Cassie Kozyrkov on Decision Intelligence

Cassie Kozyrkov spent nearly a decade at Google building something that didn’t exist before: the discipline of Decision Intelligence. As Google’s first Chief Decision Scientist, she trained over 20,000 Googlers to think better about data and AI. Her core insight? The hard part isn’t the algorithm—it’s knowing what question to ask.

The Decision Intelligence Framework

Most organizations fail at AI not because of technical problems, but because they skip the most important step: defining what success looks like before touching any data.

Kozyrkov’s framework starts backwards from how most teams operate:

1. Define the decision — What action will you take based on the outcome?
2. Set the criteria — How will you know if the answer is good enough?
3. Then, and only then — Go find or build the data and model

This inverts the typical approach where teams collect data first and figure out what to do with it later.

Statistics for Decision-Makers

Her most popular writing strips statistics down to what matters for actual decisions:

“Statistics is the science of changing your mind under uncertainty.”

She distinguishes between three types of data work:

• Analytics — Looking at what happened
• Statistics — Drawing conclusions about the unknown
• Machine Learning — Automating decisions at scale

Each requires different skills, different mindsets, and different acceptance criteria. Conflating them is why many AI projects fail.

The Pre-Mortem Approach

Before any project starts, Kozyrkov runs what she calls a “pre-mortem” — imagining the project has already failed and working backwards to identify why. This forces teams to confront uncomfortable questions:

• What decision would this analysis actually change?
• Who will take action based on this?
• What’s the minimum quality threshold that matters?

If you can’t answer these, you’re not ready to start.

Making AI Practical

Her advice for working with AI tools cuts through the hype:

On using AI well:

“Using AI well is a modern version of Pascal’s Wager. What’s the downside of asking? You lose a sentence. Maybe a minute. What’s the upside? You might save hours.”

On the real barrier: Most teams don’t need better models. They need clearer thinking about what they’re trying to accomplish.

Key Principles

Specify your defaults first. Before running any analysis, decide what you’ll do if the data is inconclusive. This removes the temptation to torture data until it confesses.

Separate exploration from confirmation. Looking at data to generate hypotheses is different from testing them. Mixing these stages leads to spurious conclusions.

Make the decision criteria explicit. If stakeholders can’t agree on what outcome would trigger action, no amount of data will help.

Resources

• Decision Intelligence course — Her LinkedIn Learning course on the fundamentals
• Statistics for People in a Hurry — Plain-English breakdown of what statistics actually does
• The Ultimate Guide to Starting AI — Step-by-step framework for AI projects

Why This Matters for Personal Systems

Decision Intelligence applies directly to personal AI systems. Before building any automation:

1. What decision am I trying to improve?
2. What would “good enough” look like?
3. What’s my default action if the system fails?

These questions prevent the common trap of building sophisticated systems that don’t actually change your behavior.

Kozyrkov now runs Data Scientific (renamed Kozyr), advising organizations including NASA, Lenovo, and Gucci on AI implementation. Her core message remains consistent: the technology is rarely the bottleneck. Clear thinking is.