04 - Business strategies for a successful generative AI solution¶
Weighted more heavily than a technical reader expects. This is the section that distinguishes the exam from every other AI certification in this repo.
Finding use cases worth doing¶
Start from the business problem, not the technology.
Good candidates share features: - A repetitive task involving language, images, or unstructured content - A measurable outcome: time saved, cost avoided, revenue enabled, quality improved - Tolerance for imperfection, or a workable human review step - Available data, if grounding or customization is needed
Poor candidates: - Anything requiring guaranteed correctness with no review, such as an unchecked regulatory filing - Problems a deterministic system solves better: arithmetic, rules-based routing, exact lookup - Cases with no measurable outcome, which cannot be justified or evaluated - Cases where the data does not exist or cannot be used
Not everything needs generative AI. The exam includes scenarios where a rules engine, a search index, or a classical ML model is the right answer, and recognizing them is part of what is being tested.
Prioritizing¶
The standard framing is value against feasibility:
| Low feasibility | High feasibility | |
|---|---|---|
| High value | Strategic bets: invest deliberately, expect a longer timeline | Start here. Quick wins that build credibility |
| Low value | Avoid | Only if genuinely trivial |
Early projects should be chosen partly for their demonstration value: a visible, quick success builds the organizational appetite that later, harder projects need.
The business case¶
What a credible case contains:
- The problem, and what it costs today
- The proposed approach, and which rung of the ladder it sits on
- Expected benefit, quantified where possible
- Total cost of ownership, not just inference
- Success criteria, defined before building
- Risks and how they are mitigated
- Who owns it after launch
Total cost of ownership components people forget: - Data preparation, cleaning, and access work, usually the largest hidden cost - Integration with existing systems - Evaluation, both initial and ongoing - Monitoring and incident response - Model version changes and re-testing - Change management and training - Human review capacity, where it is required
Inference cost is usually visible and often not the largest line.
Data readiness¶
The most common real constraint, and the exam reflects that.
Questions worth asking before committing: - Does the data exist, and is it accessible? - Is it accurate, current, and consistent enough? - Are we permitted to use it, contractually and legally? - Is it governed: classified, access-controlled, with an owner? - If we ground on it, will retrieval respect existing permissions?
An organization with poor data governance will find that generative AI surfaces the problem rather than solving it. An assistant honors existing permissions, so pre-existing oversharing becomes visible exposure on day one.
Change management and adoption¶
Deploying a tool is not adoption.
- Communicate the why, including honestly addressing job-impact concerns
- Train users on effective use, including its limits
- Redesign the workflow; bolting AI onto an unchanged process rarely delivers the benefit
- Identify champions within teams
- Collect feedback and iterate visibly
- Measure adoption, not just deployment
Organizational readiness includes skills, governance, and executive sponsorship. A technically successful pilot with no adoption plan is the common failure mode.
Responsible AI¶
Google's themes, each with a practical control:
| Theme | Means | Control |
|---|---|---|
| Fairness | Avoid creating or reinforcing unfair bias | Test outputs across affected groups; diverse evaluation data |
| Transparency | People know AI is being used and what it does | Disclosure in the interface; documentation |
| Explainability | Decisions can be explained to those affected | Citations, grounding, documented model behavior |
| Privacy | Personal data is protected throughout | Minimization, redaction, access control, retention limits |
| Safety and security | The system behaves acceptably and resists attack | Safety filters, adversarial testing, secure design |
| Accountability | A human is responsible for outcomes | Named owner, human in the loop, audit trail |
Reference frameworks the exam guide names: Google's AI Principles and the Secure AI Framework (SAIF).
Risk¶
| Risk | Shape | Mitigation |
|---|---|---|
| Hallucination in customer-facing use | A confident wrong answer given to a customer | Grounding, citations, scope limits, human review for high-stakes replies |
| Data leakage | Sensitive data reaching the wrong user or leaving the organization | Retrieval authorization on the end user's identity, DLP, enterprise data controls |
| Intellectual property | Uncertainty over generated content and training data provenance | Provider indemnity terms, review process, legal input |
| Regulatory exposure | Sector rules and emerging AI regulation | Classify systems by risk, maintain documentation, follow EU AI Act and similar regimes |
| Over-reliance | Users accepting output without scrutiny | Training, interface design that signals uncertainty, human review |
| Cost overrun | Usage growing faster than expected | Budgets, quotas, monitoring, model tiering |
| Prompt injection and misuse | Attacker-supplied text changing system behavior | Bounded permissions, output validation, see AI security |
Human in the loop¶
Where to keep a person in the decision: - High-stakes decisions: credit, employment, healthcare, legal - Anything irreversible or outward-facing - Regulated decisions requiring explanation or appeal - Any case where the cost of a rare wrong answer exceeds the aggregate benefit of automation
Where full automation is reasonable: low-stakes, high-volume, reversible tasks with good measurement in place.
Recognizing which side of that line a scenario falls on is a repeated exam pattern.
Measuring impact¶
Define before building, evaluate after: - Business metrics: time saved, cost per case, conversion, resolution rate, customer satisfaction - Quality metrics: accuracy, faithfulness, refusal rate - Operational metrics: latency, availability, cost per request - Adoption metrics: active users, proportion of eligible work handled
Then iterate. Generative AI systems degrade quietly as data, models, and usage change, so evaluation is an ongoing program rather than a launch gate.
Key terms¶
- Use case prioritization - selecting projects by business value against feasibility
- Total cost of ownership - the full lifetime cost including data preparation, integration, evaluation, and change management
- Data readiness - whether data exists, is accessible, accurate, permitted, and governed enough to use
- Change management - the organizational work of training, communication, and workflow redesign that drives adoption
- Organizational readiness - the skills, governance, and sponsorship needed for a project to succeed
- Responsible AI - the practice of building AI that is fair, transparent, explainable, private, safe, and accountable
- Google AI Principles - Google's published commitments governing its AI development and use
- Secure AI Framework (SAIF) - Google's conceptual framework for securing AI systems
- Human in the loop - requiring human review or approval before a consequential AI output is acted on
- Over-reliance - the risk that users accept AI output without appropriate scrutiny
- Explainability - the ability to explain an AI-influenced decision to the person it affects
- Adoption metric - a measure of how much of the eligible work actually flows through the new capability