Artificial Intelligence in Business

Introduction and Context

Artificial intelligence (AI) refers to computer systems designed to perform tasks that normally require human intelligence, such as recognising patterns, understanding language, making predictions, generating content and selecting actions. In business, AI is not simply a software feature. It is an organisational capability that combines data, algorithms, computing infrastructure, human expertise, business processes and governance.

AI has become contemporary business infrastructure because organisations generate large volumes of structured data—such as sales transactions and inventory records—and unstructured data—such as emails, contracts, images, audio and customer reviews. AI can convert these data into predictions, recommendations, classifications or generated outputs. However, its value depends on whether it improves a measurable business outcome while remaining accurate, secure, fair, explainable enough for its purpose and subject to appropriate human control.

This chapter focuses on four connected areas: machine learning, generative AI, natural language processing and intelligent agents. It also evaluates business applications, prompting practices, limitations, human oversight and responsible adoption.

Learning Objectives

After studying this chapter, you should be able to:

1. Explain the principles and main types of machine learning.
2. Distinguish generative AI from predictive and analytical AI.
3. Describe how natural language processing supports business activities.
4. Explain intelligent agents and their relationship with business workflows.
5. Evaluate AI use cases using value, feasibility and risk criteria.
6. Construct effective prompts and assess generated outputs.
7. Identify AI limitations and design appropriate human oversight.
8. Recommend practices for responsible AI adoption.

1. Foundations of Business AI

1.1 Data, models and decisions

A business AI system typically contains five elements:

• Data: observations used to train, test or operate a system.
• Features: measurable characteristics used by a model, such as customer tenure or order value.
• Model: a mathematical or computational representation of relationships in data.
• Inference: the model’s output when it receives new input.
• Action or decision: the business response to the output.

For example, a retailer may use customer purchase history, browsing behaviour and seasonality as inputs to predict demand. The prediction is not itself the business objective. The objective may be lower stock-outs, reduced waste or improved customer service. A model that is technically accurate but produces no useful operational change has limited business value.

A useful distinction is between automation and augmentation. Automation allows a system to perform a task with limited human intervention. Augmentation supports a person’s judgement, for example by highlighting unusual transactions for an analyst. Many high-value applications combine both: AI performs repetitive screening, while employees investigate exceptions and make accountable decisions.

1.2 Types of AI tasks

Business AI tasks can be classified as follows:

These categories may overlap. A customer-service system, for instance, may classify an enquiry, retrieve relevant policy documents, generate a response and escalate the matter if risk is high.

2. Machine Learning

2.1 Meaning and process

Machine learning (ML) is a branch of AI in which a system learns patterns or relationships from data rather than being programmed with every rule explicitly. Learning does not mean that the system understands in the human sense. It means that an algorithm adjusts parameters to perform a defined task according to an objective function.

A standard ML workflow is:

1. Define the business problem and success measure.
2. Collect and govern relevant data.
3. Clean, label and transform the data.
4. Divide data into training, validation and test sets.
5. Select and train a model.
6. Evaluate performance and business consequences.
7. Deploy, monitor and periodically review the model.

The separation of data is important. The training set supports learning, the validation set supports model selection and tuning, and the test set provides an approximately independent final assessment. Reusing test data during development can produce over-optimistic results.

2.2 Supervised, unsupervised and reinforcement learning

Supervised learning uses labelled examples. Each training observation has inputs and a known target, such as whether a loan defaulted. Classification predicts categories; regression predicts numerical values.

Unsupervised learning uses data without a specified target. Clustering can identify customer groups, while dimensionality-reduction methods can simplify complex data for analysis or visualisation. The resulting groups require managerial interpretation; an algorithmic cluster is not automatically a meaningful market segment.

Reinforcement learning involves an agent selecting actions in an environment and receiving rewards or penalties. Over time, it learns a policy—a strategy for selecting actions. It may be useful for dynamic decisions such as inventory control or resource allocation, but it requires careful definition of rewards. A poorly designed reward can encourage behaviour that improves a short-term metric while harming customers or long-term performance.

2.3 Model performance and business evaluation

For a binary classifier, a confusion matrix records true positives (TP), false positives (FP), true negatives (TN) and false negatives (FN). Important measures include:

Accuracy is the proportion of all predictions that are correct. Precision is the proportion of predicted positives that are actually positive. Recall is the proportion of actual positives that the model identifies. These measures assume a defined positive class and are most informative when the costs of errors are considered. For example, in fraud screening, missing fraud may be more costly than reviewing a legitimate transaction, whereas in marketing, excessive false alarms may unnecessarily annoy customers.

