This domain is about 15% of the exam. It is the leadership part: how to pick a use case, roll it out, measure it, keep it secure and use it responsibly. Questions are scenario-based and the right answer is usually the one that is measured, staged, human-supervised and aligned to a business goal, not the flashiest technology.
Types of gen AI solutions
| Type | Examples |
|---|---|
| Text generation | Drafting, summarising, chat assistants, translation, support replies |
| Image generation | Marketing visuals, product concepts, design variations |
| Code generation | Code completion, explanation, test writing, migration |
| Video and audio generation | Ads, training clips, voice-overs |
| Personalized experiences | Recommendations and messages tailored to each user |
| Search and discovery | Q&A across company knowledge |
| Agents | Systems that complete multi-step work with tools |
Factors that shape a gen AI need
| Business requirements | Technical constraints |
|---|---|
| The problem and goal, and who the users are | Available, clean and accessible data |
| Expected value and budget | Integration with existing systems |
| Time to market | Latency, scale and reliability needs |
| Risk tolerance, regulation and compliance | Security, privacy and data residency |
| Quality bar and acceptable error rate | Skills the team has |
| Brand and customer experience | Model capability, context window and cost |
Choosing the right solution
Work from the problem, not the technology.
- Define the business problem and success measure.
- Decide build or buy. Use a prebuilt application where one fits (Gemini for Workspace, Gemini Enterprise, Customer Engagement Suite). Use a platform (Agent Platform) to customise. Build from scratch only for unique needs.
- Check data: is it available, high quality and allowed to be used?
- Pick the model (modality, size, cost, latency, security).
- Choose the improvement technique: prompting, grounding or RAG, fine-tuning.
- Decide where humans stay in the loop.
| Situation | Direction |
|---|---|
| Common need, no AI team, want results fast | Prebuilt application or agent |
| Need company-specific answers | Grounding or RAG on company data |
| Unique workflow or product feature | Build on Agent Platform |
| Strict control over model and hosting | Open model such as Gemma |
| Single narrow capability (OCR, translation) | Pre-trained API |
Steps to integrate gen AI into an organization
- Align with strategy and identify use cases, ranked by business value and feasibility. Start with high-value, lower-risk ones.
- Prepare data and set governance and access rules.
- Pilot a focused use case with clear metrics. Prove value on a small scale.
- Build or buy and integrate with existing systems and workflows.
- Establish responsible AI and security practices early: policies, review, guardrails.
- Train people and manage change. Explain what AI does and doesn’t do, and involve staff.
- Measure and iterate: evaluate against KPIs, gather feedback, improve.
- Scale what works across teams, with continuous monitoring.
Success factors: executive sponsorship, a cross-functional team (business, IT, security, legal), clear ownership, and treating AI as a change in how people work, not just a tool install. Common failure causes: no clear problem, poor data, no measurement, ignoring users, and skipping governance.
Measuring impact
| Category | Examples of measures |
|---|---|
| Financial | Return on investment (ROI), cost savings, revenue growth, cost per task |
| Productivity and efficiency | Time saved, tasks completed per person, faster cycle time |
| Quality | Accuracy, error and rework rates, consistency |
| Customer experience | Satisfaction (CSAT), Net Promoter Score, resolution time, first-contact resolution |
| Employee experience | Adoption, satisfaction, time on higher-value work |
| Risk and safety | Incidents, policy violations, human override rates |
| Technical | Latency, uptime, token cost, model drift |
Methods: set a baseline before launch, run pilots and A/B tests against the old process, track KPIs continuously, collect qualitative feedback. ROI compares benefits to total cost (licences, usage, integration, training, oversight). Include soft benefits but don’t rely on them alone.
Secure AI
Secure AI means protecting AI systems, their data and their users from attack and misuse, across the whole lifecycle. AI adds new risks beyond normal software.
| Threat | Meaning |
|---|---|
| Prompt injection | Malicious text in a prompt or retrieved content that tricks the model into ignoring its instructions |
| Data poisoning | Tampering with training or grounding data to corrupt behaviour |
| Model theft or extraction | Stealing or copying a model or its behaviour |
| Sensitive data leakage | Model reveals private training or prompt data |
| Adversarial inputs | Crafted inputs that cause wrong outputs |
| Misuse | Using the system to produce harmful content or fraud |
| Excessive agent permissions | An agent allowed to do more than it should |
Security across the ML lifecycle
- Data: control access, encrypt, check provenance and quality, protect against poisoning.
- Training and tuning: secure environments, trusted code and datasets.
- Deployment: authenticate callers, limit permissions, encrypt traffic.
