Enterprise AI Strategy and Digital Transformation

Introduction

Artificial intelligence (AI) creates value not merely by automating isolated tasks, but by changing how organisations make decisions, deliver services, allocate resources and compete. Enterprise AI strategy is therefore a management and engineering discipline concerned with aligning AI capabilities with organisational purpose, risk tolerance, operating processes and measurable outcomes.

For private firms, objectives may include revenue growth, customer retention, productivity, resilience and innovation. For public-sector organisations, objectives may include equitable service delivery, regulatory compliance, fiscal responsibility, public safety and citizen trust. In both settings, successful AI adoption requires more than selecting a model or purchasing a platform. It requires a portfolio of use cases, an operating model, appropriate data and technology architecture, skilled people, redesigned processes, governance and sustained change leadership.

Digital transformation is broader than digitisation. Digitisation converts analogue information into digital form; digitalisation uses digital tools to improve existing processes; digital transformation redesigns capabilities, structures and value creation around digital technologies. AI can accelerate transformation, but it can also reproduce inefficient processes, amplify bias or introduce unacceptable risks if deployed without institutional controls.

Learning Objectives

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

1. Explain the relationship between AI strategy, business strategy and digital transformation.

2. Identify and prioritise enterprise and public-sector AI use cases.

3. Evaluate build, buy and partner alternatives.

4. Design appropriate AI operating models, talent structures and governance mechanisms.

5. Assess AI investments using financial, operational, social and risk-adjusted measures.

6. Explain how process redesign and change leadership influence AI outcomes.

7. Critically analyse an enterprise AI transformation case.

1. From AI Projects to Enterprise Strategy

An AI project is a bounded intervention, such as a fraud detector or document-classification service. An AI strategy is a coherent set of choices concerning where AI will be used, what capabilities the organisation will develop, how risks will be controlled and how investments will be sequenced. A portfolio perspective is essential because individual projects compete for data, talent, funding and executive attention.

A useful strategic chain is:

Organisational objective → business capability → process or decision → AI opportunity → measurable outcome → governance control.

For example, a public health agency may seek to reduce avoidable hospital admissions. This objective may depend on the capability to identify high-risk patients and coordinate interventions. The relevant process may involve analysing clinical and social data, while the AI opportunity may be a risk-stratification system. Outcomes could include reduced admissions and improved intervention timeliness. Governance must address clinical accountability, privacy, explainability and unequal performance across population groups.

AI strategy should be evaluated at three levels:

• Strategic fit: Does the initiative support the organisation’s mission and competitive or public-value priorities?
• Operational feasibility: Are data, processes, skills, infrastructure and responsible owners available?
• Responsible acceptability: Are legal, ethical, security, safety and social risks within the organisation’s tolerance?

A technically accurate model can still be strategically unsuccessful. Common causes include solving a low-value problem, producing predictions that no employee can act upon, failing to integrate outputs into workflow, or measuring model accuracy without measuring organisational outcomes.

2. Understanding Value in Enterprise AI

AI value is multidimensional. Financial value may arise from additional revenue, reduced operating cost, avoided loss or improved asset utilisation. Non-financial value may include faster service, improved safety, better accessibility, employee capability and public trust. Conversely, costs may include software, cloud infrastructure, data preparation, integration, training, process redesign, monitoring and regulatory compliance.

A basic investment model is:

\[

\text{Net Present Value (NPV)}=\sum_{t=0}^{T}\frac{CF_t}{(1+r)^t}

\]

where \(CF_t\) is the net cash flow in period \(t\), \(r\) is the discount rate expressed per period, and \(T\) is the evaluation horizon. The initial investment is normally represented by a negative cash flow at \(t=0\). NPV is applicable when expected cash flows and a suitable discount rate can be estimated. It does not, by itself, capture distributional effects, safety risks or public value.

A simpler accounting measure is return on investment:

\[

\text{ROI}=\frac{\text{Total benefits}-\text{Total costs}}{\text{Total costs}}

\]

ROI is useful for communicating an estimate, but it can be misleading if benefits occur over different time periods, if risk is ignored, or if intangible outcomes are excluded. For AI, a credible business case should distinguish:

• Gross benefit: the theoretical improvement if the system works as expected.
• Realised benefit: the improvement actually captured after adoption and process constraints.
• Net benefit: realised benefit less implementation, operation, governance and remediation costs.

