Playbook

Differentiating between AI types for a business use case.

→Use Generative AI for content creation (text, images, code). Use Traditional/Discriminative AI for classification, prediction, and analysis of existing data.

Why: Generative AI *creates* new, novel content. Traditional AI *analyzes* or *categorizes* existing data. This is a fundamental concept.

Deciding whether to build a model from scratch or use a pre-trained one.

→Leverage foundation models (e.g., Gemini) that are pre-trained on massive, diverse datasets and adapt them to specific tasks.

Why: Foundation models drastically reduce development time and resource costs by providing a powerful, general-purpose base that can be specialized via prompting or fine-tuning.

A solution needs to understand and process combinations of text, images, audio, or video.

→Use a multimodal foundation model like Gemini, which can natively reason across different data types in a single prompt.

Why: Multimodal models avoid the complexity of stitching together separate models for each data type, enabling more sophisticated, cross-domain understanding.

An LLM is confidently generating plausible but factually incorrect information.

→Implement grounding techniques, primarily Retrieval-Augmented Generation (RAG), to connect the model to verifiable data sources.

Why: Hallucinations are an inherent risk. Grounding anchors the model’s responses in a source of truth, making it the primary strategy for improving factual accuracy.

Understanding the core technology that enables modern LLMs to understand context.

→The transformer architecture, with its self-attention mechanism, allows the model to weigh the importance of all words in the input relative to each other.

Why: Self-attention is the key innovation that enables LLMs to grasp long-range dependencies and context, unlike older sequential models (RNNs).

Building a search system that understands the meaning behind queries, not just keywords.

→Use an embedding model (e.g., from Vertex AI) to convert text into numerical vectors. Store these vectors and use vector similarity search to find semantically related content.

Why: Embeddings capture semantic meaning. Queries can find conceptually similar results even if they don’t share keywords.

A creative application needs varied outputs, while a factual chatbot needs deterministic responses.

→Increase the `temperature` parameter (e.g., >0.7) for creative tasks. Decrease `temperature` (e.g., <0.3) for factual, consistent responses.

Why: Temperature controls the randomness of the output. Low temperature picks the most likely words; high temperature increases variety.

Processing a large document that exceeds the model’s token limit.

→Architect a solution using chunking, summarization, or a RAG approach to process the document in manageable pieces that fit within the context window.

Why: Models have a finite context window. Any input exceeding this limit is ignored, leading to information loss. The architecture must account for this.

Discovering, testing, and deploying a variety of foundation models on Google Cloud.

→Use Vertex AI Model Garden as the central catalog for Google’s proprietary models (Gemini), open-source models (Llama, Mistral), and partner models.

Why: Model Garden is the unified entry point for accessing a curated set of foundation models, simplifying discovery and deployment within an enterprise-grade environment.

Reference↗

An AI assistant needs to answer questions about frequently changing information, like product inventory or recent news.

→Implement a Retrieval-Augmented Generation (RAG) pattern. Connect the LLM to an external, up-to-date knowledge base (e.g., a database, document store).

Why: RAG allows the model to access real-time information at inference time, overcoming its knowledge cutoff and providing accurate, current answers.

Building an enterprise search engine or a conversational AI agent grounded in company data.

→Use Vertex AI Search and Conversation (part of Agent Builder). Point it to your data sources (websites, documents) to create a search app or chatbot.

Why: This is a managed, low-code solution for building grounded, enterprise-grade search and chat applications, significantly reducing development complexity.

A model needs to learn a highly specialized skill, terminology, or consistent behavior that prompting alone cannot achieve.

→Perform supervised fine-tuning on a foundation model using a curated dataset of high-quality examples.

Why: Fine-tuning adapts the model’s internal weights, making it an expert in a specific domain. It is more powerful than prompting for deep specialization.

Need to customize a foundation model for a specific domain but lack the resources for full fine-tuning.

→Use a Parameter-Efficient Fine-Tuning (PEFT) method like LoRA or adapter tuning available in Vertex AI.

Why: PEFT tunes only a small fraction of the model’s parameters, achieving significant customization with drastically lower computational cost and time.