For a regression or forecasting model, the mean absolute error is:

Here, n is the number of observations, yᵢ is the actual value and ŷᵢ is the predicted value. MAE is expressed in the same units as the target, such as units sold or dollars. It assumes that observations are being compared on a meaningful scale. A lower MAE indicates smaller average absolute errors, but it does not show whether errors are systematically too high or too low.

Business evaluation should go beyond model metrics. It should consider implementation cost, time saved, revenue gained, service quality, error consequences, fairness, security and employee acceptance. A model with slightly lower statistical error may be inferior if it is substantially more expensive or less interpretable for a regulated decision.

2.4 Generalisation, overfitting and bias

A model generalises when it performs effectively on new data from the relevant operating environment. Overfitting occurs when a model learns accidental details or noise in training data and performs poorly on new cases. More complex models are not automatically better; complexity must be justified by improved out-of-sample performance and business value.

Data may also contain historical or measurement bias. If past lending decisions systematically disadvantaged a group, learning from those decisions can reproduce the pattern. Removing a protected attribute does not necessarily remove discrimination because other variables may act as proxies. Responsible evaluation therefore examines outcomes across relevant groups and investigates the causes of disparities rather than relying on a single overall accuracy score.

3. Generative AI

3.1 Concept and distinction from predictive AI

Generative AI creates new outputs—such as text, images, audio, code or synthetic data—based on patterns learned from training data. Predictive AI generally estimates a label, probability or numerical value, whereas generative AI produces content in a requested form. The distinction is functional rather than absolute: a language model predicts likely next tokens, but the resulting sequence appears as an original response.

Large language models (LLMs) are generative models trained on extensive text datasets. They process text as tokens and calculate probability distributions over possible continuations. Their fluent output can be useful, but fluency is not proof of factual accuracy, reasoning reliability or source authority.

Common business uses include drafting correspondence, summarising documents, creating marketing variants, assisting software development, answering questions over approved knowledge bases and supporting brainstorming. Generative AI can reduce the cost of producing a first draft, but organisations remain responsible for accuracy, confidentiality, intellectual-property compliance and the consequences of publication or action.

3.2 Prompting as task specification

A prompt is an instruction or input provided to a generative model. Effective prompting treats the model as a probabilistic tool whose output depends strongly on context and task specification. A useful prompt normally contains:

1. Role or perspective: for example, “Act as a procurement analyst.”
2. Task: the exact activity to perform.
3. Context: relevant facts, policies, audience and background.
4. Constraints: length, tone, exclusions, legal or formatting requirements.
5. Output format: table, numbered list, JSON-like fields or executive memo.
6. Quality criteria: evidence, assumptions, uncertainty and checking requirements.

For example: “Using only the policy text below, identify clauses affecting supplier onboarding. Return a table with clause, obligation, responsible role and unresolved ambiguity. Quote the relevant wording and do not invent missing requirements.” This prompt is stronger than “Summarise this policy” because it specifies scope, evidence and output structure.

Zero-shot prompting gives instructions without examples. Few-shot prompting provides examples of desired inputs and outputs. Examples can clarify classification categories or tone, but poor examples can reproduce errors. For complex tasks, it is useful to decompose the work into stages: extract facts, check completeness, then draft. The user should ask for assumptions and uncertainty rather than accepting confident language automatically.

Prompting does not eliminate the need for data governance. Sensitive personal, financial or commercially confidential information should not be entered into an AI service unless the organisation has approved the service, contractual protections, access controls and an appropriate retention policy.

3.3 Retrieval-augmented generation

A generative model may produce incorrect or outdated information because its learned parameters are not a current database of organisational policy. Retrieval-augmented generation (RAG) addresses this by retrieving relevant documents from an approved knowledge base and supplying them as context for generation. A typical flow is: user question → document retrieval → context selection → generated answer → citations or review.

RAG can improve traceability and currency, but it does not guarantee correctness. Retrieval may select an obsolete document, fail to find a relevant clause or include conflicting sources. Documents therefore require ownership, version control, access permissions and quality review.

4. Natural Language Processing

Natural language processing (NLP) enables computers to analyse and generate human language. It combines computational methods, linguistic concepts and machine learning. Business language is difficult because meaning depends on context, ambiguity, domain vocabulary, sarcasm, multilingual variation and incomplete information.