The difference between gross and realised benefit is often substantial. A forecasting model might improve prediction accuracy, yet inventory costs may not decline because procurement policies, supplier contracts or warehouse capacity remain unchanged.

3. Use-Case Discovery and Prioritisation

Use-case discovery should begin with strategic pain points rather than fashionable technologies. Sources include customer complaints, process-performance data, audit findings, service delays, employee workarounds, fraud patterns, safety incidents and executive priorities. Interviews and observation are important because process documentation often omits informal practices on which successful work depends.

A candidate use case should specify:

1. The decision or task to be improved.

2. The users and accountable decision-makers.

3. The data required and its provenance.

4. The expected intervention or action.

5. The baseline process and performance.

6. The outcome metric and time horizon.

7. Potential harms, affected groups and controls.

8. Integration and change requirements.

Prioritisation dimensions

A practical portfolio assessment can score candidates on strategic value, feasibility, time to value, risk and reusability. Scores should be evidence-based rather than treated as objective facts. A high-value use case may be rejected if data quality is inadequate or if the organisation cannot safely act on predictions.

One illustrative scoring approach is:

\[

P=w_VV+w_FF+w_TT+w_RR-w_KK

\]

where \(P\) is a prioritisation score, \(V\) is value, \(F\) is feasibility, \(T\) is time-to-value, \(R\) is reusable capability, \(K\) is risk, and the \(w\) terms are explicitly chosen weights. All component scores must use a defined scale, such as 1 to 5, and the weights should sum to 1 if the risk term is normalised consistently. This is a decision aid, not a scientific measurement of value. Sensitivity analysis should test whether rankings change when weights vary.

Portfolio balance matters. An organisation may combine quick, low-risk automation with longer-term strategic initiatives. It should avoid a portfolio consisting only of pilots, because pilots without adoption generate learning but not transformation. It should also avoid concentrating all investment in one uncertain, high-impact programme.

Public-sector prioritisation

Public organisations should assess public value, legality, fairness, accessibility and institutional legitimacy in addition to financial efficiency. A system that reduces processing time but systematically disadvantages a protected group may be unacceptable. Consultation with affected communities and independent oversight may be necessary before deployment.

4. Build, Buy or Partner

The build-buy-partner decision concerns both the AI model and the surrounding capability. An organisation may buy a general-purpose model, build a specialised model, partner with a system integrator, or use a hybrid arrangement. The relevant object of analysis includes data pipelines, application programming interfaces, security controls, monitoring, support, intellectual property and exit options.

Build

Building is attractive when the use case is strategically differentiating, data is proprietary, requirements are highly specialised, or control and customisation are critical. It can develop internal competence and reduce dependence on a supplier. However, it requires scarce talent, sustained maintenance and responsibility for security, evaluation and compliance.

Buy

Buying can accelerate deployment and provide mature functionality, service-level commitments and vendor support. It is suitable when the capability is relatively standardised and does not create unique competitive advantage. Risks include vendor lock-in, opaque model behaviour, changing prices, data-use restrictions, limited customisation and concentration risk.

Partner

Partnerships combine organisational knowledge with external technical expertise. They are useful where implementation complexity is high or internal capability is developing. Contracts should address data ownership, confidentiality, audit rights, model updates, incident response, performance thresholds, subcontractors, portability and termination assistance.

The decision should consider total cost of ownership rather than acquisition price. A low-cost service may become expensive when integration, human review, monitoring, retraining, security assessment and migration are included. Strategic decisions should also examine whether the organisation is outsourcing a commodity capability or surrendering knowledge central to its mission.

5. The AI Operating Model

An operating model defines how work, authority and accountability are arranged. Three broad patterns are common.

Centralised model

A central AI unit owns most data science, engineering, standards and delivery. This supports consistency, reuse and strong governance, but can become distant from operational users and create a bottleneck.

Federated model

Business units maintain domain teams while a central function provides platforms, standards and assurance. This improves local responsiveness but requires mature coordination and architecture.

Hub-and-spoke model

A central hub supplies shared infrastructure, methods, talent development and governance; spokes embed specialists in business units. This often balances enterprise control with domain knowledge.

No model is universally superior. The choice depends on organisational size, regulatory exposure, data maturity, strategic urgency and the degree of similarity among use cases. Regardless of structure, accountability must be explicit. A model developer should not be the sole authority deciding whether a high-impact system is acceptable for deployment.

