Decision Intelligence, Optimization and Human–AI Collaboration

Introduction

Business analytics has traditionally been divided into three questions: what happened, what is likely to happen, and what should be done. Descriptive analytics addresses the first question, predictive analytics addresses the second, and prescriptive analytics addresses the third. Decision intelligence integrates these forms of analysis with organizational objectives, constraints, human judgment, rules, causal reasoning, simulation, and implementation processes.

A predictive model may estimate that a customer has a high probability of cancelling a subscription. That estimate does not, by itself, determine whether the firm should offer a discount, assign a service specialist, do nothing, or terminate an unprofitable relationship. The appropriate action depends on costs, benefits, capacity, fairness, regulations, timing, and the likely effects of intervention. Decision intelligence connects prediction to action.

This chapter examines how organizations design decision systems that combine predictive models, optimization, simulation, business rules, causal reasoning, and human oversight. It also considers bounded rationality, automation bias, explainability, accountability, and the conditions under which humans and AI can collaborate effectively.

Learning Objectives

By the end of this chapter, you should be able to:

1. Distinguish descriptive, predictive, causal, and prescriptive analytics.
2. Explain how decision intelligence converts information into action.
3. Formulate simple optimization problems using objective functions and constraints.
4. Select appropriately among rules, optimization, simulation, predictive models, and causal analysis.
5. Explain bounded rationality, automation bias, and the value of human oversight.
6. Evaluate a human–AI decision process using performance, fairness, robustness, and governance criteria.
7. Design a practical decision workflow for a business application.

1. From Prediction to Decision

A decision is a choice among alternatives under conditions of uncertainty and constraint. A useful conceptual sequence is:

Data → information → prediction → causal understanding → alternatives → evaluation → decision → action → feedback

Each stage answers a different question.

Prediction and prescription should not be confused. A model predicting high demand does not establish that increasing price will reduce demand by a specific amount. Causal reasoning is needed to estimate the effect of an intervention. Similarly, an accurate forecast does not automatically yield an optimal schedule because the schedule must respect labor, inventory, contractual, and operational constraints.

A decision can be represented abstractly by an action a chosen from a feasible set A, given information x and uncertain state ω. A decision rule maps information to action:

Here, δ is the decision rule. The rule may be a human policy, a business rule, an optimization algorithm, a machine-learning system, or a combination of these. The quality of a decision rule depends not only on predictive accuracy but also on the value, risk, feasibility, fairness, and consequences of the resulting actions.

2. Decision Intelligence as an Organizational Capability

Decision intelligence is broader than a software product. It is an organizational capability for designing, automating, monitoring, and improving decisions. It includes five connected elements:

1. Decision definition: specifying who or what is being decided, when, and with which alternatives.
2. Evidence generation: collecting data, forecasts, estimates, and expert knowledge.
3. Decision modeling: representing objectives, constraints, uncertainty, rules, and causal effects.
4. Execution: implementing the decision through people, workflows, or information systems.
5. Learning and governance: measuring outcomes, detecting drift, reviewing exceptions, and assigning accountability.

A decision statement should be operationally precise. “Improve customer retention” is an organizational goal, not yet a decision problem. A more precise statement might be: “For each eligible customer at the weekly review point, choose whether to offer a service intervention, a monetary incentive, or no intervention, subject to a weekly budget and fairness requirements.”

The decision statement identifies the decision owner, decision frequency, time horizon, alternatives, constraints, and outcome measures. Poorly defined decisions often cause more harm than technically imperfect models.

Decision quality and expected value

When outcomes are uncertain, a rational benchmark is expected utility. If action a can lead to outcomes y_i with probabilities p_i, expected utility is:

where U(yᵢ | a) is the utility of outcome y_i under action a, and ∑ᵢ pᵢ = 1 with p_i ≥ 0. This formulation assumes that relevant outcomes and probabilities can be represented sufficiently well and that the decision maker can compare utilities. In practice, probabilities may be estimated imperfectly, utilities may conflict across stakeholders, and some consequences—such as dignity or discrimination—are difficult to monetize. Expected utility is therefore a useful benchmark, not a complete moral or managerial theory.

3. Predictive, Causal and Prescriptive Analytics

Predictive analytics

Predictive analytics estimates an unknown or future quantity. A classification model may estimate P(Y = 1 | X = x), the probability that an event occurs given observed features. A forecasting model may estimate future demand, revenue, or machine failure. Predictive performance can be evaluated using measures such as accuracy, precision, recall, calibration, mean absolute error, or root mean squared error, depending on the task.

