Prepare and test your skills
Prepare and test your skills
Business value from generative AI (opens in a new tab) comes from using data to reach measurable business goals, not from the technology alone. Competitive differentiation grows when a business owns high-quality, varied data that makes models more accurate or enables features competitors cannot easily copy. Data quality, data variety, and proprietary datasets therefore link directly to strategic value, operational risk, and governance.
Data quality, data variety, and proprietary data are strategic assets because they decide how well an enterprise AI solution performs. Data quality is the accuracy, completeness, and relevance of data used to train, tune, and evaluate a model. Data variety is the mix of data types a model can process, including text, images, video, audio, and structured tabular data. A proprietary dataset is data a business owns or uniquely collects, and it becomes a strategic asset when it improves model effectiveness or supports capabilities competitors cannot easily copy.
Traditional AI (opens in a new tab) is particularly effective with structured data, while generative AI models can process multiple modalities such as text, images, videos, or audio. If input data or prompts are inaccurate or incomplete, the model can perform poorly or produce false outputs. High-quality, varied data therefore raises accuracy and trust, while poor data creates operational risk and weakens AI-driven decisions.
The business implications of data as an asset include:
Exam tip: A model is only as good as the data used; inaccurate or incomplete input data can lead to suboptimal performance or false model outputs.
A business-value-driven AI use case starts with measurable business goals and works backward to the AI solution. Goals may include increased efficiency, cost reduction, enhanced customer experience, or competitive advantage. Return on investment (ROI) is a key measure of AI project success and is tracked through direct financial gains, operational efficiency, and customer experience.
The decision process follows five ordered steps:
Stakeholders shape value at each step. Business leaders approve alignment with business priorities, product owners influence model behavior, domain experts improve model effectiveness, and end users must understand model output to use it for decision-making.
Data availability and variety determine whether a business uses a pre-trained model, custom training, prompt engineering, or model tuning. If a pre-trained traditional AI model meets requirements, the business uses it. If it does not, the business checks whether enough training data exists to custom train a model.
The choice between traditional AI and generative AI depends on data strengths and business fit:
| AI type | Data strength | Example use cases | Business fit |
|---|---|---|---|
| Traditional AI | Structured data such as tabular data and images | Classification, regression, time series forecasting, clustering | High control of training, explainability, and custom model behavior |
| Generative AI | Multiple modalities such as text, images, videos, and audio | Summarization, content generation, advanced transcription, chatbots | Faster go-to-market and specialized tuning with data examples |
When sufficient training data is available, the business chooses between more control and faster go-to-market (GTM). If high control is required, such as preferred algorithms, custom loss functions, explainability, model layers, learning rate, or hyperparameters, the business uses custom training of a traditional AI model. If faster GTM is the priority, the business uses generative AI and improves specialized tasks through tuning such as supervised tuning for classification, sentiment analysis, or entity extraction. If a training dataset is unavailable or too small, the business uses generative AI models with prompt engineering and tunes further with data examples.
Model tuning requires building a training dataset and selecting a method such as supervised tuning, reinforcement learning from human feedback (RLHF) tuning, or model distillation. Evaluation requires an evaluation dataset of prompt and ground truth pairs, with diverse examples that align with the task. Deployment then associates an endpoint and physical machine resources with the model for online, low-latency predictions. Tuned models must be deployed, while Google's foundation models available in generative AI on Vertex AI (opens in a new tab) already have endpoints, and deployment choices trade off traffic, latency, and budget between fully managed and self-managed environments.
Data governance for generative AI (opens in a new tab) is the practice of setting the rules, roles, and review steps that control how data is used, who approves it, and which legal or industry limits apply. Governance is not a single checkpoint; it runs through the whole AI planning process, from stating the business goal to changing the workflows that the AI solution touches. Privacy limits, stakeholder approvals, and data risks all shape which use cases move forward.
Regulatory and privacy constraints are business constraints that must be identified before a use case is approved. When a team defines its business goal and success criteria, it must also confirm that security and privacy aspects meet specific industry compliances or country regulatory requirements. These constraints sit alongside measurable goals such as increased efficiency, cost reduction, better customer experience, and competitive advantage. A use case can be blocked by privacy or compliance limits even when the financial case looks strong.
