5 Role of Data in Decision Making

5.1 Why Decisions Need Data

Decisions are the ultimate output of any organization. Every other activity (capturing transactions, building pipelines, training models) is justified only insofar as it improves the quality of those decisions. The role of data is to replace or supplement intuition with evidence, so that a decision can be defended at the time it is made and improved the next time it is made.

Decisions, data and outcomes

A decision is a commitment of resources. Data reduces the uncertainty around the likely outcome of that commitment. When the commitment is small and reversible, intuition may be adequate; when it is large and one-way, data becomes indispensable.

Three forces that raised the stakes

Three shifts since 2000 have made data-driven decision making the default in large organizations. First, transactional and digital systems now record most activity, so the marginal cost of additional evidence is low. Second, competitive cycles have shortened, so the cost of a wrong decision has risen. Third, regulatory environments (RBI for banking, SEBI for capital markets, the DPDP Act for personal data) increasingly require documented, auditable decision trails.

5.2 Anthony’s Hierarchy of Decisions

Robert Anthony’s 1965 classification remains the most widely used decomposition of managerial decisions. The three levels differ in time horizon, scope, reversibility, and the type of data required.

Strategic decisions

Long-horizon, firm-wide commitments. Examples: entering a new market (Tata Motors’ entry into the UK commercial-vehicle segment), acquiring a competitor, redesigning the operating model. The data used is largely external and forward-looking (market size, macroeconomic indicators, competitor positioning).

Tactical decisions

Medium-horizon, functional commitments. Examples: setting the price tiers for a new product line, deciding which stores to staff up for Diwali, choosing a logistics partner for a region. The data is a mix of internal performance history and external benchmarks.

Operational decisions

Short-horizon, transaction-level commitments. Examples: approving a credit-card transaction, matching a rider to a driver, stocking a warehouse bay. The data is internal, transactional, and often required in real time. Operational decisions are where analytics pays off first, because converting routine judgements into automated rules scales immediately. This is the pattern behind automated credit-card approvals at HDFC and automated dispatch at Ola and Swiggy.

Level	Horizon	Example	Data needed	Frequency
Strategic	Years	Enter a new market	External, aggregated, forward-looking	Once or rarely
Tactical	Months to quarters	Set pricing tiers	Internal history + external benchmarks	Periodic
Operational	Minutes to days	Approve a transaction	Internal, transactional, real-time	Continuous

5.3 Simon’s Decision-Making Process

Simon’s (1960) model describes the sequence through which any non-trivial decision passes, independent of its content.

flowchart LR
  A[Intelligence<br/>detect the problem] --> B[Design<br/>generate alternatives]
  B --> C[Choice<br/>pick one]
  C --> D[Implementation<br/>act and track]
  D --> E[Review<br/>learn and feed back]
  E -.-> A

Intelligence

Scanning the environment to detect situations that call for a decision. Examples: noticing a spike in call-centre escalations, detecting a drop in conversion, spotting an unusual transaction pattern. Data supports this stage through monitoring dashboards and anomaly-detection systems.

Design

Generating candidate alternatives. Modelling, what-if analysis, scenario construction, and simulation all live here. Data supports this stage by quantifying the likely consequences of each alternative.

Choice

Comparing the alternatives against criteria and selecting one. Data supports this stage through objective scoring, sensitivity analysis, and explicit risk-return trade-offs.

Implementation and Review

Executing the chosen alternative, tracking its effects, and feeding the result back into the next cycle. This is the stage most often neglected in practice and most often responsible for apparent “model failures” that are actually execution failures.

5.4 The Four Types of Analytics

A single taxonomy now dominates every industry conversation about analytics. Each type answers a different question and builds on the previous one.

Descriptive analytics: what happened

Summarises the past. Includes sums, averages, percentages, frequency distributions, dashboards, and standard reports. Typical outputs: monthly revenue by region, customer churn rate, first-call-resolution percentage. R functions: summary(), mean(), table(), ggplot2 charts.

Diagnostic analytics: why did it happen

Explains the descriptive pattern. Includes drill-down, root-cause analysis, cohort comparisons, and correlational analysis. Typical outputs: attribution of a revenue decline to a specific region and channel, identification of the driver behind a churn spike.

Predictive analytics: what will happen

Projects forward using statistical or machine-learning models. Includes regression, classification, time-series forecasting, and survival analysis. Typical outputs: next-quarter demand forecast, probability of customer churn, expected loss on a loan portfolio.

Prescriptive analytics: what should we do

Produces recommended actions by combining predictions with optimisation or simulation. Includes linear and integer programming, Monte Carlo simulation, reinforcement learning. Typical outputs: the price that maximises expected margin subject to inventory constraints, the staffing plan that minimises wait times under projected demand.

The progression is not a hierarchy of importance

Descriptive analytics is sometimes dismissed as “merely reporting”, but a large share of business value still comes from accurate descriptive reporting that people actually read. Predictive and prescriptive methods are powerful only when the underlying descriptive facts are clean and trusted.