Prediction is useful when the predicted variable is relevant to the decision and when the relationship remains sufficiently stable. A highly accurate model can still be unsuitable if it predicts an outcome that the organization cannot influence or if its predictions are systematically unfair.

Causal reasoning

Causal reasoning asks what would happen under an intervention. For a treatment or intervention T, the average treatment effect is often written as:

Here, Y(1) is the potential outcome if a unit receives the intervention, Y(0) is the potential outcome if it does not, and E denotes the population average. The formula relies on the potential-outcomes framework. Estimating it requires assumptions such as consistency, adequate treatment definition, and—depending on the study design—randomization or conditional exchangeability, positivity, and appropriate handling of interference and missing data.

A predictive risk score answers “who is likely to experience the outcome?” A causal model answers “who is likely to be helped by the intervention?” These groups need not be the same. For example, customers most likely to churn may include people who will leave regardless of any offer. A targeting strategy should often prioritize expected incremental benefit rather than risk alone.

Prescriptive analytics

Prescriptive analytics evaluates actions using forecasts, causal estimates, objectives, and constraints. It may recommend a price, route, inventory level, staffing plan, credit limit, or treatment policy. Prescriptive systems should make clear whether a recommendation is based on correlation, causal evidence, optimization assumptions, or a combination.

4. Optimization for Business Decisions

Optimization identifies the best feasible solution according to a defined objective. A general mathematical program can be written as:

Here, x is the vector of decision variables, f(x) is the objective function, g_j(x) represents constraint functions, b_j are constraint limits, and X specifies domains such as nonnegative, integer, binary, or continuous values. A solution is feasible if it satisfies all constraints. An optimal solution is a feasible solution with the best objective value according to the stated direction.

The model is only as meaningful as its assumptions. Optimization does not discover the organization’s true objective automatically. It optimizes the objective supplied by the designer, which may omit quality, fairness, resilience, or long-term effects.

Linear programming example

Suppose a bakery produces bread and cakes. Let x₁ be the number of bread units and x₂ the number of cakes. Each bread produces $30 contribution and requires 2 units of flour and 1 labor hour. Each cake produces $50 contribution and requires 4 units of flour and 3 labor hours. The bakery has 40 units of flour and 24 labor hours.

The model is:

The first constraint measures flour units and the second measures labor hours. Consider the corner points of the feasible region. The intercepts are (20, 0) and (0, 8). The intersection of the two resource boundaries satisfies:

From the second equation, x₁ = 24 − 3x₂. Substitution into the first gives 2(24 − 3x₂) + 4x₂ = 40, so 48 − 2x₂ = 40 and x₂ = 4. Therefore x₁ = 12. The objective values at relevant points are:

Under the stated linear assumptions and continuous decision variables, the optimal plan is 20 bread units and zero cakes, with contribution of $600. If production must be in batches, demand limits exist, or cakes create strategic value, the model must be revised. The result is not a universal business recommendation; it is the optimum of this particular formulation.

Common optimization approaches

• Linear programming: linear objective and constraints; suitable for many allocation and blending problems.
• Integer or mixed-integer programming: some or all variables must be integers; useful for vehicle selection, staffing, facility location, and yes/no decisions.
• Nonlinear programming: relationships include nonlinear costs, returns, or physical constraints.
• Network optimization: focuses on flows through nodes and links, such as transportation and distribution.
• Stochastic optimization: incorporates uncertain parameters or scenarios.
• Robust optimization: seeks solutions that perform acceptably across a specified range of uncertainty.
• Heuristics and metaheuristics: produce good practical solutions when exact optimization is computationally difficult.

A model should report not only the recommended solution but also binding constraints, sensitivity to assumptions, infeasibility causes, and alternative near-optimal solutions. A solution that is only marginally better but highly fragile may be inferior to a slightly less profitable and more resilient plan.

5. Rules, Simulation and Scenario Analysis

Business rules

Rules encode explicit organizational policies, such as:

• approve transactions below a delegated authority limit;
• require human review for high-value or unusual claims;
• do not use protected characteristics in a lending decision;
• maintain a minimum safety stock;
• escalate a recommendation when confidence is low.

Rules are transparent and auditable, but they can become rigid, contradictory, or outdated. A rule should have an owner, effective date, rationale, scope, exception process, and review schedule.