- Operation: monitor inputs and outputs, log activity, detect abuse and drift, respond to incidents.
- Retirement: delete data and models properly.
Secure AI Framework (SAIF)
SAIF is Google’s conceptual framework for securing AI systems. Its purpose is a shared, practical set of principles so organizations build and use AI securely. It has six core elements:
- Expand strong security foundations to the AI ecosystem, reusing existing secure-by-default infrastructure and expertise.
- Extend detection and response to bring AI into an organization’s threat universe: monitor inputs and outputs of AI systems.
- Automate defenses to keep pace with new and existing threats, using AI to defend at scale.
- Harmonize platform-level controls for consistent security across the organization.
- Adapt controls for faster feedback loops: test, red-team and update as threats change.
- Contextualize AI system risks in surrounding business processes, including end-to-end assessments.
Benefits: a common language, lower risk, consistent controls and faster response.
Google Cloud security tools
| Tool | Purpose |
|---|---|
| Secure-by-design infrastructure | Google’s hardened, layered infrastructure: encryption at rest and in transit by default, custom hardware security, global network protection |
| Identity and Access Management (IAM) | Controls who can do what on which resources. Use least privilege and service accounts for agents |
| Security Command Center | Central view of security posture: finds vulnerabilities, misconfigurations and threats across Google Cloud, including AI workloads |
| Workload monitoring tools | Cloud Logging, Cloud Monitoring and audit logs to observe activity and detect anomalies |
| VPC Service Controls | Security perimeters that stop data leaving defined boundaries |
| Sensitive data protection | Find and de-identify sensitive data |
| Model Armor and safety filters | Screen prompts and responses for injection and harmful content |
Responsible AI
Responsible AI is building and using AI in ways that are fair, safe, transparent and accountable, and that respect people’s rights. It protects customers and brand, meets regulation, builds trust and reduces harm. It is a business necessity, not an add-on. Google’s AI principles guide its own work: be socially beneficial, avoid unfair bias, be built and tested for safety, be accountable, incorporate privacy design, uphold scientific excellence.
Transparency
Tell people when they are dealing with AI and what it can do. Explain what data is used and how, label AI-generated content, and document the model’s purpose and limits.
Privacy
| Concept | Meaning |
|---|---|
| Privacy risks | Exposing personal or confidential data in prompts, training data or outputs. Misuse beyond consent |
| Data anonymization | Removing or altering identifiers so people cannot be identified (irreversible in principle) |
| Pseudonymization | Replacing identifiers with tokens or aliases, so data can be re-linked using a separate key. Lower risk but still personal data under many laws |
| Other controls | Data minimisation (collect only what you need), consent, access control, encryption, retention limits, regulation compliance (such as GDPR) |
Trap: pseudonymized data is not fully anonymous.
Data quality, bias and fairness
- Bias enters through unrepresentative or skewed data, labelling choices and design decisions, and it can be amplified by models.
- Fairness means similar people get similar outcomes. Test results across groups.
- Mitigate with diverse, representative data, bias testing and evaluation, review by diverse teams, ongoing monitoring and human oversight.
- Poor data quality (gaps, errors, outdated data) produces unreliable and unfair output.
Accountability and explainability
- Accountability: named people and processes are responsible for AI outcomes. The business owns results. “The model did it” is not an excuse. Includes governance, review boards, audit trails and human oversight.
- Explainability: being able to describe why a system produced an output, in terms users and regulators can understand. Important in high-stakes areas such as lending and healthcare. Generative models are hard to explain fully, so use citations, grounding sources, documentation and logging to help.
Common traps
- Responsible AI is not only a legal matter. It includes fairness, safety and trust.
- Security (protecting systems from attack) and responsible AI (ethical, fair use) overlap but are different topics.
- Pseudonymization is not anonymization.
- A big-bang company-wide launch is usually the wrong answer. Pilot, measure, then scale.
- Measuring only technical accuracy misses business value. Tie metrics to outcomes.
- Humans remain accountable. Automation does not transfer responsibility.
Exam checklist
- Name the types of gen AI solutions and the factors that shape a choice.
- Walk through choosing a solution: problem, build or buy, data, model, technique, humans.
- List the steps to integrate gen AI: use cases, data, pilot, integrate, governance, training, measure, scale.
- Give example measures: ROI, productivity, quality, CSAT, risk.
- Define secure AI and name threats such as prompt injection and data poisoning.
- State the purpose of SAIF and its six elements.
- Match IAM, Security Command Center, secure-by-design infrastructure and monitoring to their purposes.
- Explain transparency, privacy (anonymization vs pseudonymization), bias, fairness, accountability and explainability.