Generative AI and Foundation Models

Introduction and Context

Generative artificial intelligence (generative AI) refers to computational systems that produce new content—such as text, images, audio, video, software code, designs, and synthetic data—in response to user instructions or other inputs. Unlike conventional predictive systems, which commonly classify, rank, or forecast existing categories, generative systems construct an output that did not previously exist in precisely that form.

For entrepreneurs, generative AI is important for two related reasons. First, it can reduce the cost and time required to perform knowledge-intensive activities such as market research, customer support, content creation, software development, product design, and internal analysis. Second, it changes the basis of competition. A startup may use the same public model as a large incumbent, so competitive advantage increasingly depends on proprietary data, workflow design, customer relationships, domain expertise, trust, and execution rather than access to a model alone.

A foundation model is a large model trained on broad and diverse data that can be adapted to many downstream tasks. Large language models (LLMs) are foundation models specialized in processing and generating language, although many also process code, images, audio, or video. Multimodal systems accept or produce more than one data modality. A model may therefore interpret a photograph, answer questions about a document, generate an image, or combine text and visual information.

Generative AI is powerful but not inherently reliable. It may produce fluent yet false claims, expose confidential information, reproduce protected material, or perform inconsistently as prompts change. Entrepreneurs must therefore treat it as a socio-technical system requiring appropriate data, controls, evaluation, human judgment, and business accountability.

Learning Objectives

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

1. Explain foundation models, LLMs, multimodal models, tokens, embeddings, training, and inference.
2. Construct effective prompts using role, context, task, constraints, examples, and output specifications.
3. Distinguish prompting, retrieval-augmented generation (RAG), fine-tuning, and tool-using agents.
4. Design a basic generative-AI workflow with evaluation and human oversight.
5. Assess hallucination, reliability, latency, cost, privacy, security, and intellectual-property risks.
6. Apply generative AI responsibly to an entrepreneurial problem and justify when human review is necessary.

1. Foundation Models and Large Language Models

1.1 From task-specific AI to general-purpose models

Traditional machine-learning systems are often trained for a relatively narrow purpose: detecting fraudulent transactions, predicting demand, or classifying customer complaints. A foundation model is trained in advance on broad data and then adapted through prompting, additional training, retrieval, or tools. The same underlying model may support summarisation, translation, reasoning assistance, coding, classification, and content generation.

This creates economies of scope: one model can support many tasks. It also creates risk concentration. A defect, bias, security weakness, or licensing problem in a widely used model can affect many products simultaneously. Entrepreneurs should distinguish the model provider, the application developer, and the end user. Responsibility can be distributed across these actors, but it is not eliminated by using an external API.

1.2 How language models generate text

An LLM represents text as tokens. A token may be a word, part of a word, punctuation mark, or special symbol. The model estimates a probability distribution over possible next tokens, conditioned on the preceding context. Generation proceeds iteratively: the selected token is appended to the context, and the model predicts the next token.

The model does not retrieve a sentence from a human-like memory in the ordinary sense. Its parameters encode statistical patterns learned during training. Consequently, a generated answer can be coherent without being factually grounded. A language model is best understood as a probabilistic sequence generator whose outputs require task-appropriate verification.

The temperature parameter changes the sharpness of token selection. Lower temperature generally makes output more predictable; higher temperature generally increases variation. Temperature is not a guarantee of truth: a low-temperature answer may still be confidently wrong.

1.3 Training, adaptation, and inference

Pre-training commonly uses self-supervised learning. For a sequence of tokens, the model learns to predict missing or subsequent tokens. The model parameters are adjusted to reduce a loss function over training examples. In simplified form, cross-entropy loss for one target token can be written as:

Here, x is the input context, y is the target token, pθ(y | x) is the probability assigned to the target by a model with parameters θ, and L is the loss for that example. The logarithm is natural logarithm in the usual machine-learning convention. The formula assumes a valid probability between 0 and 1 and measures predictive error; it does not directly measure factual accuracy, usefulness, or business value.

