AI-Enabled Agile Product Management and Emerging Practice

Introduction and Context

Artificial intelligence (AI) is becoming embedded in product discovery, delivery and operations. Product teams now use generative AI to summarize interviews, identify themes, draft backlog items, create prototypes, generate software, produce test cases, detect anomalies and support experimentation. These capabilities can increase speed and broaden analytical capacity. They do not, however, eliminate the need for product judgment, customer contact, domain expertise or accountability.

AI-enabled agile product management is therefore not simply the automation of Scrum activities. It is the deliberate combination of adaptive product practices, machine-assisted analysis and human governance. The central management question is not “Where can AI be inserted?” but “Which decisions can be responsibly augmented, under what controls, and with what evidence?”

AI systems are probabilistic. They may generate plausible but incorrect statements, reproduce historical bias, expose confidential information or produce code with security weaknesses. Agile methods, by contrast, emphasize empirical learning: teams inspect evidence, adapt plans and deliver value incrementally. This makes agile a useful operating context for AI, provided that experimentation is accompanied by explicit risk controls.

Learning Objectives

After completing this chapter, the learner should be able to:

1. Explain how AI can support discovery, backlog management, prototyping, coding, testing and product analytics.
2. Distinguish AI assistance from delegated decision authority.
3. Design responsible workflows involving privacy, bias, intellectual property, security and human oversight.
4. Evaluate AI-supported product work using outcome, quality, flow and risk measures.
5. Apply continuous discovery and product operations practices to AI-enabled products.
6. Assess emerging adaptive delivery models and their implications for product leadership.

1. A Conceptual Model of AI-Enabled Agility

Traditional agile product management consists of a learning loop: understand a problem, form a hypothesis, build a small increment, observe results and adapt. AI can contribute at several points in this loop, but it does not replace the loop itself.

A useful model has four layers:

The layers are connected, but they should not be confused. A model may accurately summarize customer complaints while still being unable to determine which problem is strategically important. A code assistant may produce syntactically valid code while missing a regulatory requirement. A dashboard may identify correlation without establishing causation.

The product manager remains accountable for framing the problem, selecting evidence, making trade-offs, communicating uncertainty and ensuring that the product serves legitimate user and organizational goals.

AI as an Augmenting System

AI augmentation means that a person remains responsible for interpretation and approval while a system performs a bounded task. Examples include proposing alternative acceptance criteria, identifying duplicate backlog items or generating an initial test suite.

Automation is stronger: the system executes a defined action with limited human intervention, such as routing a low-risk support ticket. Delegation occurs when authority over a consequential decision is transferred to the system. Delegation requires substantially stronger evidence, monitoring and governance than augmentation.

A practical classification is:

2. AI Across the Product Development Lifecycle

2.1 Discovery Synthesis

Discovery produces qualitative and quantitative evidence from interviews, observations, surveys, support cases, search data, product telemetry and market research. AI can accelerate transcription, translation, coding of interview segments, clustering of complaints and comparison of themes across time.

A responsible synthesis workflow is:

1. Define the research question and population.
2. Obtain consent and remove unnecessary personal identifiers.
3. Record the source, date, method and limitations of each item.
4. Ask AI to propose summaries or themes, not to manufacture findings.
5. Compare generated themes with source evidence.
6. Examine minority, contradictory and negative cases.
7. Validate interpretations with researchers, domain experts and, where appropriate, participants.
8. Preserve an evidence trail linking conclusions to source material.

AI-generated summaries often overrepresent repeated or strongly worded comments. Frequency is not equivalent to importance: a rare accessibility barrier may matter more than a common minor inconvenience. Product teams should therefore distinguish prevalence, severity, strategic relevance and confidence.

A simple prioritization aid can use a weighted score:

Here, P_i is the priority score for opportunity i, S_i is customer severity, F_i is frequency, R_i is strategic relevance, A_i is evidence strength, and the w terms are agreed weights whose sum is 1. Scores should be treated as structured judgment, not objective truth. The scales must be defined in advance, and teams should test whether changing the weights changes the decision.

2.2 Opportunity and Backlog Support

AI can identify similar backlog items, suggest story decompositions, draft acceptance criteria, detect missing dependencies and translate technical descriptions into user-oriented language. It can also compare a proposed item with product principles, architecture constraints and known defects.