Important NLP tasks include:

• Tokenisation: dividing text into units used by a model.
• Text classification: assigning categories, such as complaint type.
• Sentiment or emotion analysis: estimating expressed attitudes; this is probabilistic and culturally sensitive.
• Named-entity recognition: identifying names, organisations, locations, dates or amounts.
• Information extraction: converting unstructured text into structured fields.
• Machine translation: converting text between languages.
• Question answering and summarisation: producing responses or shorter representations.
• Speech recognition and synthesis: converting speech to text and text to speech.

NLP supports customer-service routing, contract review, compliance monitoring, call-centre analytics, employee search and market research. Consider invoice processing: an NLP system can extract supplier name, invoice date and amount, then pass the fields to an accounting workflow. A human or rule-based control may verify exceptions before payment.

NLP outputs should be interpreted as estimates. A sentiment model may misread irony; an entity extractor may confuse a product name with a person; a translation system may alter a legally significant phrase. Evaluation should use representative business data, include multilingual and accessibility considerations where relevant, and measure errors by consequence rather than merely by average score.

5. Intelligent Agents

An intelligent agent is a system that perceives information about an environment, selects actions according to objectives and acts on that environment. In business, an agent may monitor a queue, retrieve information, call approved tools, update a system and escalate cases. An LLM may provide the language interface or planning component, but an agent also requires tools, permissions, memory or state, policies and monitoring.

A useful agent architecture contains:

• Perception: receives events, documents, messages or sensor data.
• Interpretation and planning: identifies the task and possible steps.
• Tools: accesses approved applications, databases or APIs.
• Policy and permissions: restricts what actions are allowed.
• Execution: performs an action or requests approval.
• Observation and logging: records results for monitoring and audit.

Agents differ from simple chatbots because they can pursue a multi-step objective and interact with external systems. This creates additional risk. An inaccurate answer in a chat may mislead a user; an inaccurate agent action may issue a refund, change a supplier record or send confidential information.

Agent autonomy should therefore be graduated. Low-risk agents may draft or recommend. Medium-risk agents may execute routine, reversible tasks within limits. High-impact actions—such as terminating an employee, approving a large payment or making a credit decision—should require qualified human review and, where appropriate, dual control.

6. Evaluating Business Use Cases

An organisation should not begin with the question “Where can we use AI?” It should begin with a business problem. A practical evaluation asks:

A prioritisation matrix can compare expected value with implementation difficulty and risk. Low-risk internal summarisation may be a suitable pilot. Automated eligibility decisions affecting essential services require stronger evidence, controls and governance.

The business case should include total cost of ownership: data preparation, model or service fees, integration, cybersecurity, staff training, monitoring, incident response and retirement. It should also account for opportunity cost and the possibility that employees spend time correcting poor outputs.

7. Human Oversight and Responsible AI

Responsible AI means designing and using AI in ways consistent with law, professional duties, organisational values and respect for affected people. Core principles include:

• Validity and reliability: the system performs its intended task under relevant conditions.
• Safety: foreseeable harms are identified and reduced.
• Fairness: outcomes and processes do not create unjustified discrimination.
• Transparency: users can understand the system’s role and relevant limitations.
• Privacy: data collection, use, retention and sharing are controlled.
• Security: systems resist unauthorised access, manipulation and data leakage.
• Accountability: named people and governance bodies remain responsible.

Human oversight is more than placing a person somewhere in the workflow. The reviewer must have sufficient competence, time, authority and access to evidence to challenge the system. A human who merely clicks “approve” under time pressure provides weak oversight.

Controls may include confidence thresholds, exception queues, sampling, second-level review, audit logs, version control, access restrictions, red-team testing and incident reporting. Organisations should document intended use, prohibited uses, data sources, performance by relevant groups, known limitations and escalation procedures.

Privacy laws, employment rules, consumer-protection requirements, intellectual-property law and sector-specific regulation may apply. Legal compliance is a minimum, not a complete ethical assessment. An action can be technically lawful yet unfair, deceptive or damaging to trust.

8. Limitations and Risks

AI systems have several important limitations:

1. Hallucination: a generative system may produce plausible but unsupported claims, citations or numbers.
2. Data drift: relationships change as markets, customer behaviour or policies change.
3. Distribution shift: deployment data may differ from training data.
4. Bias: data, labels, objectives or deployment decisions may disadvantage groups.
5. Opacity: complex models may be difficult to explain or audit.
6. Security threats: prompt injection, data poisoning, model extraction and unauthorised access can undermine systems.
7. Confidentiality and privacy risks: prompts or outputs may expose sensitive information.
8. Automation bias: users may trust machine outputs more than warranted.
9. Copyright and ownership uncertainty: generated content may draw on protected material or create unclear rights.
10. Environmental and financial cost: training and operating large systems consume computing resources and money.

