Data, AI and Emerging Technology Management

Introduction

Data and emerging technologies have become central to entrepreneurial competition. Firms now create value not only through physical products and services, but also through data, algorithms, digital platforms, automated processes and intelligent customer experiences. Artificial intelligence (AI), machine learning (ML), generative AI, robotics, blockchain, the Internet of Things (IoT), cloud computing and extended reality can reduce costs, improve decisions and create new business models. They can also introduce substantial risks: biased decisions, privacy violations, cybersecurity incidents, unreliable outputs, vendor dependence and regulatory exposure.

Technology management is therefore more than acquiring the newest tool. It is the disciplined process of identifying valuable opportunities, assessing organizational readiness, selecting or developing technologies, governing their use and measuring their effects. Entrepreneurs must connect technical possibilities with customer needs, operational capabilities, financial constraints and social responsibilities.

Learning Objectives

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

1. Distinguish data, AI, machine learning, generative AI, robotics, blockchain and other emerging technologies.
2. Assess an organization’s data, technical, human and governance readiness.
3. Evaluate technology use cases according to value, feasibility and risk.
4. Make reasoned build-versus-buy and human–AI collaboration decisions.
5. Explain model limitations, performance measures and responsible deployment practices.
6. Develop an implementation and governance approach for an entrepreneurial venture.

1. The Emerging Technology Landscape

Data as a strategic resource

Data are recorded observations, transactions, measurements, interactions or representations. Information is data interpreted in context, while knowledge is the ability to use information to guide action. Data create business value when they are relevant, accessible, sufficiently accurate and connected to a decision or customer outcome.

Common forms include:

• Structured data: records organized in tables, such as sales transactions.
• Unstructured data: text, images, audio and video.
• Semi-structured data: records with flexible tags or fields, such as JSON documents.
• First-party data: collected directly by an organization.
• Third-party data: obtained from an external provider.
• Real-time data: available with minimal delay.
• Synthetic data: artificially generated data used for testing, simulation or training.

Data quality is multidimensional. Accuracy concerns whether values are correct; completeness concerns missing values; consistency concerns agreement across systems; timeliness concerns currency; validity concerns conformity to rules; and uniqueness concerns the absence of unwanted duplicates. A large dataset with poor quality can produce worse decisions than a smaller, carefully governed dataset.

Artificial intelligence and machine learning

AI refers broadly to computational systems that perform tasks associated with human intelligence, including perception, prediction, language processing, planning and decision support. Machine learning is a class of AI in which a system learns patterns from data rather than being programmed only through explicit rules.

In supervised learning, a model learns from examples containing inputs and known outputs. Classification predicts categories, such as fraud or no fraud. Regression predicts numerical values, such as expected demand. In unsupervised learning, the desired output is not supplied; clustering and anomaly detection are common examples. In reinforcement learning, an agent learns through actions, feedback and rewards in an environment.

A model is a representation that maps inputs to outputs. During training, parameters are estimated from data. During inference, the trained model produces predictions for new cases. The distinction matters because strong training performance does not guarantee reliable performance in actual operations.

Generative AI

Generative AI produces new text, images, audio, video, software code or other content. Large language models generate text by estimating likely sequences of tokens from a prompt and their previous context. They are useful for drafting, summarizing, coding assistance, ideation, tutoring and conversational interfaces.

Generative systems do not inherently verify truth. They may produce hallucinations: fluent but false or unsupported statements. They may also reproduce biases in training data, expose confidential information through poor workflows, or generate content that creates copyright, reputation or security risks. Human review, grounded sources, access controls and clear usage policies are therefore essential.

Robotics and intelligent automation

Robotics combines mechanical systems, sensors, control systems and software to perform physical tasks. Robots may operate in manufacturing, logistics, healthcare, agriculture and service environments. Robotic process automation (RPA) automates structured, rule-based digital tasks, while intelligent automation adds capabilities such as document understanding or prediction.

Automation is most effective when processes are stable, measurable and sufficiently standardized. Automating a broken process may simply make errors occur faster. Entrepreneurs should first understand process variation, exception handling, worker safety and the consequences of failure.

Blockchain and distributed ledgers

A blockchain is a type of distributed ledger in which records are grouped into blocks and linked through cryptographic techniques. A consensus mechanism allows participants to agree on the ledger’s state without relying entirely on one central administrator. Smart contracts are programs that execute specified actions when conditions are met.