Simulation

Simulation represents how a system behaves over time under specified assumptions. It is valuable when randomness, queues, feedback, or complex interactions make direct calculation difficult. Monte Carlo simulation repeatedly samples uncertain inputs and records outcomes. If profit under scenario s is Π_s, the simulated average profit is:

where N is the number of simulated scenarios. The estimate becomes more stable as N increases, but simulation cannot correct an unrealistic model or biased input distributions.

Simulation supports questions such as: How often will a call center exceed a waiting-time target? How will a supply network perform when a supplier fails? What happens if demand variability increases? It is generally a tool for evaluating policies rather than proving that a policy is optimal.

Scenario analysis and stress testing

Scenario analysis evaluates coherent alternative futures, such as high inflation, a cyberattack, or a sudden demand shift. Stress testing deliberately examines severe but plausible conditions. These methods are particularly important when historical data do not contain the events that management must prepare for.

6. Bounded Rationality and Organizational Decision-Making

Classical rational choice assumes that decision makers can identify all alternatives, know relevant probabilities and consequences, and optimize consistently. Herbert Simon’s concept of bounded rationality recognizes that people face limited information, time, attention, computational capacity, and organizational authority. They often satisfice: they select an option that is good enough rather than globally optimal.

Bounded rationality is not simply a personal weakness. It is a property of decision environments. A purchasing manager may not compare every supplier because search is costly. An executive may rely on a dashboard because attention is scarce. A frontline employee may use a rule because rapid action is required.

AI can reduce some cognitive burdens by searching large datasets, generating forecasts, and evaluating scenarios. It can also introduce new burdens: users must interpret model outputs, monitor data quality, challenge recommendations, and manage exceptions. Automation is therefore not a complete solution to bounded rationality. It changes the location and nature of human effort.

Effective systems use selective automation. Routine, low-risk, reversible decisions may be automated. High-impact, ambiguous, novel, or irreversible decisions should usually receive stronger human review. The appropriate level depends on consequence, uncertainty, contestability, and the availability of safe recovery.

7. Human–AI Collaboration

Human–AI collaboration is a sociotechnical arrangement, not merely a person receiving a score. The central design question is how authority, information, responsibility, and intervention are distributed.

Collaboration patterns

1. Human-only: people make the decision without algorithmic recommendation.
2. AI as adviser: AI recommends; an authorized person decides.
3. Human-in-the-loop: human approval is required before execution.
4. Human-on-the-loop: AI operates within boundaries while humans monitor and can intervene.
5. AI automation with escalation: routine cases are automated and exceptions are routed to people.
6. Joint deliberation: humans and AI compare multiple models, scenarios, and rationales.

The best pattern depends on the decision. Requiring manual approval for every low-value transaction can create delay and rubber-stamping. Fully automating a consequential employment or healthcare decision can remove necessary contextual judgment and appeal rights.

Automation bias and overreliance

Automation bias is the tendency to favor an automated recommendation, even when contradictory evidence exists. It can arise from perceived objectivity, time pressure, authority gradients, unfamiliarity with model limitations, or repeated exposure to apparently successful predictions. A related problem is complacency: people monitor less carefully when they believe the system is reliable.

Countermeasures include:

• displaying uncertainty and relevant evidence rather than only a final score;
• requiring reviewers to record an independent assessment before seeing the recommendation in selected contexts;
• using structured challenge prompts;
• training users with realistic failure cases;
• sampling apparently correct automated decisions for audit;
• defining override authority and escalation paths;
• measuring whether overrides improve or worsen outcomes;
• preventing performance targets from rewarding unquestioning acceptance.

Human oversight is meaningful only when reviewers have time, authority, competence, access to evidence, and protection from retaliation when they challenge a system. A nominal “human in the loop” who can only click approve is not effective oversight.

Complementarity rather than replacement

AI tends to be useful at scale, consistency, pattern detection, and rapid comparison. Humans contribute contextual knowledge, moral judgment, negotiation, sense-making under novelty, and responsibility for social consequences. Complementarity is not guaranteed. If AI and humans share the same data errors or biases, combining them may reinforce rather than correct mistakes.

8. Designing a Governed Decision Workflow

A professional decision workflow can be organized as follows:

Step 1: Define the decision and authority

Specify the decision owner, affected parties, action alternatives, timing, and appeal mechanism. Distinguish recommendations from authorized decisions.

Step 2: Identify objectives and constraints

Include financial, operational, legal, ethical, customer, employee, and environmental objectives where relevant. Identify hard constraints and preferences. Check whether the objective encourages harmful proxy behavior.

