Module 1 is the vocabulary module. It sets up the distinction the rest of the course keeps leaning on — gen AI is a technology that gets built into products, not a product itself — and then gives you the two things that make that technology usable: foundation models, which supply broad capability, and prompts, which aim it at a specific job. It closes on strategy: who in an organization decides where to point it, and where the line falls between work gen AI takes over and work it merely assists.1
What generative AI is
The course builds its definition in nested layers, and the exam tests whether you can place a given thing at the right layer.
- AI — computer systems performing tasks that would normally require human intelligence. It’s the outer category, defined by the outcome rather than by any particular technique.
- Machine learning — the way AI is actually achieved in practice: systems learn patterns from data instead of following rules a programmer wrote out. This matters because it explains why data quality dominates everything downstream — the system’s behavior comes from the data, not from code someone can inspect and fix line by line.
- AI model — the trained artifact itself: it takes an input and produces an output based on what it learned. A spam filter that takes an email and returns “spam” or “not spam.”
- Generative AI — the subset of AI focused on creating new content rather than analyzing or classifying content that already exists. The word “new” is load-bearing: output that didn’t previously exist, not a label attached to an existing input, and not a number predicted from a table.
Exam note: the correct definition of gen AI is always the one about creating new and original content — text, images, audio, or code — that did not previously exist. Distractors describe predictive analytics or rule-based automation, which are AI but not generative AI.
Gen AI is a technology, not an application. It gets integrated into applications. The course’s own naming is the worked example, and it’s an easy trap: the Gemini app is Google’s gen AI chatbot — a product you open. Gemini is the underlying model that powers many products, including that app, Workspace with Gemini, Gemini in Looker, and Gemini in BigQuery. Same word, two different layers of the stack.
Four primary ways to use gen AI
Google groups business use cases into four verbs. The taxonomy is worth memorizing because scenario questions expect you to sort an example into one of them.
- Create — write articles, emails, and social posts; produce images, video, and audio; generate code.
- Summarize — condense long documents, turn complex data into concise reports, extract takeaways from meetings and presentations.
- Discover — surface hidden patterns and insights in data, search across resources and documents, monitor real-time events.
- Automate — convert formats, produce documentation, trigger notifications and alarms.
Multimodal gen AI
A modality is a type of data — text, image, audio, video, sensor readings. Multimodal gen AI processes and creates content across several modalities at once, rather than being confined to one. The significance is that a single model can reason across sources that previously needed separate systems and a human to reconcile them.
The slides give four pairings, and they double as the answer bank for scenario questions:
- text + image — a marketing team generates on-brand social posts.
- text + PDF — a legal team summarizes a contract and flags risks.
- text + video + data — a market research firm analyzes sentiment across video testimonials and survey responses.
- sensor + video — a manufacturing team analyzes camera feeds alongside sensor readings to spot safety hazards.
Exam note: the canonical multimodal example is analyzing customer sentiment in video testimonials and survey data — it combines video, text, and structured data. A scenario that touches only one data type isn’t multimodal no matter how sophisticated it sounds.
Foundation models
A foundation model is a large-scale, general-purpose model trained on a vast and diverse body of data, capable of adapting to many different tasks. The contrast that defines it:
- Traditional AI model — trained for a single purpose on data specific to that purpose. A spam filter, trained on labeled emails, that does nothing but classify emails.
- Foundation model — trained on massive, diverse data (text, images, code) and adaptable to tasks nobody enumerated at training time. One model that drafts an email, explains a code snippet, and summarizes a report, having been trained for none of those specifically.
The practical payoff is that you stop building a new model per task. A foundation model develops a deep general understanding of its data, adapts to a wide range of downstream tasks, streamlines processes, and unlocks capabilities across business functions.
Three key features, which the exam asks for as a set:
- Trained on diverse data — wide variety of data means general patterns and relationships that transfer across tasks.
- Flexible — one model supports many use cases.