Potential applications include traceability, digital ownership, settlement and shared records among organizations that do not fully trust one another. Blockchain is not automatically appropriate. A conventional database may be cheaper, faster and easier to correct when one trusted organization controls the system. Important questions include governance, identity, privacy, interoperability, transaction capacity and legal enforceability.

Other enabling technologies

Cloud computing provides on-demand access to computing, storage and software resources. IoT connects physical objects to networks through sensors and communication systems. Edge computing processes data near the point of collection, reducing latency and bandwidth requirements. Digital twins represent physical assets or processes for monitoring and simulation. Extended reality combines virtual, augmented and mixed reality experiences.

These technologies often operate together. For example, an agricultural venture may use IoT sensors, edge processing, cloud analytics, ML-based crop prediction and autonomous equipment. Technology management must therefore consider architecture and integration rather than isolated tools.

2. From Technology Possibility to Business Use Case

A use case describes who uses a technology, what problem it addresses, how value is created and what risks arise. A strong use case begins with a consequential business problem rather than a fascination with a tool.

A practical screening framework considers five dimensions:

A simple financial measure is return on investment:

Here, B is the relevant benefit over a defined period and C is the relevant cost over the same period. Both must use the same currency and time period. This measure is useful for initial comparison but does not capture uncertainty, strategic option value, distributional effects or long-term capability building. Benefits should include only attributable benefits, and costs should include data preparation, integration, training, monitoring, maintenance and decommissioning—not merely the purchase price.

For risk-adjusted decisions, expected monetary value may be used:

where p_i is the probability of outcome i, V_i is its monetary value, and the probabilities sum to 1. This calculation assumes that relevant outcomes can be identified and assigned credible probabilities. It should support, not replace, qualitative assessment of catastrophic or irreversible risks.

Use-case discovery process

1. Identify a customer or operational pain point.
2. Describe the current process and its measurable weaknesses.
3. Determine whether data or automation can address the cause rather than merely the symptom.
4. Define the decision, prediction or action to be improved.
5. Establish success measures and unacceptable outcomes.
6. Test feasibility with a small, reversible pilot.
7. Decide whether to scale, redesign or stop.

Useful performance measures may include conversion rate, cycle time, error rate, customer retention, cost per transaction, service availability and employee workload. AI-specific measures should be linked to business consequences; accuracy alone is not a business outcome.

3. Data Readiness and Technology Readiness

Data readiness is the extent to which data are available, relevant, legally usable, sufficiently high quality and technically accessible for a proposed use. A readiness assessment should examine:

• Availability: Are enough observations available for the intended task?
• Representativeness: Do the data reflect the populations and conditions in deployment?
• Label quality: If supervised learning is used, are outcomes defined and labeled consistently?
• Provenance: Can the organization identify where data came from and how they changed?
• Access and security: Are permissions, encryption and retention controls in place?
• Interoperability: Can systems exchange data using stable definitions and interfaces?
• Governance: Are ownership, stewardship, privacy and accountability assigned?

A model trained on historical decisions may learn historical discrimination. A dataset can be technically complete yet strategically unsuitable if it omits important variables or represents only existing customers. Data readiness is therefore a business and ethical question, not merely an engineering checklist.

Technology readiness includes infrastructure, integration capability, cybersecurity, talent, support arrangements and the ability to operate a system after launch. Entrepreneurial firms often underestimate operational costs: monitoring, model updates, vendor management, incident response and user support continue after implementation.

4. Selecting, Building and Buying Technology

The build-versus-buy decision compares internal development with acquisition from a vendor, often including a hybrid option. The decision should not be reduced to purchase price.

A hybrid approach might buy a foundation model or cloud platform while building proprietary data pipelines, prompts, evaluation procedures and user workflows. Contract review should address data ownership, confidentiality, service levels, audit rights, portability, intellectual property, subcontractors, security incidents and termination assistance.

Total cost of ownership includes acquisition, implementation, integration, training, governance, support, usage charges, upgrades and exit costs. A vendor that appears inexpensive per transaction may be costly at scale or difficult to replace.

5. Human–AI Collaboration and Organizational Design

Human–AI collaboration is not simply a choice between full automation and no automation. Common patterns include:

1. Human-only: technology provides information, but a person makes and executes the decision.
2. Human-in-the-loop: AI recommends or drafts; a person approves consequential outputs.
3. Human-on-the-loop: AI acts within defined boundaries while people monitor performance and intervene when necessary.
4. Automated execution: the system acts without routine human approval, appropriate only where risks are low, decisions are reversible and controls are strong.