Step 3: Establish the evidence base

Assess data provenance, quality, representativeness, timeliness, missingness, label quality, and access rights. Separate training, validation, and monitoring data where appropriate.

Step 4: Select the analytical method

Use rules for explicit policies, predictive models for forecasting, causal methods for intervention effects, optimization for constrained choice, and simulation for dynamic uncertainty. Hybrid systems are often appropriate.

Step 5: Test before deployment

Evaluate predictive accuracy, calibration, subgroup performance, feasibility, sensitivity, robustness, security, and failure modes. Conduct scenario and stress testing.

Step 6: Design the user interaction

Present the recommendation, confidence or uncertainty, key evidence, constraints, alternatives, and reasons for escalation. Avoid interfaces that create false precision or conceal disagreement among models.

Step 7: Deploy with controls

Use access controls, logging, versioning, approval thresholds, rollback procedures, and incident-response processes. Establish who can modify rules, data pipelines, and model versions.

Step 8: Monitor outcomes and drift

Monitor data drift, concept drift, performance, fairness, override rates, complaints, downstream effects, and unintended incentives. A model can remain statistically accurate while becoming operationally inappropriate if the decision environment changes.

Step 9: Review and learn

Compare outcomes with objectives, investigate errors, update assumptions, and retire systems that no longer provide sufficient value or cannot be governed safely.

9. Applied Case: Retail Inventory Replenishment

The following is an illustrative case, not a report of a specific real company.

Background

A regional supermarket chain experiences stockouts in fresh products and excessive waste in short-life items. Store managers currently place orders using spreadsheets, recent sales, and personal experience. The company wants an AI-supported replenishment process.

Problem

The organization must choose daily order quantities for each store and product. Demand is uncertain, shelf capacity is limited, deliveries have fixed schedules, and unsold fresh products lose value. The company also wants managers to retain control during promotions, local events, and unusual weather.

Analysis

A demand model estimates a probability distribution rather than a single number. Let D be demand during the replenishment horizon, q the order quantity, p the selling price, c the unit acquisition cost, and s the salvage value of leftover inventory. A simplified expected-profit expression is:

where (q−D)⁺ = max(q−D,0). The model assumes that all demand is either served from available stock or lost, that the unit values are known, and that the demand distribution is appropriate. Real systems may need to include substitutions, backorders, spoilage timing, service-level requirements, and promotion effects.

Prediction alone is insufficient. The chain must also account for truck capacity, warehouse availability, minimum order quantities, and managerial overrides. A causal analysis may be needed to estimate the effect of promotions on demand rather than relying solely on historical correlations.

Approach and solution

The chain implements a hybrid workflow:

1. A forecasting model estimates demand and uncertainty using sales history, promotions, holidays, weather, and store-level features.
2. A causal evaluation estimates whether promotional displays produce incremental demand.
3. An optimization model recommends order quantities subject to capacity, delivery, and minimum-order constraints.
4. Business rules prevent orders above shelf capacity and flag products with unreliable forecasts.
5. Managers see the recommendation, demand range, expected waste and stockout trade-off, and the main drivers.
6. Managers may override the recommendation by selecting a reason such as local event, supplier issue, or data error.
7. High-risk exceptions are escalated to regional planners.

The system does not treat every override as an error. Overrides are analyzed to identify local information that the model lacks. At the same time, the company monitors whether managers systematically override recommendations for particular stores or products, which may indicate model bias, poor usability, or inappropriate incentives.

Outcome

In a pilot, the chain evaluates stockout rate, waste, gross margin, service level, forecast calibration, override quality, and employee workload. It does not judge success using forecast accuracy alone. A model with slightly lower forecast error may be less valuable if it recommends infeasible orders or increases waste.

Learning points

The case demonstrates that decision intelligence requires integration. Forecasting supports the decision, optimization respects constraints, rules encode policy, causal reasoning improves intervention estimates, simulation tests uncertainty, and human oversight contributes local knowledge and accountability. The system’s success depends on workflow and governance as much as on model architecture.

10. Ethical, Legal and Managerial Issues

Decision systems can distribute opportunities, costs, risks, and explanations. Ethical review should therefore ask:

• Who benefits and who bears the risk?
• Are affected groups represented in data and evaluation?
• Is the decision necessary and proportionate?
• Can a person understand, challenge, and appeal the result?
• Are sensitive attributes used lawfully and responsibly?
• Does the system create disparate error rates or unequal access?
• Is personal data collected and retained only for legitimate purposes?
• Can responsibility be assigned when a recommendation causes harm?