- Adaptable — can be specialized for a particular domain through additional targeted training.
Exam note: “specialized to specific tasks” is not a feature of foundation models — that describes traditional AI, and it appears as a distractor precisely because “adaptable” sounds adjacent to it.
Google’s examples:
- Gemini — trained on a massive multimodal dataset spanning text, images, code, audio, and video; performs tasks across domains.
- Nano Banana — trained on images plus text descriptions; generates, edits, and understands images.
- Chirp — trained on a large multilingual audio dataset; built for speech recognition.
Don’t confuse these: foundation models vs. LLMs
All LLMs are foundation models, but not all foundation models are LLMs. An LLM is the language-specialized kind. Foundation models also cover image, audio, video, and multimodal models — Nano Banana and Chirp are foundation models and neither is an LLM. Both directions of this relationship get tested, often as a select-two question.
Also worth holding onto: decision trees and linear regression are not foundation models. They’re classic ML, trained narrowly, and they show up as distractors in “which of these are foundation models” questions.
Prompting
A prompt is the input you give a model to trigger an output. Every AI model takes some input, but foundation models — multimodal ones especially — accept far more flexible inputs than traditional models did, which is what makes prompting a skill rather than a fixed API call.
The mental model the course wants: the foundation model supplies a vast knowledge base; the prompt is how you guide the model to apply that knowledge to your situation. Neither half creates value alone. This framing is the answer to “how do foundation models and prompt engineering create value together.”
Prompt-to-output examples: a question yields the answer; a text description yields an image; a request to summarize a file yields the summary; code containing errors yields corrected code.
Prompting is developed through practice and experimentation. Module 4 covers the named techniques — zero-shot, one-shot, few-shot, role prompting, prompt chaining — in Module 4.
Google’s gen AI ecosystem: five pillars
The strategic claim behind all five is one sentence: businesses can leverage Google’s AI advances without starting from scratch or managing the underlying infrastructure themselves.
- Individual productivity and efficiency — Google Search for faster information; Workspace with Gemini for drafting in Gmail, generating Slides, summarizing in Docs, automating Sheets; the Gemini app; Gemini for Google Cloud for building apps and services.
- Continuous improvement — automatic model upgrades, early access to new Gemini features, security patches and updates. You get better models without a migration project.
- Responsible AI — SAIF (Secure AI Framework), tools and best practices for building secure AI systems; Mandiant, threat intelligence for protecting AI systems; AI Principles, Google’s published guidelines for AI development; and the Responsible AI toolkit for building fair, unbiased systems.
- Enterprise ready — Agent Platform, the unified platform for building and deploying ML models with enterprise-grade security, scalability, and compliance; Google Cloud’s security infrastructure (encryption, access control, network security); and compliance certifications across industry standards.
- Open approach — contributions to TensorFlow, PyTorch, and JAX; open-source models and datasets released to researchers; support for open standards for AI interoperability.
Exam note: two things the ecosystem does not do — it does not guarantee complete accuracy, and it does not eliminate the need to build internal gen AI knowledge. Both appear as tempting “benefits.” The three benefits that are correct: access to pre-trained models like Gemini, automatic upgrades and security patches, and enterprise-grade security and compliance.
Building an AI strategy
Successful implementations start with a clear vision and a focus on genuinely impactful use cases, and they track results — tracking is what converts a pilot into a competitive advantage. The course’s central strategic point is that direction has to flow both ways.
- Top-down (executives and senior management) — create a clear, compelling AI vision aligned with strategic business priorities, identify the most important use cases and workflows, and make sure initiatives get adequate resources, org-wide support, and permission to experiment. Executive sponsorship is described as essential.
- Bottom-up (mid-level managers and individual contributors) — proximity to daily operations and end users is the asset here. People close to the work can see which problems are worth solving and which solutions are actually feasible; they champion adoption by running small experiments and sharing what worked.