AI governance depends on more than the data scientists and machine learning engineers who commonly lead model selection. Four other groups hold distinct responsibilities, and each group reviews a different part of the decision. Approval flows from business leaders while model behavior is steered by product owners and domain experts.
| Stakeholder | Responsibility in the governance process |
|---|---|
| Business leaders and decision-makers | Approve the selection when it aligns with business priorities |
| Product owners | Influence or take more control of model behavior to align it with product priorities |
| Domain experts | Apply domain expertise to improve model effectiveness |
| End users | Understand the model's output and how to incorporate it for more informed decision-making |
Because these roles control different gates, a use case can be technically sound and still be blocked at approval if it conflicts with business priorities or regulatory constraints.
Data availability is a common governance risk because a business may not have enough usable data to train a model. When a training dataset is unavailable, or the available datasets are not large enough to custom train a model, the recommended path is to use generative AI models with prompt engineering (opens in a new tab) instead of custom training. Those models can then be tuned further for specialized tasks by using data examples, such as supervised tuning for classification, sentiment analysis, or entity extraction.
The choice also depends on what the business values more. If the priority is high control over training, such as choosing the algorithm, writing custom loss functions, using specific explainability features, and setting layers, learning rate, and other hyperparameters, then custom training of a traditional AI model is the fit. If the priority is faster go-to-market, generative AI is the fit.
Exam tip: When training data is missing or too small to custom train a model, the documented path is prompt engineering with generative AI, not custom training.
Generative AI can act as a governance tool by scanning business data for problems. It extracts and analyzes content from reports, invoices, receipts, financial transactions, and contracts to highlight possible errors or compliance issues, identify potential risks, and uncover anomalies that indicate fraud. It can analyze sentiment in user-generated content such as social media posts and product reviews, and it can process transcribed call center conversations to surface patterns such as the most common reasons customers give low ratings.
The same approach applies to security data. Generative AI analyzes cybersecurity data such as threat reports, articles, and repositories to extract key threat indicators. This analysis supports proactive defense by summarizing and prioritizing mitigation strategies. It can translate complex attack graphs into plain-text explanations of exposure, simulate possible attack paths to highlight impacted assets, and recommend mitigations before those assets can be exploited.
Return on investment (ROI) is one of the key measures of AI project success and shows whether governance and investment choices worked. ROI is tracked through three categories of metrics. Direct financial gains cover increased revenue or reduced costs. Operational efficiency covers faster time-to-market or quicker issue resolution. Customer experience covers increased satisfaction scores or improved retention. Governance decisions feed into these measures because privacy and compliance constraints shape which use cases are approved in the first place.
Adapting business processes is the final governance step that puts the approved AI use case into daily work. The business must identify the changes it has to make to existing processes or workflows to adapt to the generative AI or traditional AI use case. This step closes the loop because the approved model, its privacy constraints, and its stakeholder owners all feed back into how work is actually performed.
Data types determine which kind of AI (opens in a new tab) to use and how a model learns to produce relevant results. Structured data fits neatly into fixed rows and columns, semi-structured data mixes labeled fields with free text, and unstructured data includes free-form text, images, video, and audio. Generative AI models learn patterns from large collections of these types and use additional company data to stay accurate and current.
A hierarchy splitting AI data into structured, semi-structured, and unstructured types, showing each type's examples and which AI approach it serves: structured data trains traditional AI for prediction, semi-structured data powers extraction and fraud detection, and unstructured multi-modal data trains generative AI for summarization and content creation.