The appropriate pattern depends on task complexity, error costs, explainability, reversibility and the availability of skilled reviewers. Human oversight is ineffective if employees lack time, authority or expertise to challenge the system. Automation bias may cause people to accept algorithmic recommendations too readily, while algorithm aversion may lead them to reject useful assistance after a visible error.

Good design allocates responsibility explicitly. It specifies who owns the decision, who monitors the system, who handles exceptions and who communicates with affected stakeholders. Employees should receive training not only in tool operation but also in limitations, escalation and responsible use.

6. Understanding and Evaluating Model Performance

For a binary classifier, the confusion matrix records true positives, true negatives, false positives and false negatives. Common measures include:

Here, TP is a true positive, TN a true negative, FP a false positive and FN a false negative. Accuracy is informative when classes and error costs are reasonably balanced. Precision asks how many positive predictions were correct; recall asks how many actual positive cases were detected. The measures are unitless proportions and apply to a specified dataset and decision threshold.

A fraud detector, for example, may prioritize recall to detect more fraud, but excessive false positives can inconvenience legitimate customers. Threshold selection should therefore reflect economic and social consequences.

For regression, mean absolute error is:

where n is the number of observations, y_i is the actual value and ŷ_i is the prediction. MAE has the same units as the predicted variable and treats all absolute errors proportionally. Evaluation should use data that represent future deployment and should be separated from training data to avoid overly optimistic results.

Important limitations include:

• Overfitting: the model captures noise in training data and performs poorly on new data.
• Data drift: the relationship between inputs and outcomes changes over time.
• Concept drift: the meaning or distribution of the target changes.
• Distribution shift: deployment conditions differ from training conditions.
• Spurious correlation: a statistical association is mistaken for a causal relationship.
• Limited explainability: complex systems may be difficult to interpret.
• Calibration problems: predicted probabilities do not match observed frequencies.

No performance measure proves that a model is fair, safe or causally valid. Evaluation must include subgroup analysis, robustness tests, security testing and monitoring after deployment.

7. Responsible and Governed Deployment

Responsible technology management combines legal compliance, ethical reasoning and operational control. Core principles include:

• Purpose limitation: collect and use data for specified legitimate purposes.
• Data minimization: use no more personal data than necessary.
• Transparency: communicate meaningful information about system use and limitations.
• Human accountability: assign responsibility to identifiable people and institutions.
• Fairness: investigate disparate impacts and avoid unjustified discrimination.
• Privacy and security: protect confidentiality, integrity and availability.
• Safety and reliability: test foreseeable failure modes and provide fail-safe procedures.
• Contestability: allow affected people to request explanation, correction or review where appropriate.

Organizations should maintain an inventory of AI systems, classify applications by risk, document data sources and intended use, conduct impact assessments, retain evaluation records and establish incident reporting. Access controls should follow least privilege. Sensitive information should not be entered into public tools without authorization. Model and prompt changes should be version-controlled and tested.

Regulatory obligations differ by jurisdiction and sector. Areas commonly affected include privacy and data protection, consumer protection, employment, intellectual property, financial services, product safety and automated decision-making. Compliance is not static; ventures should monitor applicable legislation, contractual obligations and industry standards.

8. Implementation Roadmap

A practical implementation roadmap has seven stages:

1. Strategic alignment: define the problem, affected stakeholders and desired outcome.
2. Readiness assessment: examine data, infrastructure, skills, governance and process maturity.
3. Risk classification: identify privacy, safety, fairness, security, legal and reputational risks.
4. Pilot design: select a limited scope, baseline measures, approval criteria and rollback plan.
5. Development or procurement: choose architecture, vendor or internal team and document requirements.
6. Validation and launch: test technical performance, user experience, security, subgroup outcomes and operational resilience.
7. Monitoring and retirement: track outcomes, drift, incidents, costs and user feedback; update or withdraw the system when appropriate.

A baseline is essential. If a new forecasting system reduces errors from 20% to 15%, the absolute reduction is 5 percentage points, while the relative reduction is 25%. Both descriptions are correct but communicate different things. Measurement should also consider unintended effects, such as increased employee workload or exclusion of customers who cannot use a digital channel.

Illustrative Case: FarmLink’s AI Advisory Service

Background

FarmLink is a fictional early-stage venture that provides crop advice to small farms. It proposes combining weather data, soil sensors, satellite imagery and generative AI to provide irrigation and planting recommendations.