A generated user story is not automatically a good backlog item. The product owner must verify:

• the user or beneficiary is real and sufficiently understood;
• the problem is distinguished from a preferred solution;
• acceptance criteria are observable and testable;
• non-functional requirements are included;
• dependencies, constraints and risks are visible;
• the item is small enough to learn from or deliver within the team’s flow.

AI should not be used to inflate a backlog with large numbers of plausible but unvalidated ideas. A smaller evidence-based backlog is generally more valuable than a large machine-generated inventory.

2.3 Prototyping and Experimentation

Generative systems can create interface alternatives, service blueprints, sample data, conversational flows and low-fidelity code prototypes. This supports rapid comparison of concepts before expensive implementation.

The product manager must distinguish prototype fidelity from evidence quality. A polished interface can create false confidence. Every prototype should state its purpose: for example, testing navigation, explaining a service concept or measuring willingness to perform a task. It should also identify what it does not test, such as performance, accessibility, security or operational feasibility.

AI-generated synthetic users or synthetic data may be useful for exploratory design, but they cannot substitute for representative users and real-world validation. Synthetic data can reproduce the assumptions and biases embedded in its generating process.

2.4 Coding and Technical Delivery

AI coding assistants can suggest functions, explain unfamiliar code, generate documentation, translate between programming languages and propose refactoring. These benefits are strongest when the task is well specified and the code is reviewed by someone competent in the relevant technology.

Important controls include:

• prohibit submission of confidential source code to unapproved services;
• use enterprise settings that define retention and training policies;
• require code review and automated checks;
• scan generated code for vulnerabilities, secrets and license concerns;
• test edge cases rather than only the example supplied in the prompt;
• preserve authorship and provenance records where required;
• avoid accepting code merely because it compiles.

The productivity effect of AI coding tools is not identical to the value effect. Faster code production can increase maintenance burden, defects or architectural inconsistency if review capacity does not grow correspondingly. Teams should measure delivered outcomes and quality, not only lines of generated code or number of accepted suggestions.

2.5 Testing and Quality Assurance

AI can generate unit-test ideas, boundary cases, test data, exploratory test charters, accessibility checks and defect summaries. It can also identify patterns in logs and suggest likely fault locations.

A robust testing strategy remains layered:

1. Unit and component tests for local behavior.
2. Integration tests for interactions and contracts.
3. End-to-end tests for critical user journeys.
4. Security, privacy, accessibility and performance testing.
5. Human exploratory testing for unexpected behavior and usability.
6. Production monitoring and incident learning.

AI-generated tests may mirror the implementation and fail to challenge its assumptions. Test diversity is therefore important. Teams should ask what invalid inputs, unusual users, language variants, accessibility needs and failure conditions have not been represented.

For AI-based product features, testing must also consider model behavior. Relevant dimensions include accuracy, robustness, calibration, refusal behavior, harmful output, fairness across groups, latency, cost and degradation over time. A model can be statistically accurate overall while failing systematically for a smaller population.

2.6 Analytics and Decision Support

AI-supported analytics can forecast demand, classify events, detect anomalies, personalize experiences and estimate likely churn. These tools can improve attention allocation, but predictive power does not establish a managerial prescription.

Product teams should separate:

• descriptive analytics: what happened;
• diagnostic analytics: what may explain it;
• predictive analytics: what may happen;
• prescriptive analysis: what action may be considered.

A prediction such as “this account is likely to cancel” is not evidence that a particular intervention will prevent cancellation. Controlled experiments, quasi-experimental designs or careful qualitative follow-up may be required to estimate causal impact.

Useful product metrics include activation, retention, task success, conversion, time to value, support burden, reliability and customer-reported outcomes. AI initiatives should also be assessed using flow and quality indicators such as cycle time, escaped defects, rework, review time and incident rates.

For a cohort-based retention rate, a common definition is:

R_t is retention at time t, N₀ is the number of users in the starting cohort, N_t is the number of those original users still active at time t, and N_new,t is the number of new users acquired during the period who must be excluded. The formula assumes a clearly defined cohort and activity rule. It should not be used to compare products whose definitions of “active” differ.

3. Responsible AI Governance in Agile Teams

Governance should be integrated into everyday product work rather than treated as a final approval gate. A lightweight but serious governance system includes risk classification, data controls, documentation, review responsibilities, monitoring and escalation.