An effective operating model commonly includes:

• An executive sponsor responsible for strategic alignment.
• A portfolio or AI investment committee.
• Product owners accountable for user outcomes.
• Data owners responsible for quality, access and stewardship.
• Data scientists and machine-learning engineers.
• Software, cloud, cybersecurity and platform engineers.
• Domain specialists and process owners.
• Legal, privacy, risk, compliance and ethics specialists.
• Change, training and service-design professionals.
• Independent assurance for high-risk systems.

6. Data, Architecture and Model Lifecycle

Enterprise AI depends on data that is relevant, lawful, sufficiently representative and available at the time decisions are made. Data governance should address ownership, classification, quality, lineage, retention, access, consent or legal basis, and permitted use. Data quality is multidimensional: accuracy, completeness, timeliness, consistency, validity and uniqueness may matter differently for different applications.

A production architecture often includes source systems, ingestion, storage, preparation, feature or retrieval services, model serving, application integration, monitoring and feedback. The architecture should support versioning and reproducibility. A model that cannot be linked to the data, code, parameters and evaluation results used to create it is difficult to audit or safely maintain.

The lifecycle typically includes:

1. Problem definition and impact assessment.

2. Data acquisition and preparation.

3. Baseline selection, such as a rule-based or human process.

4. Model development and validation.

5. User and security testing.

6. Controlled release and monitoring.

7. Periodic review, retraining or retirement.

Model performance is not constant. Data drift, concept drift, changes in user behaviour and policy changes may reduce effectiveness. Monitoring should therefore include technical metrics, outcome metrics, subgroup performance, human override patterns, incidents and unexpected uses.

For generative AI, additional controls may be necessary for hallucination, prompt injection, confidential-data disclosure, copyright concerns, unsafe content and inappropriate automation of authoritative advice. Retrieval augmentation can improve access to approved information, but it does not guarantee factuality or eliminate the need for evaluation.

7. Process Redesign and Human Work

AI rarely creates value when inserted into an unchanged process without examining roles and decision rights. Process redesign asks which activities should be automated, augmented, eliminated or retained for human judgement. It also examines exception handling, escalation, accountability and the consequences of errors.

A useful distinction is between:

• Human-in-the-loop: a person must approve or intervene before an action occurs.
• Human-on-the-loop: a system acts with monitoring and the ability for a person to intervene.
• Human-out-of-the-loop: no routine human intervention occurs.

The appropriate pattern depends on risk, reversibility, speed and the quality of available oversight. A low-risk email classification task may tolerate extensive automation; a medical, employment or liberty-affecting decision requires stronger safeguards and meaningful human review.

Human review is not automatically protective. If reviewers face excessive workload, unclear guidance or automation bias, they may approve system outputs without genuine scrutiny. Effective redesign therefore includes training, workload analysis, interface design, override procedures and auditability.

8. Talent and Change Leadership

AI transformation requires interdisciplinary capability rather than data scientists alone. Technical skills include machine learning, software engineering, data engineering, cloud operations, cybersecurity and evaluation. Organisational skills include product management, process analysis, service design, finance and procurement. Responsible-AI capability includes privacy, fairness, safety, accessibility, law and ethics.

Organisations can acquire capability through hiring, reskilling, universities, professional communities and carefully governed partners. A mature talent strategy defines critical roles, career pathways, competency expectations and retention mechanisms. It also recognises that domain experts are essential: they understand the meaning, consequences and operational limits of predictions.

Change leadership should communicate:

• Why the change is necessary.
• Which problems are being solved.
• How roles and accountability will change.
• What will not change.
• How employees and affected communities can provide feedback.
• How failures and concerns will be handled.

Resistance is not always irrational opposition. It may identify unsafe assumptions, unrealistic productivity targets, loss of professional autonomy or poor system usability. Leaders should treat responsible challenge as a source of risk detection. Adoption measures may include active usage, task completion, user trust calibrated to actual performance, training completion, exception rates and realised outcomes.

9. Governance, Ethics and Regulation

AI governance establishes principles, decision rights, controls and evidence. It should be proportionate to risk rather than identical for every application. A low-risk recommendation tool and a system influencing eligibility for public benefits require different assessment intensity.

Core governance questions include:

• What is the intended and prohibited use?
• Who is accountable for the outcome?
• What data and model documentation is required?
• How will users and affected people be informed?
• Can decisions be explained or contested where necessary?
• What security threats and failure modes exist?
• How will bias and subgroup performance be assessed?
• What monitoring, audit and incident processes apply?
• When must the system be paused or retired?