Legal requirements vary by jurisdiction and sector. Organizations should maintain records of purpose, data sources, model versions, validation, decisions, human interventions, incidents, and retention. Privacy, employment, consumer protection, financial regulation, safety law, and emerging AI regulation may all apply.

Fairness is not a single metric. Possible measures include demographic parity, equalized error rates, calibration, individual similarity, and outcome-based assessments. These criteria can conflict, especially when base rates differ. The organization must select criteria according to the decision context, legal requirements, and ethical justification rather than presenting one metric as universally decisive.

11. Limitations and Contemporary Developments

Decision intelligence faces several enduring limitations:

• Model uncertainty: estimates may be wrong even when validation performance is strong.
• Distribution shift: future conditions may differ from historical data.
• Objective misspecification: optimizing a measurable proxy can damage the real goal.
• Feedback loops: decisions change the data later used for learning.
• Strategic response: customers, employees, or competitors may adapt to the system.
• Correlation without intervention validity: predictive relationships may not support action.
• Automation complacency: human review may become superficial.
• Computational and organizational constraints: the theoretically best solution may be too slow or difficult to implement.

Contemporary systems increasingly combine machine learning with optimization, retrieval of organizational knowledge, simulation, digital twins, and generative interfaces. These developments may improve accessibility, but they also create risks of fabricated explanations, inconsistent recommendations, hidden assumptions, and unclear accountability. Generative AI should generally be used as an interface or analytical assistant only within controlled workflows, with source verification, permission controls, and human review for consequential outputs.

Best Practices

1. Start with a clearly defined decision, not with a fashionable model.
2. Separate prediction from causal claims and from optimization.
3. Make objectives, constraints, assumptions, and trade-offs explicit.
4. Use uncertainty ranges and scenarios rather than false point precision.
5. Include feasibility, fairness, resilience, and long-term effects in evaluation.
6. Automate proportionately to risk and reversibility.
7. Give reviewers authority, time, evidence, and training to challenge AI.
8. Log recommendations, overrides, outcomes, and model versions.
9. Monitor both technical performance and organizational consequences.
10. Treat governance as a continuing operating process, not a one-time approval.

Chapter Summary

Decision intelligence connects data and analysis to organizational action. Descriptive, predictive, causal, and prescriptive analytics answer different questions and should not be treated as interchangeable. Prediction estimates what may happen; causal reasoning estimates the effect of intervention; optimization selects among feasible alternatives; simulation examines uncertain and dynamic outcomes; and rules encode explicit policies.

Optimization requires a defined objective, decision variables, constraints, and assumptions. Its result is optimal only relative to the model. Simulation and stress testing help evaluate robustness, while sensitivity analysis reveals dependence on uncertain parameters. Hybrid decision systems are often more useful than a single algorithm.

Human–AI collaboration must address bounded rationality and automation bias. AI can expand analytical capacity, but humans remain essential for context, values, accountability, exception handling, and contestability. Effective oversight requires genuine authority and competence, not merely nominal approval.

Responsible decision intelligence includes problem definition, evidence assessment, model selection, validation, deployment controls, monitoring, documentation, and learning. Its success should be judged by decision outcomes, feasibility, fairness, resilience, and stakeholder impact—not by predictive accuracy alone.

Key Terms

Decision intelligence: An organizational capability for designing, executing, monitoring, and improving data-informed decisions.

Predictive analytics: Analysis that estimates future or unknown outcomes from available information.

Causal reasoning: Analysis of how an intervention changes an outcome relative to a suitable counterfactual.

Prescriptive analytics: Analysis that recommends actions using objectives, evidence, uncertainty, and constraints.

Decision rule: A mapping from available information to an action.

Objective function: A mathematical representation of the quantity an optimization model seeks to maximize or minimize.

Constraint: A restriction defining which decisions are feasible.

Feasible solution: A solution satisfying all specified constraints.

Linear programming: Optimization with linear objective and constraint expressions.

Mixed-integer programming: Optimization in which some variables are required to take integer or binary values.

Simulation: Representation of system behavior under specified assumptions and uncertain inputs.

Scenario analysis: Evaluation of decisions across coherent alternative situations.

Bounded rationality: Decision-making limited by information, time, cognitive, and computational constraints.

Satisficing: Selecting an option that meets an acceptable threshold rather than optimizing globally.