The three data types differ in form, examples, and primary role:
| Data type | What it looks like | Examples in AI work | Primary role |
|---|---|---|---|
| Structured data | Fixed rows, columns, and values | Tabular data, purchase history, house features and location, sales records | Train traditional AI to classify and predict |
| Semi-structured data | Labeled fields mixed with free text | Invoices, receipts, contracts, emails, financial transactions | Extract fields, check compliance, and detect anomalies |
| Unstructured data | No fixed schema, multiple modalities | Documents, articles, reviews, social posts, call transcripts, threat reports, product images, video, audio | Train and ground generative AI for summary and creation |
Structured data is data organized in a fixed format such as tabular data with defined rows and columns. Common examples include historical purchase data, house features and location, and sales and demand records. Traditional predictive AI is particularly effective with structured data because the model can learn clear historical patterns from these consistent fields. That training supports classification, regression, forecasting, and clustering, such as filtering email as spam or not spam, predicting house prices or customer revenue, forecasting sales, and performing customer segmentation.
Semi-structured data is data that combines identifiable fields with free-form content, such as invoices, receipts, contracts, and financial transactions. An invoice may contain set fields like dates and totals alongside variable descriptions, notes, and layouts. Generative AI supports document and data understanding on this type by extracting data and analyzing content to highlight errors or compliance issues, identify risks, and uncover anomalies that indicate fraud. Because the format is partly consistent and partly variable, this data is also useful for focused tuning such as entity extraction and sentiment analysis.
Unstructured data is data without a fixed table structure, including text, images, video, and audio. Examples include documents, articles, customer feedback, social media posts, product reviews, transcribed call center conversations, threat reports, and creative media. Foundation models (opens in a new tab) are large models trained on vast amounts of this multi-modal data, learning language patterns for text and diffusion patterns for images. Generative AI uses that learning for summarization, content generation, advanced transcription, conversational search, and code assistance, including summaries from multiple sources and answers that combine text with related images.
Exam tip: Use traditional AI when the task is prediction or classification on structured data, and use generative AI when the task needs summarization, creation, or reasoning across text, images, video, or audio.
Training, fine-tuning, and grounding are three ways data shapes a generative AI model to keep outputs contextually relevant. Training builds the broad capability of a foundation model from vast amounts of multi-modal data. Fine-tuning adapts that base model to a specialized task by training it further on curated data examples, such as company terminology for classification, sentiment analysis, or entity extraction. Grounding connects the model's output at response time to verifiable sources like internal documentation, project data, and communication records, which reduces hallucinations.
Grounding often uses retrieval-augmented generation (RAG) (opens in a new tab), which relies on embeddings and vector search to add facts to the prompt:
Structured data uses fixed rows, columns, and values, semi-structured data mixes labeled fields with free text, and unstructured data has no fixed schema and spans multiple modalities such as text, images, video, and audio. Structured data such as tabular data and sales records trains traditional AI to classify and predict, semi-structured data such as invoices, receipts, contracts, and financial transactions supports field extraction, compliance checks, and anomaly detection, and unstructured data such as documents, reviews, call transcripts, and threat reports trains and grounds generative AI for summary and creation.
Use traditional AI when the task is prediction or classification on structured data such as classification, regression, time series forecasting, and clustering. Use generative AI when the task needs summarization, content generation, or reasoning across multiple modalities such as text, images, video, or audio. Traditional AI offers high control of training, explainability, and custom model behavior, while generative AI offers faster go-to-market and specialized tuning with data examples.
When training data is unavailable or too small to custom train a model, the documented path is to use generative AI models with prompt engineering instead of custom training. Those models can then be tuned further for specialized tasks by using data examples, such as supervised tuning for classification, sentiment analysis, or entity extraction.
Retrieval-augmented generation grounds responses by connecting the model output at response time to verifiable sources, which reduces hallucinations. It converts the user query into an embedding, compares that embedding to document embeddings where closer proximity means similar meaning, and retrieves the most semantically relevant information to add to the prompt so the response is grounded in facts and current information.
A retail organization wants to deploy an enterprise generative AI solution to assist customer service representatives and optimize supply chain operations. While standard public foundation models provide broad language capabilities, leadership wants to establish a sustainable competitive advantage and ensure responses reflect the company's specific product catalog, policies, and operational history.
How does integrating high-quality proprietary datasets into their generative AI strategy deliver business value and competitive differentiation?
Last updated