Important principles include lawfulness, fairness, transparency, accountability, privacy, security, safety, accessibility and sustainability. These principles can conflict. Greater transparency may reveal security-sensitive information; a highly accurate model may be less interpretable; extensive data collection may improve prediction while increasing privacy risk. Governance therefore requires reasoned trade-offs and documented justification.

Organisations must comply with applicable data-protection, sectoral, employment, consumer-protection, accessibility, procurement and AI-specific requirements. Legal compliance is a minimum condition, not a complete ethical evaluation. Cross-border data transfers, third-party model terms and differing regulatory regimes create additional complexity.

10. Measuring ROI and Portfolio Performance

A credible evaluation begins with a baseline. If an organisation claims that AI reduced processing time, it should specify the previous process, measurement method, population, period and relevant confounding factors. Controlled pilots, phased rollouts, matched comparisons or interrupted time-series designs may improve causal interpretation, although operational constraints may prevent experimental designs.

A balanced scorecard can combine:

• Financial: cost per transaction, revenue, avoided loss, NPV.
• Operational: cycle time, throughput, reliability, forecast error and service-level performance.
• Customer or citizen: satisfaction, accessibility, complaint rates and equitable outcomes.
• Workforce: adoption, workload, skill development and safety.
• Risk: incidents, privacy events, security findings, subgroup disparities and unresolved exceptions.
• Portfolio: reuse of platforms and components, delivery time, retirement rate and benefits realised.

Metrics should not become targets detached from purpose. Optimising average accuracy can conceal poor performance for a minority group. Maximising automation can encourage unsafe removal of human review. A portfolio review should ask whether the system is producing the intended outcome, at acceptable risk, for the affected population.

11. Illustrative Case: Regional Public-Service Agency

Background

This illustrative case concerns a regional public-service agency responsible for processing housing-support applications. Processing delays were causing hardship, while staff spent substantial time checking documents and responding to repetitive enquiries. The agency proposed an AI system to classify submitted documents, identify missing information and provide a draft response for staff review.

Problem analysis

The initial proposal focused on building a sophisticated language model. Process analysis revealed that the main causes of delay were inconsistent document naming, repeated requests for information and a lack of clear escalation rules. The agency therefore defined the primary objective as reducing avoidable processing delay without reducing fairness or access.

Baseline measurement showed that applications were processed through several manual queues. Staff performed document classification, completeness checks and eligibility assessment. The first two tasks were suitable for assistance; eligibility decisions involved legal interpretation and remained subject to authorised staff.

Strategic assessment

The use case scored highly on service value and feasibility because the agency had historical documents and clear workflow owners. Risks were significant: documents contained sensitive personal information, errors could delay support, and applicants with limited digital access might be disadvantaged. The initiative was approved only as an assistive system, not an autonomous decision-maker.

Approach

The agency chose a partner for initial document-processing components but retained ownership of workflow design, data policy and acceptance criteria. A federated operating model was established: a central digital team supplied security and platform services, while housing-service specialists defined labels, exceptions and user requirements.

The redesigned process included:

1. Secure document upload and malware screening.

2. Automated classification with confidence indicators.

3. Human review for low-confidence or sensitive documents.

4. A completeness suggestion presented to staff, not directly sent to applicants.

5. Plain-language draft communication reviewed by an authorised employee.

6. An appeal and correction route for applicants.

Evaluation compared processing time, error rates, escalation rates and outcomes across relevant applicant groups. The agency monitored whether applicants using assisted or non-digital channels experienced different outcomes. Staff received training on automation bias and were instructed that confidence scores were not substitutes for judgement.

Outcome

The pilot reduced routine document-handling effort and improved queue visibility, but it did not eliminate processing delays because legal review remained the principal bottleneck for complex cases. This finding prevented an inaccurate claim that the AI system had transformed the entire service. The agency expanded the system for administrative assistance while investing separately in policy clarification and workforce capacity.

Learning points

First, process diagnosis is more valuable than technology enthusiasm. Second, assistive deployment may be strategically preferable to autonomous decision-making in high-impact services. Third, benefits must be measured at service level rather than inferred from model accuracy. Fourth, procurement contracts and operating ownership are as important as model selection. Finally, public value depends on accessibility, contestability and trust as well as efficiency.