During inference, the trained parameters are held fixed while the system produces an output for a new input. Inference cost and latency depend on factors such as input length, output length, model size, hardware, provider pricing, and system architecture.

After pre-training, models may be adapted through supervised fine-tuning, preference optimisation, instruction tuning, or domain-specific training. These approaches can improve task behaviour but do not automatically make a model current, unbiased, or truthful.

1.4 Embeddings and semantic search

An embedding is a numerical vector representing an item such as a sentence, product description, image, or customer question. Similar meanings tend to be located near one another in the embedding space, although “similarity” reflects the training and embedding method rather than human understanding.

A common similarity measure is cosine similarity:

The vectors a and b represent two items, a · b is their dot product, and ||a|| and ||b|| are their Euclidean lengths. The measure is defined for non-zero vectors and is unit-free. It is often used for semantic retrieval, but a high similarity score is not proof that a document answers a question.

2. Prompting as Entrepreneurial Workflow Design

A prompt is an instruction and its supporting context. Effective prompting is not merely clever wording; it is the design of an interaction between a user, a model, business information, and an evaluation process.

A practical prompt structure contains:

• Role or perspective: the expertise or function required.
• Task: the specific action to perform.
• Context: relevant facts, audience, data, and assumptions.
• Constraints: length, tone, exclusions, legal or brand requirements.
• Examples: representative input-output pairs when format or classification matters.
• Output schema: headings, table fields, JSON keys, or decision categories.
• Quality criteria: evidence requirements, uncertainty disclosure, and conditions for refusal.

For example, a weak request is “Write a marketing plan.” A stronger request identifies the product, target segment, geography, budget, channel constraints, evidence standard, and required structure. The stronger prompt reduces ambiguity, but it does not eliminate the need to validate market claims.

2.1 Zero-shot and few-shot prompting

Zero-shot prompting asks a model to perform a task without examples. Few-shot prompting supplies a small number of examples showing the desired mapping or format. Examples are useful when a startup needs consistent tone, labels, or structured outputs. However, poor examples can transmit errors or unintended bias. Examples should therefore be representative, reviewed, and version-controlled.

2.2 Structured outputs and decomposition

Entrepreneurial tasks are often improved by decomposing them into stages: extract facts, classify the issue, generate alternatives, compare alternatives, and produce a recommendation. Decomposition makes errors easier to locate and enables separate checks. It should not be confused with a guarantee that the model’s hidden reasoning is correct. The entrepreneur should request concise, inspectable evidence or intermediate artefacts rather than relying on unsupported confidence.

Structured output—such as a fixed table or schema—supports automation. Yet syntactic validity is not semantic validity: a response can be valid JSON while containing an incorrect price, invented citation, or nonsensical category.

3. Retrieval-Augmented Generation

Retrieval-augmented generation combines information retrieval with generation. Instead of asking a model to answer solely from its parameters, the system retrieves relevant material from an approved knowledge base and places that material in the prompt. The model then generates an answer grounded, ideally, in the retrieved sources.

A typical RAG pipeline is:

1. Collect and govern documents.
2. Remove duplicates, obsolete versions, and unauthorised material.
3. Split documents into meaningful chunks.
4. Create embeddings for chunks and store them in an index.
5. Convert a user question into a search representation.
6. Retrieve and, where useful, rerank candidate passages.
7. Place passages and source identifiers in the model context.
8. Generate an answer with citations or traceable references.
9. Evaluate retrieval, answer quality, and safety.

RAG is particularly useful for product manuals, internal policies, changing prices, support knowledge, and regulatory guidance. It is not a universal solution. If the knowledge base is incomplete, contradictory, badly chunked, or unauthorised, retrieval may provide misleading material. The model can also misread a relevant passage or cite a source that does not support its conclusion.

A good RAG application instructs the system to answer only from supplied evidence when the task requires that restriction, identify insufficient evidence, distinguish quoted facts from inferences, and display document dates or versions. Access controls must be applied before retrieval, not merely after generation, because confidential text can be exposed through the model context.