Bias and Fairness

Bias can enter through sampling, labels, historical decisions, proxy variables, model design, interface choices or deployment context. A model trained on past hiring decisions, for example, may reproduce past exclusion even if protected attributes are removed, because other variables may act as proxies.

Teams should ask:

• Which groups may be underrepresented in the data?
• Who benefits and who bears the risk of error?
• Are error rates or service quality materially different across relevant groups?
• Is the intended use legitimate and proportionate?
• Can users challenge or correct an adverse result?

Fairness has no single universal metric. Different definitions can conflict, especially when base rates differ across groups. Product teams should choose measures in relation to the harm, domain and legal context, and document the rationale rather than claiming that one metric proves fairness.

Privacy and Data Protection

Privacy risk concerns not only data leakage but also inappropriate collection, inference, retention, linkage and reuse. Responsible practices include data minimization, purpose limitation, access controls, encryption, retention schedules, de-identification where appropriate, supplier assessment and transparent user communication.

Removing names does not guarantee anonymity. Unique combinations of location, time, behavior or employment information may re-identify individuals. Teams should avoid sending personal or confidential data to public models unless an approved contractual and technical arrangement exists.

Privacy impact assessments are especially important where AI profiles people, makes decisions about access or eligibility, processes sensitive information or combines datasets for a new purpose.

Intellectual Property and Provenance

AI outputs may incorporate or resemble material from training data, and the legal treatment of generated content varies by jurisdiction and circumstance. Organizations should define acceptable sources, review obligations, attribution requirements and ownership expectations.

Product teams should maintain provenance for important artifacts: the source data, model or tool used, prompt or configuration where material, human edits, validation performed and final approver. Provenance supports licensing review, incident investigation and reproducibility.

Security and Reliability

AI introduces familiar security risks and new ones. Prompt injection may manipulate a system that reads untrusted content. Data poisoning can corrupt training or retrieval sources. Excessive permissions can allow an agent to perform actions beyond its intended scope. Model outputs may expose secrets or create insecure code.

Controls include least-privilege access, input and output filtering, isolation of tools, approval for consequential actions, secrets management, logging, red-team testing and a rapid rollback mechanism. An AI agent should not receive broad production authority merely because it can call an API.

Human Oversight

“Human in the loop” is meaningful only if the human has adequate competence, time, information and authority to disagree with the system. A reviewer who merely clicks approve is not effective oversight.

Oversight design should specify:

• which decisions require review;
• what evidence the reviewer sees;
• how uncertainty and confidence are communicated;
• when escalation is mandatory;
• how disagreements are recorded;
• who can suspend the system;
• how affected users can appeal or obtain correction.

4. Continuous Discovery and Product Operations

Continuous discovery is the ongoing practice of conducting small research activities throughout product development rather than running discovery only at the beginning of a project. A dual-track model often separates discovery activities—understanding opportunities and testing assumptions—from delivery activities—building and releasing validated increments. The tracks interact continuously rather than forming a strict sequence.

AI can support continuous discovery by monitoring support themes, summarizing new research, detecting changes in user behavior and suggesting questions for the next research cycle. However, automated listening can create a narrow view of users who generate digital traces. Teams must deliberately include non-users, dissatisfied users, low-volume segments and people affected by exclusion.

Product operations, or product ops, provides the systems that allow product teams to work consistently at scale. It may include research repositories, taxonomy and instrumentation standards, portfolio reporting, experimentation governance, tool administration, enablement and cross-functional operating rhythms.

AI makes product operations more powerful but also increases the need for data stewardship. A product organization should define a common metric dictionary, event ownership, data quality checks, model inventories, approved tools and incident processes. Without these foundations, AI can make inconsistent data appear authoritative.

A useful operating cadence is:

5. Evaluating AI-Enabled Product Work

Evaluation should cover four dimensions: value, quality, flow and responsibility.

Value

Does the product solve a meaningful problem and improve an outcome that matters to users or the organization? Possible measures include task completion, retention, revenue, cost avoided, time saved, satisfaction and accessibility.

Quality

Does AI-assisted work meet functional, security, performance, reliability and usability expectations? Escaped defects, rollback frequency, incident severity and support contacts can reveal hidden costs.

Flow

Does the team move validated work more effectively? Lead time, work in progress, review time and blocked time are useful, but local optimization should be avoided. Faster delivery of low-value features is not product success.