A model is failing at tasks that require complex, multi-step reasoning (e.g., math problems, logic puzzles).

→Use chain-of-thought (CoT) prompting. Instruct the model to "think step by step" before giving the final answer.

Why: CoT encourages the model to break down a problem, which has been shown to significantly improve its reasoning ability and final answer accuracy on complex tasks.

The model needs to consistently generate output in a specific format (e.g., JSON, a certain writing style).

→Use few-shot prompting. Provide 2-5 examples of the desired input-output pattern directly in the prompt.

Why: Providing examples is more effective than just describing the format. The model learns the pattern and applies it to the new request.

Choosing the right Gemini model variant for a specific use case.

→Use Gemini Pro for complex, high-quality reasoning. Use Gemini Flash for high-volume, low-latency, and cost-sensitive tasks. Use Gemini Nano for on-device applications.

Why: Selecting the right model size is a critical trade-off between capability, speed, and cost. Using the smallest model that meets the requirement is a best practice.

Automating the extraction of structured data (e.g., line items, dates, totals) from unstructured documents like invoices or receipts.

→Use Google Cloud Document AI. Utilize its pre-trained processors for common document types or build a custom processor for unique formats.

Why: Document AI is a purpose-built service that goes beyond simple OCR to understand document structure and semantics, providing much higher accuracy for data extraction tasks.

Reference↗

Applying generative AI capabilities (e.g., summarization, sentiment analysis) to data stored in a BigQuery data warehouse.

→Use BigQuery ML to call Vertex AI foundation models directly with SQL commands. Process the data in place without moving it.

Why: This simplifies the architecture, improves security by keeping data within BigQuery, and allows data analysts to leverage AI using familiar SQL syntax.

Boosting productivity for business users within their existing tools like Gmail, Docs, and Sheets.

→Integrate Gemini for Google Workspace. This provides AI assistance directly inside the Workspace applications for tasks like drafting emails, summarizing documents, and analyzing data.

Why: This brings AI capabilities to users in their familiar workflow, accelerating adoption and providing immediate productivity benefits without context switching.

Improving developer velocity and code quality.

→Provide developers with Gemini Code Assist, which integrates into IDEs to offer code completion, generation, explanation, and test creation.

Why: AI code assistants reduce time spent on boilerplate code, help understand complex codebases, and improve overall developer productivity.

Choosing the right tool for generative AI experimentation and development.

→Use Google AI Studio for quick, no-cost web-based prototyping with Gemini models via an API key. Use Vertex AI Studio for enterprise-grade development with GCP integration, security controls, and MLOps capabilities.

Why: Google AI Studio is for rapid prototyping; Vertex AI Studio is the path to production, offering enterprise security, data governance, and scalability.

An AI agent needs to adopt a specific persona, follow rules, and maintain a consistent tone across conversations.

→Define the agent's behavior using a system prompt. This instruction is provided to the model separately from the user query to guide its overall conduct.

Why: A system prompt is the most effective way to establish durable, consistent behavioral guidelines without having to repeat them in every user-facing prompt.

A solution requires a common, specific AI capability like translation, speech-to-text, or text-to-speech.

→Use the pre-trained, purpose-built APIs: Cloud Translation API, Speech-to-Text API, or Text-to-Speech API.

Why: These managed APIs are highly optimized for their specific task and are more cost-effective and simpler to implement than using a general-purpose LLM for the same function.

An AI system is used for a high-stakes process where errors are costly or dangerous (e.g., medical summaries, financial reports).

→Implement a Human-in-the-Loop (HITL) workflow. AI generates a draft, which is then reviewed, edited, and approved by a human expert.

Why: HITL combines AI’s speed with human judgment and accountability, which is essential for mitigating risks in critical applications.

An AI model’s performance degrades over time after being deployed to production.

→Implement continuous monitoring to track model performance and detect data drift or concept drift.

Why: The real world changes. Data drift occurs when production data no longer resembles the training data. Monitoring is critical to know when retraining or updating is necessary.

Forecasting and managing the operational cost of a generative AI service.

→Understand that Vertex AI GenAI services are priced on a pay-per-use basis, typically per 1,000 input and output characters or tokens.