Problem

The founders initially assume that a sophisticated AI model will create a defensible product. A readiness review reveals inconsistent sensor calibration, limited historical data for certain crops, weak rural connectivity and no formal process for correcting incorrect advice. Some farmers speak local languages not represented in the proposed training data.

Analysis

The team maps the decision rather than beginning with the technology. The immediate opportunity is not fully autonomous farming; it is earlier identification of water stress and clearer communication of recommended actions. The team identifies risks: inaccurate predictions could damage crops, automated advice could be mistaken for professional agronomic guidance, and location data could reveal sensitive commercial information.

Approach and solution

FarmLink chooses a staged, human-supported approach. It purchases weather and satellite services but builds its own data-quality pipeline and farmer feedback system. A conventional statistical baseline is established before testing ML models. Recommendations are presented with evidence, confidence information and an invitation to consult a human agronomist. The system begins with low-risk alerts rather than automatic irrigation control.

Sensors are tested and flagged when readings are implausible. Data-use notices explain collection and retention. Pilot farmers represent different regions, farm sizes and languages. Performance is evaluated by crop and location, not only in aggregate. A rollback process allows the venture to suspend recommendations after severe anomalies.

Outcome

The pilot does not produce the originally imagined autonomous platform, but it achieves a more credible service. Farmers receive earlier alerts, agronomists spend less time on routine monitoring, and the company learns which data are valuable. The staged design reduces safety and trust risks while generating evidence for future investment.

Learning points

1. A narrow, high-value use case is often better than an ambitious technology showcase.
2. Data quality and representativeness constrain model value.
3. Human expertise remains important where errors have material consequences.
4. Buying infrastructure while building differentiated workflows can be rational.
5. Responsible design can improve commercial adoption rather than merely add cost.

Challenges and Contemporary Developments

Emerging technologies develop faster than many organizations’ governance processes. Foundation models are increasingly multimodal and can be integrated into enterprise workflows, but their scale does not eliminate hallucination, bias or security concerns. Smaller specialized models may offer lower cost, greater privacy and better domain control. Edge AI is expanding where latency, connectivity or data sovereignty matters. Synthetic data and simulation can support rare-event testing, although synthetic data may reproduce the assumptions and biases of the process used to generate it.

Other developments include autonomous agents that plan and execute sequences of actions, privacy-enhancing computation, federated learning and advances in quantum computing research. These technologies should be assessed by demonstrated capability and business fit, not publicity. Pilot evidence, independent evaluation and reversibility are especially important when technology claims are uncertain.

Entrepreneurs should also recognize concentration risks. Dependence on a small number of cloud, model or data providers can create pricing, availability and geopolitical vulnerabilities. Interoperability, data portability, contractual safeguards and alternative suppliers improve resilience.

Best-Practice Checklist

Before deploying an emerging technology, an organization should be able to answer:

• What specific problem and measurable outcome justify the system?
• What is the baseline alternative?
• Are the data lawful, relevant, representative and sufficiently reliable?
• Who is accountable for decisions and incidents?
• What errors are most harmful, and how will they be detected?
• Where is human review required?
• How will users understand, challenge or correct outputs?
• What are the total costs and vendor lock-in risks?
• How will performance, drift, security and fairness be monitored?
• Under what conditions will the system be paused, redesigned or retired?

Chapter Summary

Data and emerging technologies create entrepreneurial opportunities when they are connected to meaningful customer and operational problems. AI includes a broad family of capabilities, while machine learning learns patterns from data and generative AI produces new content. Robotics, blockchain, IoT, cloud and edge computing extend these capabilities into physical and distributed environments.

Effective technology management requires use-case discipline, data and organizational readiness assessment, careful build-versus-buy analysis and explicit human–AI responsibility. Model performance must be evaluated with measures suited to the decision, while recognizing overfitting, drift, bias, weak representativeness and the difference between correlation and causation. Responsible deployment requires privacy, security, fairness, transparency, accountability, contestability and ongoing monitoring. A staged pilot with clear baselines, rollback procedures and stakeholder feedback is usually more valuable than an uncontrolled technology launch.

Key Terms

Artificial intelligence: Computational systems performing tasks associated with human intelligence.

Machine learning: AI in which patterns and parameters are learned from data.

Generative AI: Systems that create new text, images, code, audio or other content.

Data readiness: Suitability of data for a proposed use in terms of quality, access, legality and governance.

Structured data: Data organized in predefined fields or tables.

Data drift: Change over time in input data or their distributions.

Concept drift: Change in the relationship between inputs and outcomes.