Automation bias: The tendency to favor or accept an automated recommendation despite contrary evidence.

Human-in-the-loop: A design in which human approval is required before an AI-supported action is executed.

Model drift: A decline in suitability caused by changes in data, relationships, or the operating environment.

Robustness: The ability of a decision or model to perform acceptably under variation and stress.

Override: An authorized human departure from an automated recommendation.

Decision governance: Structures and processes that assign accountability, control risk, and monitor decision systems.

Review Questions

5 Short Answer Questions

1. What is the difference between predictive and prescriptive analytics?
2. Why does a predictive model not necessarily identify the best intervention?
3. Define a feasible solution in an optimization problem.
4. What is automation bias?
5. State two conditions required for meaningful human oversight.

5 Long Answer Questions

1. Explain how descriptive, predictive, causal, and prescriptive analytics can be integrated in a business decision workflow.
2. Formulate and interpret a simple resource-allocation optimization problem, including its objective function and constraints.
3. Discuss the contribution and limitations of simulation and scenario analysis in business planning.
4. Explain bounded rationality and analyze how AI may reduce and also intensify its effects.
5. Discuss the ethical and governance requirements for a high-impact automated decision system.

3 Analytical / Application Questions

1. A bank uses a machine-learning model to predict loan default. Explain why default prediction alone is insufficient for deciding whom to offer a repayment modification. Identify the additional causal, optimization, fairness, and governance considerations.
2. Design a human–AI workflow for hospital appointment scheduling that accounts for uncertainty, capacity constraints, emergency exceptions, and human oversight.
3. An organization reports that its AI recommendations are accurate but that employees almost never override them. Develop a diagnostic plan to determine whether this indicates excellent performance, automation bias, poor interface design, or inadequate oversight.

Multiple Choice Questions with Answers

Question 1

Which type of analytics primarily answers the question, “What should we do?”

A. Descriptive analytics

B. Predictive analytics

C. Prescriptive analytics

D. Diagnostic analytics

Correct Answer: C. Prescriptive analytics

Question 2

What does the constraint in an optimization model primarily define?

A. The historical accuracy of the data

B. The set of feasible decisions

C. The probability of every outcome

D. The user interface design

Correct Answer: B. The set of feasible decisions

Question 3

A model estimates that a customer is likely to churn. What additional analysis is most directly needed to determine whether a discount will reduce that customer’s probability of churning?

A. Causal analysis

B. Descriptive reporting

C. Data compression

D. Random password generation

Correct Answer: A. Causal analysis

Question 4

In the bakery example, what is the contribution of producing 20 bread units and zero cakes?

A. $400

B. $500

C. $560

D. $600

Correct Answer: D. $600

Question 5

Which statement best describes simulation?

A. It always proves that a policy is mathematically optimal.

B. It represents system behavior under specified assumptions and uncertain inputs.

C. It replaces the need to define objectives and constraints.

D. It can use only deterministic data.

Correct Answer: B. It represents system behavior under specified assumptions and uncertain inputs.

Question 6

Bounded rationality suggests that decision makers often:

A. Have unlimited information and computational capacity

B. Choose randomly without objectives

C. Select satisfactory options under practical limitations

D. Always follow optimization results

Correct Answer: C. Select satisfactory options under practical limitations

Question 7

Which situation is most likely to require stronger human review?

A. A reversible, low-value routine reorder

B. A high-impact decision with ambiguous evidence and limited appeal

C. A fully specified formatting transformation

D. A calculation with verified inputs and no discretion

Correct Answer: B. A high-impact decision with ambiguous evidence and limited appeal

Question 8

What is automation bias?

A. The tendency to reject every algorithmic recommendation

B. The tendency to favor an automated recommendation despite contrary evidence

C. The technical failure of a computer processor

D. The use of automation only for repetitive work

Correct Answer: B. The tendency to favor an automated recommendation despite contrary evidence

Question 9

Which practice most directly supports meaningful human oversight?

A. Allowing reviewers to approve recommendations without seeing evidence

B. Removing all override capabilities

C. Providing authority, time, evidence, and training to challenge recommendations

D. Measuring success only by the number of accepted recommendations

Correct Answer: C. Providing authority, time, evidence, and training to challenge recommendations

Question 10

A decision system performs well on historical data but poorly after market conditions change. What problem is most likely involved?

A. Model or concept drift

B. A binding constraint

C. A perfectly calibrated model

D. A fixed business rule

Correct Answer: A. Model or concept drift