Risk is context-dependent. A spelling suggestion and a medical triage recommendation should not be governed identically. A proportionate approach matches controls to impact, reversibility, affected rights and scale.

9. Implementation Roadmap

A responsible adoption programme can follow these stages:

1. Discover: identify problems, stakeholders, affected groups and existing processes.
2. Classify risk: determine whether the proposed use is low, moderate or high impact.
3. Prepare data: establish lawful access, quality checks, retention and ownership.
4. Design controls: define permissions, review points, fallback procedures and audit records.
5. Pilot: test with representative cases and compare against a meaningful baseline.
6. Evaluate: assess technical metrics, business outcomes, fairness, user experience and incidents.
7. Deploy gradually: limit scope, monitor performance and train users.
8. Review continuously: detect drift, revise prompts or models, investigate complaints and retire the system when it no longer provides justified value.

AI literacy should extend beyond technical staff. Managers need to understand appropriate use, employees need practical guidance, and affected customers or citizens may need clear explanations and routes for contesting significant decisions.

Applied Case: AI-Assisted Customer-Service Triage

Background

Illustrative case: Northstar Telecom receives approximately 40,000 customer messages each month through email and web forms. Messages concern billing, technical faults, cancellations and suspected fraud. Response times are inconsistent, and urgent service outages are sometimes mixed with routine enquiries.

Problem

Management proposes an AI system that classifies messages, extracts key details, drafts responses and routes cases to teams. The company must improve response time without allowing the system to make unsupported promises or mishandle vulnerable customers.

Analysis

The classification task can be evaluated using precision and recall for urgent cases. Missing an outage report has a greater cost than sending an extra case to a specialist, so recall is important. However, excessive false positives can overload the specialist team. Historical labels are reviewed because previous routing practices may contain inconsistent categories.

NLP extracts account identifiers, service type and stated problem. A retrieval system supplies current service policies to a generative model, which drafts a response. The draft must be clearly labelled as machine-assisted and cannot include compensation or cancellation commitments without authorised review.

Approach and solution

Northstar pilots the system on a limited set of categories. It prohibits the model from making final decisions about fraud, termination, refunds above a threshold or complaints involving vulnerable customers. Every draft includes links to the policy sources used. Cases with low confidence, conflicting information or sensitive indicators go directly to trained staff.

Managers monitor routing recall, precision, average response time, customer complaints, correction rates and outcomes across language groups. Employees can override the classification and record the reason. A governance committee reviews incidents monthly and checks whether policies or retrieval documents are current.

Outcome and learning points

The pilot improves routing speed and reduces repetitive drafting, but it also reveals that informal customer language is frequently misclassified. The organisation expands training examples, improves escalation rules and retains human review for high-impact matters. The central lesson is that value comes from redesigning the workflow—not from adding a model alone. Clear boundaries, current knowledge sources, measurable outcomes and accountable employees are essential.

Contemporary Developments

Multimodal systems increasingly process combinations of text, images, audio and video. Smaller specialised models may be cheaper and easier to control than general-purpose models. Organisations are also adopting model-risk management, AI inventories, synthetic data and automated evaluation tools. These developments do not remove fundamental responsibilities: data quality, security, fairness, human accountability and continuous monitoring remain necessary.

Chapter Summary

AI in business combines data, computational models, organisational processes and accountable decisions. Machine learning learns patterns for classification, prediction, clustering or action selection; its performance must be evaluated on new data and in business context. Generative AI creates text and other content, but fluent outputs can be inaccurate, biased or confidential. Prompting improves results by specifying task, context, constraints, format and quality criteria, while retrieval-augmented generation can connect outputs to approved documents without guaranteeing correctness.

Natural language processing supports classification, extraction, translation, summarisation, speech and customer-service activities, but language ambiguity requires representative evaluation. Intelligent agents add planning and tool use, making permissions, logging and graduated autonomy especially important. Responsible adoption evaluates value, data, feasibility, risk, human roles and measurable outcomes. Human oversight must be meaningful, and governance must address reliability, fairness, privacy, security, transparency and accountability. Successful adoption is an ongoing sociotechnical process rather than a one-time technology purchase.

Key Terms

Artificial intelligence: Computer-based capability to perform tasks associated with human intelligence.

Machine learning: AI method that learns patterns from data to perform a defined task.

Supervised learning: Learning from examples with known target labels or values.

Unsupervised learning: Learning patterns in data without predefined target labels.

Reinforcement learning: Learning actions through rewards and penalties in an environment.