For average throughput over a period:

The unit might be items per week, but items must be sufficiently comparable for interpretation. Throughput is not a measure of customer value and should not be used alone to rank individuals.

Responsibility

Are privacy, fairness, security, explainability and accountability acceptable for the context? A balanced scorecard can prevent teams from celebrating speed while ignoring harm.

Evaluation should include a baseline. If a team introduces an AI coding tool, it might compare a pre-adoption period with a carefully selected post-adoption period, while accounting for changes in team composition, product complexity and release policy. Simple before-and-after comparisons can mislead because many variables change simultaneously.

6. Emerging Adaptive Delivery Models

AI-Augmented Scrum

In AI-augmented Scrum, the existing accountabilities and events remain, but AI assists preparation and analysis. For example, it may summarize review feedback or identify impediments before a Daily Scrum. The Scrum Team still owns the increment and decisions.

Continuous Delivery with Adaptive Governance

Teams increasingly release small changes through automated pipelines and use feature flags, telemetry and experiments to learn after deployment. Governance shifts from approving every release manually to defining risk-based controls, automated evidence and explicit thresholds for escalation.

Human-Agent Collaboration

An emerging model uses AI agents to perform bounded sequences such as retrieving documentation, drafting a change, running tests and opening a pull request. The agent operates within permissions and a workflow; humans approve consequential actions. This model demands clear interfaces between agent capabilities, team responsibilities and operational controls.

Outcome-Based Product Operating Models

Organizations are moving from feature output targets toward outcomes and evidence. Teams may receive a problem space and measurable objectives rather than a fixed list of features. AI can increase exploration capacity, but it can also increase the number of experiments and therefore the need for disciplined portfolio choices.

Adaptive, Not Autonomous, Delivery

The most credible future is not fully autonomous product development. It is adaptive delivery in which systems continuously sense signals, propose options and execute low-risk tasks, while humans set purpose, interpret context, resolve value conflicts and accept accountability. The division of labor will vary by domain. Healthcare, finance, education, employment and public services require stronger safeguards than low-consequence internal productivity tools.

7. Illustrative Case: AI-Assisted Support Triage at Northstar Learning

Background

Northstar Learning, a fictional subscription education platform, receives approximately 18,000 support requests monthly. Response times are long, and product managers spend substantial time reading tickets to identify recurring product problems. The company proposes an AI system that classifies tickets, drafts responses and identifies possible product defects.

Problem

Early testing shows high overall classification accuracy, but errors are concentrated among users writing in less common languages and among accessibility-related complaints. The system also occasionally includes personal information in prompts sent to an external provider. Support agents report that draft responses sound confident even when the model is uncertain.

Analysis

The product team maps risks across the workflow. Classification is relatively low risk if it only routes tickets and allows correction. Drafted responses are more consequential because inaccurate advice can frustrate learners or cause them to lose access to coursework. Accessibility complaints have elevated importance because misclassification can delay accommodation.

The team establishes separate measures for overall accuracy, per-language performance, accessibility-ticket recall, response time, correction rate and privacy incidents. It also reviews representative samples rather than relying on the aggregate score.

Approach and Solution

Northstar adopts a staged design:

1. The system classifies and routes tickets but does not send messages automatically.
2. Personal identifiers are removed or masked before external processing.
3. A confidence threshold sends uncertain cases directly to trained agents.
4. Agents see the relevant source article and the evidence used for the suggested classification.
5. Accessibility and account-access categories require specialist review.
6. Every correction is logged and used for periodic evaluation, subject to data governance.
7. A dashboard reports performance by language and issue category.
8. The model can be disabled immediately if incident thresholds are exceeded.

The product manager runs a limited pilot and compares it with the existing process. The pilot measures not only average handling time but also reopening rates, complaint escalation, customer satisfaction and fairness-related disparities.

Outcome

The pilot reduces agent triage time and improves visibility into recurring defects. The company does not automate outbound responses because the evidence does not justify that risk. Product discovery improves because classified themes are linked to ticket samples and validated by support specialists. The system is treated as a decision-support service, not an autonomous support employee.

Learning Points

• Aggregate accuracy can conceal serious subgroup failures.
• A staged rollout can capture value without accepting unnecessary risk.
• Human oversight must include authority, evidence and escalation paths.
• Privacy controls must be designed before experimentation with real data.
• Product operations turns a promising tool into a governable capability.