12. Challenges and Contemporary Developments

Enterprise AI programmes face persistent challenges. Data may be fragmented across legacy systems, and historical data may encode earlier discrimination. Benefits may be uncertain while implementation costs are immediate. Skilled staff are scarce, and vendor markets change rapidly. Cyberattacks may target training data, model interfaces or connected workflows. Organisational incentives may reward launching pilots rather than sustaining reliable products.

Contemporary developments include foundation models, multimodal systems, AI agents, synthetic data, edge AI and specialised accelerators. These technologies can broaden capability but also increase governance complexity. Agentic systems that plan and execute multiple steps require explicit permissions, bounded tools, transaction controls, logging and recovery procedures. Synthetic data may support testing or privacy protection, but it can reproduce biases or fail to represent rare cases.

Sustainable strategy requires architectural modularity, documented dependencies, model and data inventories, clear retirement criteria and continuous reassessment of whether AI remains the best solution. Sometimes a rule, redesigned form, better training or improved data collection is more effective and less risky than machine learning.

Best-Practice Principles

1. Start with mission and user outcomes, not model novelty.

2. Establish a measurable baseline before deployment.

3. Treat AI initiatives as a governed portfolio.

4. Prioritise actionability, not prediction alone.

5. Assess build, buy and partner choices using total lifecycle cost and strategic dependence.

6. Assign accountable owners for data, products, risks and outcomes.

7. Redesign processes and roles before automating them.

8. Use proportionate, documented risk controls.

9. Evaluate performance across relevant groups and operating conditions.

10. Monitor deployed systems and provide mechanisms for correction, appeal and retirement.

11. Measure realised benefits rather than celebrating pilot activity.

12. Preserve organisational learning and avoid irreversible vendor dependence.

Chapter Summary

Enterprise AI strategy aligns AI investments with organisational mission, business capability and public or customer outcomes. Effective strategy moves beyond isolated pilots to portfolio governance, prioritised use cases and measurable value. Use-case selection should assess strategic fit, feasibility, time-to-value, reusability and risk, while recognising that public-sector decisions must include fairness, accessibility and legitimacy.

Build, buy and partner choices should account for total cost, differentiation, control, security, skills, contractual rights and exit options. Operating models may be centralised, federated or hub-and-spoke, but all require clear accountability and interdisciplinary talent. AI value depends on data governance, architecture, process redesign, human oversight and change leadership. ROI and NPV can support investment analysis, but financial metrics must be complemented by operational, social, ethical and risk measures.

Responsible governance covers privacy, fairness, transparency, security, safety, accessibility, contestability and regulatory compliance. The illustrative public-service case demonstrates that process diagnosis, assistive deployment and outcome measurement are more important than model sophistication. Sustainable transformation treats AI as an evolving organisational capability requiring monitoring, learning, adaptation and, when necessary, retirement.

Key Terms

AI strategy: A coordinated set of choices for using and governing AI to achieve organisational objectives.

Digital transformation: Redesign of organisational capabilities and value creation through digital technologies.

Use case: A defined application of AI to a task, decision or process for a specified outcome.

AI portfolio: The collection of AI initiatives managed as an interdependent set of investments.

Strategic fit: The degree to which an initiative supports organisational mission and priorities.

Realised benefit: The value actually captured after deployment and adoption constraints.

Net Present Value: The present value of expected cash flows less the initial investment, using a specified discount rate.

Build-buy-partner decision: The choice among internal development, external acquisition and collaborative delivery.

Operating model: The arrangement of people, processes, technology and decision rights.

Federated model: An arrangement combining central standards or platforms with distributed domain teams.

Data provenance: The documented origin, transformation and use history of data.

Model drift: Decline in model relevance or performance caused by changes in data or the environment.

Human-in-the-loop: A design in which human review or approval is required before a specified action.

Automation bias: The tendency to over-rely on system outputs despite contrary evidence.

Responsible AI: The design and use of AI consistent with legal, ethical, social and safety obligations.

Contestability: The ability of affected people to question, challenge or appeal an AI-assisted outcome.

Portfolio governance: Oversight for selecting, funding, monitoring and retiring AI initiatives.

Total cost of ownership: The full lifecycle cost of acquiring, integrating, operating, governing and retiring a system.

Process redesign: Deliberate restructuring of work, roles and decision flows before or alongside technology adoption.

Public value: Benefits to society, including service quality, equity, legitimacy and collective welfare.