Classification: Predicting a category or class.

Regression: Predicting a numerical value.

Overfitting: Learning training-specific noise so that performance on new data declines.

Generative AI: AI that produces new content such as text, images, audio or code.

Large language model: A model trained on large text datasets to predict and generate token sequences.

Prompt: An instruction or input provided to a generative AI system.

Retrieval-augmented generation: Generation supported by retrieving relevant external documents.

Natural language processing: Computational analysis and generation of human language.

Named-entity recognition: Identifying entities such as people, organisations, dates or locations in text.

Intelligent agent: A system that perceives an environment, selects actions and acts toward objectives.

Hallucination: An unsupported or false output presented in a plausible form.

Human oversight: Meaningful human ability and responsibility to review, challenge or control AI outputs.

Data drift: Change over time in the data or relationships on which a model depends.

Responsible AI: AI development and use consistent with safety, fairness, privacy, accountability and applicable obligations.

Review Questions

Short Answer Questions

1. What is the difference between supervised and unsupervised learning?
2. Define generative AI and give one business example.
3. What is the purpose of a prompt’s output-format instruction?
4. Identify two business uses of natural language processing.
5. Why do intelligent agents require permission controls?

Long Answer Questions

1. Explain the machine-learning workflow from problem definition through monitoring.
2. Compare predictive AI and generative AI, including their different risks.
3. Explain how retrieval-augmented generation can improve business responses and why it does not guarantee accuracy.
4. Discuss the principles of responsible AI adoption in an organisation.
5. Explain why human oversight must include competence, authority and access to evidence.

Analytical / Application Questions

1. A bank wants to use AI to screen loan applications. Identify relevant data, possible benefits, risks, performance measures and required human controls.
2. Design a prompt for an AI system that summarises supplier contracts for procurement managers. Specify context, constraints, evidence requirements and output format.
3. A retailer proposes an autonomous agent that changes prices hourly. Analyse the agent’s permissions, data requirements, failure modes, monitoring indicators and escalation rules.

Multiple Choice Questions with Answers

Question 1

Which statement best describes supervised learning?

A. It learns from examples with known target outcomes.

B. It operates without data or predefined objectives.

C. It only generates creative text.

D. It selects actions solely through random trial.

Correct Answer: A. It learns from examples with known target outcomes.

Question 2

If TP = 80 and FP = 20, what is the precision of the classifier?

A. 20%

B. 50%

C. 80%

D. 100%

Correct Answer: C. 80%

Question 3

What is a defining characteristic of generative AI?

A. It only stores historical records.

B. It creates new content from learned patterns.

C. It eliminates the need for human review.

D. It can only classify numerical data.

Correct Answer: B. It creates new content from learned patterns.

Question 4

Which prompt practice most directly improves traceability?

A. Requesting an answer with no context

B. Asking the model to be creative without constraints

C. Providing approved source material and requesting citations

D. Removing all output-format requirements

Correct Answer: C. Providing approved source material and requesting citations

Question 5

What is a common purpose of named-entity recognition?

A. Identifying people, organisations or dates in text

B. Calculating inventory carrying costs

C. Encrypting a database

D. Routing physical delivery vehicles

Correct Answer: A. Identifying people, organisations or dates in text

Question 6

What makes an intelligent agent different from a basic text generator?

A. It cannot receive external information.

B. It can perceive conditions, use tools and take authorised actions.

C. It always produces mathematically exact answers.

D. It does not need an objective.

Correct Answer: B. It can perceive conditions, use tools and take authorised actions.

Question 7

Which situation is an example of overfitting?

A. A model performs similarly on training and new data.

B. A model performs well on training data but poorly on new data.

C. A model uses a validation set.

D. A model is monitored after deployment.

Correct Answer: B. A model performs well on training data but poorly on new data.

Question 8

Why is human oversight especially important for a high-impact AI decision?

A. High-impact decisions have no possible errors.

B. Human review can provide accountability and a route to challenge errors.

C. Human oversight guarantees that the model is unbiased.

D. It removes the need for data governance.

Correct Answer: B. Human review can provide accountability and a route to challenge errors.

Question 9

Which measure is expressed in the same units as a regression target?

A. Mean absolute error

B. Precision

C. Recall

D. Accuracy

Correct Answer: A. Mean absolute error

Question 10

Which is the strongest first step in responsible AI adoption?

A. Deploying the most complex model available

B. Defining the business problem, affected stakeholders and success measures

C. Removing all human review

D. Using confidential data in a public tool

Correct Answer: B. Defining the business problem, affected stakeholders and success measures