8. Practical Governance Checklist

Before introducing an AI capability, a product team should document:

1. The user problem and expected benefit.
2. The decision or task AI will support.
3. The data sources, purpose and retention period.
4. The affected users and foreseeable harms.
5. The risk classification and applicable law or policy.
6. Human review points and named accountability.
7. Evaluation data, baselines and subgroup checks.
8. Security, access and supplier arrangements.
9. Intellectual-property and provenance requirements.
10. Monitoring, incident response, rollback and retirement criteria.

The checklist should be revisited as the system changes. A low-risk prototype can become high-risk when connected to production data or given authority to act.

Challenges and Limitations

AI-enabled agility has several structural limitations. First, output quality depends on problem framing and data quality; automation can accelerate a poorly chosen direction. Second, generated material can create review overload. Third, tools may encourage shallow discovery by privileging easily captured digital signals. Fourth, vendor dependence can create cost, availability, confidentiality and portability risks. Fifth, productivity gains may be unevenly distributed, affecting skills, roles and professional development.

There are also epistemic limitations. A language model generates likely sequences rather than guaranteed truths. A prediction can be useful without being causal. A dashboard can create measurement confidence without measuring the intended construct. Product leaders must therefore preserve skepticism, triangulate evidence and make uncertainty visible.

Best Practices

Effective AI-enabled agile product management follows these principles:

• Start with a meaningful user or operational problem, not with a fashionable tool.
• Use AI for bounded assistance before considering automation.
• Keep source evidence available and require human validation of important claims.
• Minimize sensitive data and use approved tools and contracts.
• Test performance across relevant user groups and failure conditions.
• Measure outcomes, quality, flow and responsibility together.
• Give humans real authority to reject, correct or suspend AI behavior.
• Maintain provenance, auditability and rollback capability.
• Involve product, engineering, design, security, legal, research and operations stakeholders.
• Review whether the system should be continued, changed or retired.

Chapter Summary

AI can support nearly every stage of agile product management, including discovery synthesis, opportunity analysis, backlog refinement, prototyping, coding, testing, analytics and product operations. Its proper role depends on task risk, reversibility, data sensitivity and the consequences of error.

AI augmentation is not equivalent to autonomous decision-making. Product teams remain responsible for problem framing, evidence quality, trade-offs and outcomes. Responsible practice requires attention to bias, privacy, intellectual property, security, provenance and meaningful human oversight. Evaluation must combine product value with quality, delivery flow and risk indicators.

Continuous discovery and product operations provide the organizational foundations for responsible use. Emerging models—AI-augmented Scrum, adaptive governance, human-agent collaboration and outcome-based delivery—are likely to expand product capacity, but they will not remove the need for judgment. The future of agile product management is best understood as adaptive collaboration between people, data, software and bounded AI systems.

Key Terms

AI augmentation: Use of AI to assist human work while accountability remains with a person or team.

Artificial intelligence: Computational systems that perform tasks associated with human cognitive capabilities, such as prediction, classification or generation.

Bias: A systematic pattern that produces unfair or inaccurate results for particular people, groups or contexts.

Continuous discovery: Ongoing research and experimentation conducted throughout product development.

Data minimization: Collecting and using only the data necessary for a defined purpose.

Drift: Change over time in data, user behavior or model performance that reduces reliability.

Human oversight: Deliberate review, intervention and accountability by competent people.

Intellectual-property provenance: A record of the sources, tools, transformations and approvals associated with an artifact.

Model confidence: An estimate associated with a model output; it is not necessarily the probability that the output is correct.

Outcome: A change in user, business or societal condition resulting from product activity.

Product operations: Systems, standards and services that help product organizations work consistently and effectively.

Prompt injection: An attempt to manipulate an AI system through instructions contained in input data or content.

Risk-based governance: Applying controls in proportion to the likelihood and consequence of harm.

Synthetic data: Artificially generated data designed to resemble characteristics of real data.

Telemetry: Data collected about product use, performance and operational behavior.

Traceability: The ability to connect a decision or output to its sources, process and approver.

Validation: Checking whether an artifact, assumption or product behavior is appropriate and supported by evidence.

Workflow automation: Software execution of defined steps with limited human intervention.