4. Fine-Tuning and Model Selection

Fine-tuning updates model parameters using a curated dataset for a particular task or style. It may be appropriate when the desired behaviour is stable, repeated at high volume, difficult to express through prompts, or dependent on many labelled examples. Examples include classifying support tickets into a proprietary taxonomy or producing consistent outputs in a specialised format.

Fine-tuning is less suitable for frequently changing facts. A current price list or policy is normally better handled through retrieval. Fine-tuning also introduces risks of overfitting, memorisation, catastrophic forgetting, data leakage, and maintenance burden. The organisation must retain rights to training examples and ensure that personal data is processed lawfully.

The choice among prompting, RAG, fine-tuning, and a new model should be based on the problem rather than fashion. A useful decision sequence is:

Model selection should consider accuracy on the target task, context capacity, latency, availability, data handling, price, auditability, and vendor dependence. A larger model is not automatically better for every task.

5. Multimodal Systems and Agents

Multimodal systems process combinations of text, images, audio, video, or structured data. An entrepreneurial use case might involve analysing a photograph of damaged inventory, extracting fields from an invoice, generating a product concept from a sketch, or supporting a voice-based service desk.

Multimodal outputs require modality-specific controls. An image model may produce an attractive but unusable product design; an optical character recognition system may misread a decimal point; an audio system may misunderstand an accent. Sensitive biometric, health, or identity information requires heightened governance.

An agent is a system that uses a model to select steps, call tools, observe results, and continue toward a goal. Tools may include search, a database, a calculator, inventory software, a calendar, or a payment system. Agents introduce action risk: a wrong answer in a draft is different from a wrong answer that changes a customer’s account or sends money.

A safe agent design uses least-privilege permissions, allow-listed tools, input and output validation, transaction limits, audit logs, sandbox testing, confirmation for consequential actions, and a clear stop condition. The model should not be granted unrestricted access merely because it can produce natural language.

6. Evaluation: Hallucination, Reliability, and Cost

6.1 Hallucination and factuality

A hallucination is an output that is unsupported, fabricated, or materially incorrect, including invented citations, nonexistent products, false calculations, and misrepresented quotations. Hallucination is not simply a poor writing style. It is a failure relative to the task’s evidence requirements.

Evaluation should use a representative test set containing ordinary, difficult, ambiguous, adversarial, and out-of-scope cases. Useful measures include factual accuracy, completeness, citation correctness, refusal quality, consistency, latency, and user success. Automated metrics can assist, but human or expert review is necessary for high-impact domains.

For a binary test set, accuracy is:

Accuracy is appropriate only when the classes and error costs are reasonably balanced. In a support system where missing a safety complaint is much worse than escalating an ordinary complaint, precision, recall, cost-weighted measures, and human review may be more informative.

6.2 Reliability and uncertainty

Reliability means that a system performs acceptably under defined conditions, not that it is always correct. Reliability testing should examine prompt variation, document updates, long inputs, missing information, multilingual use, malicious instructions, and provider changes. Confidence language generated by a model should not be treated as calibrated probability unless calibration has been empirically demonstrated.

A production system should define thresholds for automatic handling, escalation, and refusal. For example, low-risk product FAQs may be automated, while legal commitments, medical advice, credit decisions, employment decisions, and payments require specialist review or tightly controlled workflows.

6.3 Cost and unit economics

The cost of an AI feature includes model charges, embedding and storage costs, retrieval infrastructure, engineering, monitoring, human review, security, compliance, support, and failure remediation. A simplified variable-cost estimate is:

Here, n_i and n_o are input and output token counts, c_i and c_o are the corresponding per-token prices, and C_fixed includes fixed or non-token costs over the chosen period. Units must be consistent—for example, if a provider quotes price per million tokens, token counts must be divided by 1,000,000 before multiplication. This estimate excludes unexpected retries, failed calls, human review, and infrastructure unless they are included in C_fixed.