Supervised learning: Learning from examples with known target outcomes.

Unsupervised learning: Learning patterns without supplied target labels.

Reinforcement learning: Learning through actions, rewards and environmental feedback.

Blockchain: A distributed ledger using linked records and a consensus process.

Smart contract: Software that executes specified actions when conditions are met.

Human-in-the-loop: A design in which a person reviews or approves AI outputs.

Overfitting: Learning noise or peculiarities that reduce performance on new data.

Precision: The proportion of positive predictions that are true positives.

Recall: The proportion of actual positive cases detected by a model.

Model hallucination: A plausible but false or unsupported generative-AI output.

Build-versus-buy: The decision to develop a capability internally, acquire it externally or combine both.

Total cost of ownership: The full cost of acquiring, operating, maintaining and exiting a technology.

Review Questions

Short Answer Questions

1. Distinguish data, information and knowledge.
2. What is the difference between supervised and unsupervised learning?
3. Identify three dimensions of data quality.
4. Why may a conventional database be preferable to blockchain?
5. What is meant by human-in-the-loop deployment?

Long Answer Questions

1. Explain how an entrepreneur should move from a business problem to an AI use case.
2. Discuss the main criteria in a build-versus-buy decision.
3. Explain why model accuracy alone is insufficient for responsible deployment.
4. Describe a data and technology readiness assessment for a start-up.
5. Discuss the governance requirements of generative AI in an organization.

Analytical / Application Questions

1. A retailer’s fraud model has high accuracy but misses most fraudulent transactions. Which performance measure should be examined, and what additional business trade-off should be considered?
2. Design a staged human–AI workflow for an online lending venture. Identify where human review, monitoring and customer appeal should occur.
3. Evaluate whether an entrepreneurial food-traceability venture should use blockchain. State the conditions under which it would add value and when a conventional database would be preferable.

Multiple Choice Questions with Answers

Question 1

Which statement best describes machine learning?

A. A system that only follows manually written rules

B. A method in which systems learn patterns from data

C. A physical machine that performs manufacturing tasks

D. A distributed ledger used to record transactions

Correct Answer: B. A method in which systems learn patterns from data

Question 2

Which data-quality dimension concerns whether data are current enough for the intended decision?

A. Timeliness

B. Uniqueness

C. Validity

D. Completeness

Correct Answer: A. Timeliness

Question 3

In a human-in-the-loop system, what normally happens?

A. The system is prohibited from making recommendations

B. A person reviews or approves consequential outputs

C. The system operates without monitoring

D. Only historical data are used

Correct Answer: B. A person reviews or approves consequential outputs

Question 4

What is a major limitation of generative AI?

A. It cannot produce new content

B. It always requires blockchain infrastructure

C. It may produce fluent but false information

D. It can only process numerical tables

Correct Answer: C. It may produce fluent but false information

Question 5

Which situation most strongly supports buying rather than building a technology?

A. The capability is the venture’s unique competitive advantage

B. The firm has highly specialized internal talent

C. A mature standard product meets requirements and speed is important

D. The workflow requires proprietary algorithms unavailable externally

Correct Answer: C. A mature standard product meets requirements and speed is important

Question 6

If a classifier has TP = 80 and FN = 20, what is its recall?

A. 20%

B. 50%

C. 80%

D. 100%

Correct Answer: C. 80%

Question 7

Which problem occurs when a model performs well on training data but poorly on new data?

A. Overfitting

B. Encryption

C. Interoperability

D. Consensus

Correct Answer: A. Overfitting

Question 8

Which consideration is most relevant when deciding whether blockchain is appropriate?

A. Whether a shared ledger is needed among parties with limited mutual trust

B. Whether the venture wants the newest technology regardless of need

C. Whether all records should be impossible to correct

D. Whether a centralized database is always prohibited

Correct Answer: A. Whether a shared ledger is needed among parties with limited mutual trust

Question 9

What should normally be established before assessing whether a new AI system improves performance?

A. A baseline using the current process or alternative

B. A larger marketing budget

C. A decision to automate every task

D. A promise that errors will never occur

Correct Answer: A. A baseline using the current process or alternative

Question 10

Which practice best supports responsible AI deployment?

A. Avoiding performance monitoring after launch

B. Allowing unrestricted access to sensitive data

C. Documenting purpose, accountability, evaluation and incident procedures

D. Measuring only aggregate accuracy

Correct Answer: C. Documenting purpose, accountability, evaluation and incident procedures