Why: Cost is directly tied to usage. Architects must design systems to manage prompt and response lengths to control operational expenses.

An AI application experiences high latency or errors during peak user traffic.

→Scale the model deployment. For Vertex AI Prediction endpoints, increase the number of machine replicas or use higher-performance machine types.

Why: Inference performance is not infinitely scalable. The underlying infrastructure must be provisioned to handle the expected request volume.

A generative AI solution must process sensitive data that is subject to regional data sovereignty regulations (e.g., GDPR).

→Configure Vertex AI to use regional endpoints. Integrate with VPC Service Controls to create a service perimeter that prevents data exfiltration.

Why: Google Cloud provides explicit controls to ensure data is processed within a specific geographic region and is isolated from public networks, which is mandatory for many compliance regimes.

An application handles a mix of simple and complex queries, and using a single large model is cost-prohibitive.

→Implement a model router. Pre-classify incoming prompts and route simple requests to a small, fast, and cheap model (e.g., Gemini Flash) and complex requests to a powerful model (e.g., Gemini Pro).

Why: This pattern optimizes the cost-performance trade-off by using the most appropriate resource for each task, significantly reducing overall operational costs.

Launching a new generative AI initiative within the organization.

→Start by identifying a high-value business problem or use case. Do not start with the technology and search for a problem.

Why: Successful AI projects are those that deliver measurable business value. A clear problem statement ensures focus and aligns the project with strategic goals.

An AI model is showing biased behavior against certain demographic groups.

→Address bias throughout the ML lifecycle: audit and curate training data for fairness, test the model for disparate impacts, and implement post-deployment monitoring for biased outcomes.

Why: Bias primarily originates from the data. It cannot be fixed with a single technical solution; it requires a comprehensive, ongoing process of testing and mitigation.

An enterprise needs to scale its use of AI responsibly across multiple departments.

→Establish a cross-functional AI governance committee. Create clear policies for AI development, risk assessment, ethical review, deployment, and monitoring.

Why: Centralized governance ensures consistency, manages risk, and promotes the responsible use of AI, preventing a chaotic "wild west" of unmanaged AI projects.

Preventing a public-facing chatbot from generating harmful, hateful, or inappropriate content.

→Enable the built-in safety filters in Vertex AI. Configure thresholds for categories like hate speech, harassment, and dangerous content.

Why: These pre-trained classification models provide a critical first line of defense against generating unsafe content, forming a core part of responsible AI deployment.

Justifying an AI investment to executive leadership.

→Measure ROI holistically. Track efficiency metrics (e.g., time saved, cost reduction) and effectiveness metrics (e.g., revenue lift, quality improvement, customer satisfaction).

Why: A comprehensive ROI analysis goes beyond just cost savings to capture the full business value, including improvements in quality and new revenue opportunities.

Deploying an AI system in a regulated industry (e.g., finance, healthcare) that requires decision transparency.

→For traditional ML, use Vertex AI Explainability. For GenAI, use RAG with source attribution to provide citations and justifications for generated answers.

Why: Transparency builds trust and is a legal requirement in many domains. Providing citations for GenAI responses is the primary method for explainability.

Developing a corporate strategy for securing AI systems from new types of threats.

→Adopt the principles of Google's Secure AI Framework (SAIF), which provides recommendations for securing the AI supply chain, model, and deployment.

Why: SAIF offers a structured, conceptual guide for extending traditional cybersecurity practices to the unique challenges of AI, such as prompt injection and data poisoning.

Reference↗

Rolling out a new AI tool to the workforce to ensure successful adoption.

→Implement a structured change management program. Secure executive sponsorship, communicate clearly about AI’s role, provide comprehensive training, and integrate AI into existing workflows gradually.

Why: Technology is only part of the solution. Successful AI adoption depends on people and process, requiring deliberate effort to build skills, trust, and new ways of working.

Using customer data to train or run a generative AI model.

→Ensure strict compliance with data privacy regulations (e.g., GDPR). Use data minimization principles, anonymize PII where possible, and carefully review the AI provider’s data-use policies.

Why: Using customer data with AI creates significant privacy and compliance risks. Data governance and privacy must be core design considerations from the start.