Worked example. Suppose one interaction uses 2,000 input tokens and 500 output tokens. The prices are $2 per million input tokens and $8 per million output tokens. The model cost is:

For 10,000 interactions, the model-only cost is $80, before infrastructure and review. If 5% of interactions require four minutes of human review at $24 per hour, review cost is 500 × (4 ÷ 60) × $24 = $800. Thus, ignoring other costs, the combined estimate is $880. The example demonstrates why unit economics must include human oversight rather than comparing API prices alone.

7. Intellectual Property, Privacy, and Governance

Entrepreneurs should identify what data enters a model, where it is processed, how long it is retained, whether it is used for provider training, and who can access outputs. Personal data requires a lawful and transparent processing basis, purpose limitation, data minimisation, appropriate security, and rights management under applicable law. Sensitive data should not be pasted into an unapproved public tool.

Intellectual-property risks include unauthorised training data, copyrighted input documents, output similarity, trademarks, confidential information, and unclear ownership or licensing terms. An output being newly generated does not guarantee that it is free of legal risk. Businesses should maintain provenance records, use licensed or authorised data, review provider terms, avoid false claims of originality, and obtain human review for commercially important material.

Security threats include prompt injection, data exfiltration, insecure tool use, model manipulation, poisoned retrieval documents, excessive agency, and indirect instructions embedded in webpages or files. A retrieved document should be treated as data, not as an authorised command. System instructions, user content, and retrieved content should be separated and assigned explicit trust levels.

Governance should assign an accountable owner and define:

• permitted and prohibited use cases;
• data classification and retention rules;
• testing before deployment and after model updates;
• human-approval requirements;
• incident reporting and rollback procedures;
• documentation of model, prompt, data, and version changes;
• accessibility, bias, and non-discrimination checks.

Relevant legal obligations vary by jurisdiction and application. Entrepreneurs should obtain qualified legal advice rather than assuming that a general AI policy satisfies sector-specific requirements.

8. Illustrative Case: GreenBasket’s AI Support Service

Background

GreenBasket, a fictional online retailer of sustainable household products, receives approximately 8,000 monthly customer-support requests. Questions concern delivery, product composition, returns, and product suitability. The founders want to reduce response time without making unsupported environmental or safety claims.

Problem

A prototype chatbot answers quickly but invents certification labels and sometimes relies on outdated return rules. The founders must decide whether to deploy it, redesign it, or abandon the project.

Analysis

The problem is not solved by selecting a larger model alone. Product information and policies change, so current documents are needed. Certification claims have reputational and legal consequences. Returns may involve refunds, making unrestricted automated action inappropriate. The team therefore separates low-risk information from consequential decisions.

Approach and solution

GreenBasket creates an approved knowledge base containing versioned product specifications, certification evidence, delivery information, and return policies. Documents are chunked with titles, dates, and source identifiers. A RAG system retrieves relevant passages and requires citations in draft answers. The prompt instructs the model to say when evidence is insufficient and never infer certification from marketing language.

The system classifies requests into informational, complaint, safety-related, and transactional categories. Informational answers can be sent automatically only after evaluation. Safety-related questions are escalated to a trained employee. Refunds and account changes require human confirmation. The company tests 300 historical and synthetic cases, including outdated documents, ambiguous product names, prompt injection, and deliberately unsupported claims.

Outcome

In the pilot, response time falls for routine questions, while escalations increase for ambiguous cases. The business does not measure success only by automation rate; it tracks grounded-answer rate, citation correctness, customer resolution, complaint rate, cost per resolved request, and harmful-error incidents. A monitoring process pauses deployment if a new model or knowledge-base update causes unacceptable regression.

Learning points

The case illustrates that value comes from workflow redesign, not text generation alone. RAG supports currency but depends on document governance. Human oversight should be concentrated where error consequences are high. Evaluation must measure business outcomes and harm, and a refusal or escalation can be a successful system behaviour.

