Decision Intelligence

Decision intelligence (DI) is an engineering discipline that applies data science, decision science, and social science to improve organizational decision-making outcomes by structuring decisions systematically, rather than focusing solely on prediction or analysis. The term was coined by Cassie Kozyrkov while at Google, where she led the Decision Intelligence team as Chief Decision Scientist. ¹⁾

Kozyrkov defines DI as “the fusion of traditional data science with decision science,” distinguishing it from pure analytics by prioritizing actionable choices over models alone. Her framework at Google structures decision-making into six stages: ²⁾

1. Frame the decision: Define what needs to be decided, who decides, and what constraints exist.
2. Gather intelligence: Collect relevant data and qualitative inputs.
3. Generate options: Create a diverse set of possible actions.
4. Evaluate options: Assess each option against criteria using data, models, and judgment.
5. Choose: Make the decision with appropriate confidence.
6. Learn: Monitor outcomes and feed results back into the process.

This approach treats decisions as engineered products — scalable via repeatable processes rather than ad hoc intuition.

DI categorizes decisions along two key dimensions:

Reversibility:

• Reversible decisions: Allow iteration with low long-term cost (e.g., A/B testing pricing). These can be made quickly and corrected if wrong.
• Irreversible decisions: Demand exhaustive modeling and analysis before commitment (e.g., factory closures, major acquisitions). ³⁾

Stakes:

• High-stakes decisions: Involve significant impact and require auditable rules, explainability, and careful governance (e.g., regulatory compliance in finance). ⁴⁾
• Low-stakes decisions: Suit quick automation with minimal human oversight.

DI orchestrates multiple analytical capabilities into coherent decision flows:

• Machine learning: Provides predictive capabilities (forecasting demand, identifying patterns).
• Optimization algorithms: Enable resource allocation and operational efficiency.
• Natural language processing: Extracts insights from unstructured data like customer feedback.
• Explainable AI: Ensures transparency in automated decisions for auditing and trust.
• Statistics: Enables risk assessment and uncertainty quantification.

Unlike siloed tools, DI composes these capabilities in a decision-centric manner, monitoring outcomes for continuous learning. ⁵⁾

Gartner defines decision intelligence platforms (DIPs) as “software used to create solutions that support, automate, and augment decision making of humans or machines, powered by the composition of data, analytics, knowledge, and artificial intelligence techniques.” ⁶⁾

Key platforms include:

• Stravito: Insights management with AI personas for evidence-backed enterprise decisions. ⁷⁾
• Domo: Real-time dashboards with low-code automation for non-technical teams. ⁸⁾
• FICO: ML combined with rules engines for auditable logic in high-stakes finance (credit scoring, fraud detection). ⁹⁾
• Qlik: Active intelligence with data integration and automation for operational decisions. ¹⁰⁾
• SAS Decision Manager: Rule-based systems with regulatory integration for finance and pharma. ¹¹⁾
• H2O.ai: Open-source ML for predictions in finance and operations. ¹²⁾

Led by Kozyrkov until her departure, Google's DI team pioneered the discipline internally for strategic and operational choices. The team embedded DI in products like Google Cloud, focusing on scalable frameworks for both reversible low-stakes decisions (ad bidding) and irreversible high-stakes decisions (infrastructure investments). ¹³⁾

Gartner predicts rapid DI platform growth, recognizing them in Market Guides for their decision modeling, execution, governance, and learning features. DIPs address data silos and organizational distrust, with adoption surging in enterprises seeking outcome-focused solutions. Full platforms include collaboration and audit capabilities alongside analytical tools. ¹⁴⁾

DI incorporates behavioral economics to counter cognitive biases like confirmation bias, anchoring, and overconfidence. Kozyrkov's framework prompts explicit bias checks during the option evaluation stage, blending human psychology awareness with data-driven analysis. This integration helps organizations make better decisions by designing systems that account for how humans actually think rather than how they ideally should. ¹⁵⁾

Decision automation fully executes rules and ML models for high-volume, well-defined cases. FICO's credit scoring is a prominent example, where decisions are made entirely by algorithms with human oversight limited to exception handling. ¹⁶⁾

Decision support augments humans with insights and recommendations while preserving human judgment for complex scenarios. Qlik's alert systems and Stravito's insight management exemplify this approach. ¹⁷⁾

The choice between automation and support depends on decision stakes, reversibility, regulatory requirements, and the maturity of available models.

DI engineering treats decisions as code: model them explicitly, orchestrate flows, govern outcomes, and iterate via feedback loops. The discipline spans data engineering, ML operations, and decision operations, fostering composable analytics like scenario modeling and what-if analysis. ¹⁸⁾

• Human-in-the-Loop
• Chief AI Officer
• AI-Native Organization