Review Questions

Short Answer Questions

1. What is the difference between AI augmentation and AI delegation?
2. Why can frequency of customer complaints be an insufficient basis for prioritization?
3. Identify three controls relevant to AI-assisted coding.
4. What is the purpose of a model or artifact provenance record?
5. Why should product teams evaluate AI performance across user groups?

Long Answer Questions

1. Explain how AI can support the product development lifecycle from discovery through operations. Include limitations at each stage.
2. Discuss the relationship between continuous discovery, product operations and responsible AI governance.
3. Compare automation, recommendation and analysis-only roles for AI. When is each role appropriate?
4. Explain why aggregate accuracy may be inadequate for evaluating an AI product feature.
5. Evaluate the likely benefits and risks of human-agent collaboration in future adaptive delivery models.

Analytical / Application Questions

1. A team wants to send customer interview transcripts to a public generative AI service for synthesis. Design a responsible workflow, identifying privacy, consent, evidence-validation and provenance controls.
2. An AI coding assistant reduces cycle time by 20% but escaped defects and review time increase. Analyze possible explanations and recommend a measurement and intervention plan.
3. Design a pilot for an AI support-triage system in a regulated service. Specify the risk classification, human review points, outcome metrics, subgroup evaluation and rollback conditions.

Multiple Choice Questions with Answers

Question 1

Which statement best describes responsible AI augmentation in product management?

A. The AI system makes all product decisions without review

B. A person remains accountable while AI performs a bounded assistance task

C. The product backlog is generated without customer evidence

D. AI output is accepted whenever it is grammatically correct

Correct Answer: B. A person remains accountable while AI performs a bounded assistance task

Question 2

Why should AI-generated discovery themes be checked against source evidence?

A. AI cannot process any qualitative information

B. Repeated comments are always irrelevant

C. Generated summaries can omit contradictions or minority experiences

D. Source evidence is legally required in every product decision

Correct Answer: C. Generated summaries can omit contradictions or minority experiences

Question 3

Which activity is an example of product operations?

A. Writing every user story manually

B. Maintaining instrumentation standards and a product metric dictionary

C. Replacing all customer research with synthetic users

D. Allowing each team to define metrics independently

Correct Answer: B. Maintaining instrumentation standards and a product metric dictionary

Question 4

What is the principal limitation of aggregate model accuracy?

A. It cannot be calculated mathematically

B. It may conceal materially worse performance for a subgroup

C. It always overestimates performance by exactly 10%

D. It measures only software delivery speed

Correct Answer: B. It may conceal materially worse performance for a subgroup

Question 5

Which control is most appropriate for an AI agent that can change production systems?

A. Give it unrestricted administrator access

B. Remove all logging to improve speed

C. Use least privilege and require approval for consequential actions

D. Permit changes whenever the agent reports high confidence

Correct Answer: C. Use least privilege and require approval for consequential actions

Question 6

In the retention formula R_t = (N_t - N_new,t)/N₀ × 100\%, why is N_new,t subtracted?

A. To exclude newly acquired users from the original cohort

B. To calculate the number of product defects

C. To convert the result into currency

D. To estimate the model’s confidence

Correct Answer: A. To exclude newly acquired users from the original cohort

Question 7

Which statement about AI-generated code is most accurate?

A. Code that compiles is necessarily secure

B. Generated code removes the need for architectural review

C. Generated code should undergo review, testing and security scanning

D. AI-generated code cannot contain licensing concerns

Correct Answer: C. Generated code should undergo review, testing and security scanning

Question 8

What is a defining feature of continuous discovery?

A. Research occurs only before development begins

B. Small research and learning activities continue throughout development

C. Customer evidence is replaced by stakeholder opinion

D. Discovery and delivery must never interact

Correct Answer: B. Small research and learning activities continue throughout development

Question 9

Which metric category assesses whether an AI initiative improves the intended user or business condition?

A. Outcome

B. Syntax

C. Permission

D. Prompt length

Correct Answer: A. Outcome

Question 10

Why is meaningful human oversight more than a formal approval click?

A. The reviewer must have information, competence, time and authority to disagree

B. The reviewer must accept every model recommendation

C. Human oversight applies only to low-risk tasks

D. The reviewer must rewrite the model from scratch

Correct Answer: A. The reviewer must have information, competence, time and authority to disagree