Review Questions

Short Answer Questions

1. Distinguish an AI project from an enterprise AI strategy.

2. What is the difference between gross benefit and realised benefit?

3. State two advantages and two risks of a federated AI operating model.

4. Why is a baseline necessary when evaluating an AI initiative?

5. Define automation bias and explain why it matters.

Long Answer Questions

1. Explain how an organisation can prioritise AI use cases while balancing value, feasibility and risk.

2. Compare build, buy and partner strategies using strategic control, cost, capability and vendor-dependence criteria.

3. Discuss the relationship among process redesign, human oversight and successful AI adoption.

4. Explain how an AI governance framework should address privacy, fairness, security, accountability and contestability.

5. Critically evaluate the limitations of ROI as the primary measure of public-sector AI value.

Analytical / Application Questions

1. A university wants to deploy AI to predict students at risk of dropping out. Design a prioritisation assessment covering value, data, process actionability, fairness and governance.

2. An insurance firm can purchase a claims model or develop one internally. Recommend a decision framework and identify the evidence required before selecting an option.

3. Using the illustrative public-service case, propose three additional metrics that would test whether the system improved outcomes without creating unequal access or unsafe automation.

Multiple Choice Questions with Answers

Question 1

Which sequence best represents a sound enterprise AI strategy logic?

A. Model selection → data purchase → automation → strategic review

B. Organisational objective → capability or process → AI opportunity → measurable outcome → governance

C. Vendor selection → pilot launch → staff replacement → compliance exemption

D. Data volume → algorithm complexity → infrastructure expansion → business justification

Correct Answer: B. Organisational objective → capability or process → AI opportunity → measurable outcome → governance

Question 2

Which statement best describes realised benefit?

A. Theoretical improvement if a model performs perfectly

B. Total expenditure before deployment

C. Value actually captured after adoption and process constraints

D. Model accuracy measured on a test dataset

Correct Answer: C. Value actually captured after adoption and process constraints

Question 3

What is a principal advantage of a federated AI operating model?

A. It eliminates the need for central standards

B. It combines domain responsiveness with shared enterprise capabilities

C. It guarantees that all units use identical models

D. It transfers all accountability to external suppliers

Correct Answer: B. It combines domain responsiveness with shared enterprise capabilities

Question 4

Which factor is most important when evaluating a high-impact public-sector AI system?

A. Whether the model is the newest available architecture

B. Whether it maximises average accuracy regardless of subgroup performance

C. Whether it supports lawful, fair, accountable and contestable decisions

D. Whether it removes all human involvement

Correct Answer: C. Whether it supports lawful, fair, accountable and contestable decisions

Question 5

In the NPV equation, what does the discount rate represent?

A. The number of model parameters

B. The rate used to convert future cash flows into present values

C. The percentage of inaccurate predictions

D. The annual rate of data collection

Correct Answer: B. The rate used to convert future cash flows into present values

Question 6

Which is an example of process redesign rather than simple automation?

A. Copying an existing manual form into a digital document

B. Installing a larger server for the same workflow

C. Changing roles, escalation rules and decision rights around an AI-assisted process

D. Increasing the number of historical records in a database

Correct Answer: C. Changing roles, escalation rules and decision rights around an AI-assisted process

Question 7

What is automation bias?

A. A model’s inability to process automated data

B. The tendency to over-rely on system outputs despite contrary evidence

C. A legal requirement for every process to be automated

D. A method for removing bias from training data

Correct Answer: B. The tendency to over-rely on system outputs despite contrary evidence

Question 8

Which contractual provision is especially important when partnering for an enterprise AI system?

A. A ban on documenting model updates

B. Unlimited vendor discretion over all organisational data

C. Audit rights, incident responsibilities and data or model portability

D. A requirement to prevent system retirement

Correct Answer: C. Audit rights, incident responsibilities and data or model portability

Question 9

Why should AI portfolio governance include retirement criteria?

A. All AI systems become more accurate automatically over time

B. Some systems become unsafe, ineffective, obsolete or unjustified by their costs

C. Retirement criteria increase model complexity

D. They eliminate the need for monitoring

Correct Answer: B. Some systems become unsafe, ineffective, obsolete or unjustified by their costs

Question 10

Which evaluation approach best supports a credible claim that an AI system improved service outcomes?

A. Reporting only the model’s training accuracy

B. Comparing outcomes with a defined baseline and monitoring relevant groups and risks

C. Counting the number of employees attending AI training

D. Measuring the number of algorithms purchased

Correct Answer: B. Comparing outcomes with a defined baseline and monitoring relevant groups and risks