Exam note: the recommended approach is always to combine both, never to pick one. A related question asks why mid-level managers and ICs matter — the answer is their proximity to workflows, which lets them identify impactful solutions leadership can’t see from above.
Six strategy factors, each viewed from both altitudes. The pattern: the executive column is about vision, governance, and resourcing at org scale; the IC column is about hands-on experimentation and demonstrating local impact.
| Factor | Executive view | Mid-level / IC view |
|---|---|---|
| Strategic focus | Focus on one high-impact area then expand; prioritize use cases that are feasible, actionable, affordable, high-ROI | Identify pressing pain points; understand user needs and workflows; explore low-risk use cases |
| Exploration | Empower employees to experiment; build a collaborative environment for sharing findings | Experiment with different gen AI tools; share findings with colleagues |
| Responsible AI | Establish ethical guidelines and safety mechanisms; institute data governance; set content moderation policy; monitor impact | Adhere to company AI standards; think proactively about pitfalls; test for bias, safety, and ethics |
| Resourcing | Create a robust data strategy; leverage existing tools and platforms; invest in AI talent | Use resources available in the team; make the case for more by showing value and ROI |
| Impact | Set clear goals and KPIs; communicate progress to stakeholders | Demonstrate tangible impact of experiments; articulate contribution to business goals; track KPIs |
| Continuous improvement | Embrace iterative development; set a regular evaluation cadence; implement feedback mechanisms | Continuously test, measure, and refine from user feedback; update solutions; keep learning |
Google agent product categories
From the creative-matrix exercise — a fictional floral business maps product categories to business goals. The categories themselves are the takeaway; the floral examples make them concrete.
- Google Workspace with Gemini — AI-powered workspace tools. Sales teams writing personalized client emails through Gemini in Gmail → increases efficiency.
- Agent Conversation — conversational AI and chatbots. An AI chatbot answering floral-care questions around the clock → enhances customer experience.
- Agent Search — AI-powered search. An image-upload tool that finds similar floral arrangements → drives innovation.
- Agent Studio — AI model development and customization. Fine-tuning a model to analyze customer data and predict demand → increases efficiency.
Augmentation vs. automation
This is the module’s most-tested distinction, and the sorting rule is simpler than the scenarios make it look: ask whether a human still exercises judgment on the output. If yes, augmentation. If the task runs end to end without one, automation.
Gen AI augments — it assists without replacing — in four areas: critical thinking and problem solving, creativity and innovation, relationship building and collaboration, and strategic planning and vision. In each, gen AI supplies data, insights, ideas, and forecasts, but humans interpret them, decide, build trust, and set long-term direction.
Gen AI automates two kinds of task:
- Repetitive and rule-based — data entry, information retrieval, content formatting, basic code generation.
- Time-consuming and resource-intensive — research, data analysis, content summarization, first-draft creation.
Exam note: gen AI augments critical thinking by providing data and insights that humans then interpret to make informed decisions — not by deciding autonomously. The distractor is always some version of the AI making the call itself.
Practice scenarios, with the reasoning:
| Scenario | Verdict | Why |
|---|---|---|
| Chatbot handles basic pricing and availability questions | Automation | Repetitive task runs unattended; frees the sales team for complex interactions |
| Financial analyst uses gen AI to analyze data and generate reports, then advises clients | Augmentation | The analyst still interprets and advises — AI enhances the role |
| Online retailer auto-generates product descriptions from attributes | Automation | Labor-intensive task executed at scale without per-item review |
| Dev team generates standard code with gen AI, then reviews and integrates it | Augmentation | Human oversight remains essential to the workflow |
Human-in-the-loop
Human involvement isn’t confined to reviewing output — it runs through the whole pipeline, in four stages:
- Data selection and preparation — ensure training data is high-quality and representative.
- Prompt design and refinement — craft prompts that produce accurate, useful responses.
- Output evaluation and refinement — review and edit generated content for accuracy, relevance, and brand alignment.
- Continuous monitoring and feedback — report on performance and identify areas to improve.