet · pin this
The memo's estimates section demands numbers with honest uncertainty attached - and before you can give any, session 2 measures what your "90 percent sure" is actually worth: ten questions, verified answers, your own hit rate computed live. Across the classic studies, untrained estimators land between 40 and 60 percent. Bring humility; leave with a trainable skill.