9. Best Practices for Entrepreneurial Adoption

1. Begin with a specific user problem and measurable outcome, not with a model.
2. Map data flows before selecting a provider or uploading information.
3. Use the least powerful and least autonomous system that can meet the requirement.
4. Prefer retrieval for changing facts and fine-tuning for stable specialised behaviour.
5. Build a representative evaluation set before public launch.
6. Test both ordinary use and adversarial misuse.
7. Require citations, uncertainty statements, or human review where evidence matters.
8. Log prompts, retrieved sources, model versions, outputs, approvals, and incidents subject to privacy rules.
9. Calculate full cost per successful outcome, including review and failure costs.
10. Keep a fallback process so customers are not trapped by an unreliable system.
11. Review accessibility, cultural context, bias, and language coverage.
12. Re-evaluate after changes to models, prompts, tools, policies, or data.

Contemporary Developments and Strategic Implications

Current development is moving toward smaller specialised models, multimodal interfaces, longer-context systems, model routing, on-device inference, synthetic-data workflows, and increasingly capable tool use. These developments may reduce cost or latency, but they do not remove the need for governance. Longer context can increase the amount of irrelevant or malicious material a system receives. Autonomous agents can increase productivity while also increasing the scale of mistakes.

For startups, defensibility may arise from proprietary workflows, high-quality permissioned data, integration into customer operations, measurable reliability, and trusted domain expertise. A thin interface over a general model is easy to imitate. A well-governed system that solves a costly problem and earns user trust is more difficult to replace.

Chapter Summary

Foundation models are broad, adaptable models that support many tasks, while LLMs generate language token by token and multimodal systems work across several data types. Their fluency should not be confused with factuality. Prompting improves task specification; RAG supplies current or private evidence; fine-tuning adapts stable behaviours; and agents connect models to tools and actions.

Entrepreneurs should select an approach according to the business problem, data conditions, risk, cost, and required control. Evaluation must measure accuracy, groundedness, citation quality, refusal behaviour, latency, cost, and user outcomes. Hallucination, privacy breaches, prompt injection, excessive autonomy, intellectual-property disputes, and vendor dependence are material risks. Human oversight, least-privilege access, auditability, version control, and escalation procedures are essential for consequential applications. Sustainable advantage comes from reliable workflows, proprietary knowledge, domain expertise, and trust—not from model access alone.

Key Terms

Foundation model: A broadly trained model adaptable to many downstream tasks.

Large language model (LLM): A model trained to process and generate language tokens.

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

Token: A unit of text processed by a language model.

Inference: The process of using a trained model to produce an output.

Embedding: A numerical vector representation used to compare items by semantic relationship.

Prompt: An instruction and context supplied to a generative model.

Few-shot prompting: Prompting that includes examples of the desired task or format.

Multimodal model: A model that processes or generates more than one data modality.

Retrieval-augmented generation (RAG): A method that retrieves external material before generation.

Fine-tuning: Additional training that adapts model parameters to a specialised task.

Agent: A model-based system that selects steps and uses tools to pursue a goal.

Hallucination: Unsupported, fabricated, or materially incorrect generated content.

Grounding: Connecting an answer to reliable, relevant evidence.

Prompt injection: An attempt to manipulate system behaviour through malicious or conflicting instructions.

Human-in-the-loop: A design in which a person reviews, approves, or corrects system activity.

Model drift: A change in performance caused by changing data, models, users, or environments.

Intellectual property: Legal rights relating to creations, inventions, brands, and confidential information.

Least privilege: Granting only the access and permissions necessary for a task.

Review Questions

5 Short Answer Questions

1. What distinguishes a foundation model from a task-specific machine-learning model?
2. Why can an LLM produce a fluent but false answer?
3. What problem does RAG address, and what problem does it not automatically solve?
4. State two controls appropriate for a tool-using AI agent.
5. Why should human review costs be included in an AI system’s unit economics?