5.5 Data Across the Decision Lifecycle

Different analytics types support different stages of Simon’s process.

Simon stage	Primary analytics type	Typical artefact
Intelligence	Descriptive + diagnostic	Monitoring dashboard, anomaly alert
Design	Predictive	Forecast, scenario model
Choice	Prescriptive	Optimisation or simulation output
Implementation	Descriptive (real-time)	Operational dashboard, log stream
Review	Diagnostic + predictive	Post-mortem, recalibrated model

5.6 Evidence-Based Management

Evidence-based management (Pfeffer and Sutton 2006; Barends, Rousseau and Briner 2014) argues that consequential managerial decisions should triangulate four sources of evidence.

The four evidence streams

Scientific evidence from peer-reviewed research. 2. Organizational data from the firm’s own systems. 3. Stakeholder values and concerns from employees, customers, regulators. 4. Practitioner expertise from the judgement of experienced decision-makers.

Why triangulation matters

No single stream is sufficient. Scientific evidence may not generalise to the firm’s context, organizational data may be biased by past practice, stakeholder views may conflict, and practitioner expertise may simply encode past mistakes. Combining the four reduces the chance that any one blind spot dominates.

5.7 Cognitive Biases Data Can and Cannot Fix

Kahneman (2011) catalogues the systematic errors that dominate intuitive judgement. Data can correct some and amplify others.

Biases data can usually correct

Availability bias (recent or vivid examples dominate): data forces attention to base rates. Confirmation bias (we seek evidence that supports our prior view): a complete summary of the evidence, including disconfirming cases, weakens the pull. Anchoring (the first number dominates): structured comparison across alternatives reduces the effect.

Biases data can amplify

Sampling bias: a dataset that is unrepresentative of the population (survivorship, self-selection) will produce systematically wrong conclusions with apparent precision. Algorithmic bias: a model trained on historical decisions will inherit the biases of those decisions. Goodhart’s Law: once a metric becomes a target, it ceases to be a good measure (gaming the KPI replaces achieving the underlying goal).

The implication for managers

Using data well requires the same scepticism about data that experienced managers bring to opinion. “The model says so” is not, by itself, a complete argument. A defensible decision states what the data shows, what it does not show, and what would change the conclusion.

5.8 Decision Support Systems

The tools that deliver data to decision makers span a spectrum from passive reporting to active automation.

Class	Purpose	Examples
Business intelligence	Dashboards, standard reports, self-service exploration	Power BI, Tableau, Qlik, Looker, ThoughtSpot
Predictive analytics platforms	Model development and deployment	Databricks, Dataiku, DataRobot, AWS SageMaker
Optimisation and simulation	Solving for the best decision under constraints	Gurobi, IBM CPLEX, AnyLogic, Simul8
Decision-intelligence platforms	Integrated pipeline from data to recommended action	Palantir Foundry, RapidMiner, Google Vertex AI

Indian industry context

Most listed Indian firms run a BI stack (typically Power BI or Tableau) for descriptive and diagnostic work; a smaller set (large banks, telecoms, e-commerce platforms) additionally run ML platforms in production. Fully prescriptive systems remain rare outside supply chain, pricing, and fraud domains.

5.9 A Reference Decision-Analytics Flow

flowchart TD
  A[Source data<br/>transactions, logs, sensors] --> B[Preparation<br/>clean, join, transform]
  B --> C1[Descriptive<br/>what happened]
  B --> C2[Diagnostic<br/>why it happened]
  B --> C3[Predictive<br/>what will happen]
  B --> C4[Prescriptive<br/>what to do]
  C1 --> D[Decision makers]
  C2 --> D
  C3 --> D
  C4 --> D
  D --> E[Action]
  E --> F[Outcomes]
  F -.-> A

Each of the four analytics types plugs into the decision step; the outcomes of the action flow back into the source data and feed the next cycle. The quality of the loop, not the sophistication of any single stage, determines how much the organization actually learns.

5.10 Summary

Summary of data concepts introduced in this chapter
Concept	Description
Decision Levels
Strategic	Long-horizon, firm-wide commitments
Tactical	Medium-horizon, functional commitments
Operational	Short-horizon, transaction-level commitments
Simon's Stages
Intelligence	Detect the problem worth solving
Design	Generate candidate alternatives
Choice	Select one alternative against criteria
Implementation	Execute and track the chosen action
Review	Learn from the outcome and feed back
Analytics Types
Descriptive	Summarise what happened
Diagnostic	Explain why it happened
Predictive	Project what will happen next
Prescriptive	Recommend what action to take
Decision Biases
Availability	Recent or vivid examples dominate judgement
Confirmation	Preference for evidence that confirms prior beliefs
Anchoring	The first number seen exerts undue pull
Sampling	Unrepresentative data produces confidently wrong answers

Data earns its place in decision making when it changes what a manager would otherwise have done. The rest of this book is a set of tools for producing data of that standard: clean enough to trust, summarised well enough to read, modelled well enough to forecast, and optimised well enough to act on.