5 Long Answer Questions

1. Explain the differences among prompting, RAG, fine-tuning, and agent-based design. Give one entrepreneurial use case for each.
2. Describe a complete evaluation plan for a customer-support LLM, including test data, metrics, escalation, and monitoring.
3. Analyse the privacy, intellectual-property, and security risks of using a public generative-AI service in a startup.
4. Explain how multimodal systems expand entrepreneurial opportunities and why their risks differ from text-only systems.
5. Discuss why proprietary data and workflow integration may create stronger competitive advantage than access to a general-purpose model.

3 Analytical / Application Questions

1. A startup wants an AI assistant to answer questions from policies that change weekly. Should it begin with fine-tuning, RAG, or prompting alone? Justify the choice and specify controls.
2. Design a safe agent for processing supplier invoices. Identify its tools, permissions, validation checks, human-approval points, and failure-recovery process.
3. Using the cost equation in this chapter, estimate the model cost for 25,000 interactions if each uses 1,600 input tokens and 400 output tokens, with prices of $1.50 per million input tokens and $6 per million output tokens. Explain which additional costs would be needed for a business decision.

Multiple Choice Questions with Answers

Question 1

Which statement best describes a foundation model?

A. A model trained only for one narrowly defined classification task

B. A broadly trained model that can be adapted to multiple downstream tasks

C. A database containing only verified business facts

D. A rule-based system that cannot generate content

Correct Answer: B. A broadly trained model that can be adapted to multiple downstream tasks

Question 2

What is the primary purpose of retrieval-augmented generation?

A. To replace all human review

B. To increase model temperature

C. To supply relevant external evidence to support generation

D. To permanently change the model’s parameters

Correct Answer: C. To supply relevant external evidence to support generation

Question 3

Which use case is generally most suitable for fine-tuning?

A. A frequently changing price list

B. A stable specialised classification behaviour used at high volume

C. A one-time request requiring current legal information

D. A confidential document that the organisation has no right to use

Correct Answer: B. A stable specialised classification behaviour used at high volume

Question 4

What does cosine similarity compare?

A. The legal ownership of two documents

B. The semantic orientation of two non-zero vectors

C. The financial cost of two model calls

D. The number of tokens in two prompts only

Correct Answer: B. The semantic orientation of two non-zero vectors

Question 5

Which control is most appropriate for an AI agent that can issue refunds?

A. Unlimited permissions to reduce latency

B. Removal of transaction logs

C. Human confirmation and transaction limits

D. Higher temperature for more creative decisions

Correct Answer: C. Human confirmation and transaction limits

Question 6

Using the chapter’s cost formula, what is the model-only cost of one interaction using 2,000 input tokens at $2 per million and 500 output tokens at $8 per million?

A. $0.0008

B. $0.008

C. $0.08

D. $8.00

Correct Answer: B. $0.008

Question 7

Which statement about hallucination is correct?

A. It means the output is grammatically incorrect

B. It occurs only when temperature is high

C. It is unsupported, fabricated, or materially incorrect generated content

D. It is eliminated by using a larger model

Correct Answer: C. It is unsupported, fabricated, or materially incorrect generated content

Question 8

Why is a valid structured format not sufficient evidence that an AI answer is correct?

A. Structured formats cannot be generated by language models

B. Syntactic validity does not guarantee semantic or factual validity

C. Tables always increase hallucination

D. Models cannot process schemas

Correct Answer: B. Syntactic validity does not guarantee semantic or factual validity

Question 9

Which practice best protects confidential information in a RAG system?

A. Retrieve all documents for every user

B. Apply access controls before documents enter the model context

C. Publish the knowledge base publicly

D. Treat retrieved text as trusted instructions

Correct Answer: B. Apply access controls before documents enter the model context

Question 10

For a weekly changing policy knowledge base, which approach is usually the best starting point?

A. RAG with versioned and governed documents

B. Fine-tuning once and never updating it

C. Prompting without supplying policy content

D. Allowing the model to invent missing policies

Correct Answer: A. RAG with versioned and governed documents