10.1 Gen AI Solution Types and Choosing the Right One
Key Takeaways
- The exam guide asks leaders to recognize gen AI solution types, identify the factors influencing need, and choose the right solution for a specific business requirement.
- The build-versus-buy ladder runs from turnkey SaaS through configured platform services to fully custom development, and cost, time, and differentiation rise together at each step.
- Business requirements and technical constraints are the two factor families the exam names: accuracy tolerance, latency, volume, data sensitivity, regulatory exposure, and available skills.
- Buying is correct when the capability is not a source of competitive differentiation, which is true of most enterprise use cases.
- The most common leadership error is selecting a solution type before defining the acceptable failure mode for the use case.
10.1 Gen AI Solution Types and Choosing the Right One
Executive Summary: The exam guide's business-strategy section opens by asking leaders to "recognize the different types of gen AI solutions," "identify the key factors that influence gen AI needs (e.g., business requirements, technical constraints)," and "describe how to choose the right gen AI solution for a specific business need." This is the exam's build-versus-buy question, and it has a disciplined answer: choose the least complex solution that meets the requirement, where complexity is measured in time, cost, and the specialist skills the organization must sustain.
The Solution Types
| Type | What it is | Google Cloud examples | Time to value | Differentiation it creates |
|---|---|---|---|---|
| Turnkey SaaS | Licensed product, configured not built | Gemini for Google Workspace, Gemini Enterprise app, Gemini Code Assist | Days | None - competitors can license it too |
| Configured platform service | Managed service pointed at your data | Prebuilt RAG with Agent Search, Customer Engagement Suite, pre-trained APIs | Weeks | Moderate - your data is the differentiator |
| Custom-built on the platform | Your logic, orchestration, and agents on managed components | Custom agents on Agent Platform, RAG APIs, AutoML, tuned models | Months | High - encodes proprietary process |
| Fully custom / self-hosted | You own the model and serving stack | Self-hosted Gemma on GKE, custom training | Quarters | Highest, and highest ongoing burden |
The default should be the highest row that meets the requirement. Enterprises consistently over-select downward, building custom systems for capabilities that a licensed product delivers on Monday.
The Factors That Actually Decide It
Business requirements
- Acceptable failure mode. This is the most important and most frequently skipped question. What happens when the system is wrong? A wrong marketing draft is edited. A wrong dosage recommendation harms a patient. The tolerance determines whether human review is mandatory, and that in turn constrains the whole design.
- Accuracy threshold. "Better than the current process" and "provably correct" are wildly different bars with wildly different costs.
- Time to value. A regulatory deadline in eight weeks eliminates every custom option regardless of its elegance.
- Differentiation. If the capability is not something customers would choose you for, buy it.
- Volume and economics. Ten queries a day and ten million a day are different architectures; at high volume, unit cost dominates.
Technical constraints
- Data sensitivity and residency. Regulated or sovereign data may force particular regions, encryption arrangements, or self-hosting.
- Latency. A real-time voice agent has a budget in hundreds of milliseconds; an overnight batch summarization does not.
- Integration surface. Systems the solution must read from and write to, and whether connectors already exist.
- Available skills. A solution the organization cannot staff for three years is not a solution.
- Existing estate. Data already in BigQuery, identity already in Google Cloud IAM, and documents already in Workspace all shorten the path.
The Selection Method
- Classify the use case by pattern - create, summarize, discover, or automate. This narrows the candidate services immediately.
- Define the acceptable failure mode and therefore whether human review is mandatory.
- Apply the hard filters - residency, latency, volume, regulatory. These eliminate options rather than rank them.
- Ask whether the capability differentiates. If not, buy.
- Select the least complex surviving option.
- Define the success metric before building, so the decision to continue, fix, or retire is evidence-based.
Worked Portfolio: One Company, Four Different Answers
A mid-size insurer evaluates four requests. Applying the method yields four different solution types - which is the point.
| Request | Pattern | Decisive factor | Solution type chosen |
|---|---|---|---|
| Help staff draft internal emails and documents faster | Create | Not differentiating; needed immediately | Turnkey SaaS - Gemini for Google Workspace |
| Let staff find answers across policy documents in Drive, SharePoint, and the claims system | Discover | Cross-system grounding with per-user permissions | Configured platform service - Gemini Enterprise app with connectors |
| Automate first-notice-of-loss intake, extracting fields and triaging severity | Automate | Encodes proprietary underwriting rules; writes to core systems | Custom-built on the platform - custom agent plus Document AI, with human-in-the-loop above a threshold |
| Run claims triage inside a jurisdiction that forbids external model calls | Automate | Hard regulatory constraint on data leaving the boundary | Self-hosted - Gemma on GKE inside the permitted boundary |
The fourth is the most expensive and most burdensome option, and it is correct only because a hard regulatory filter eliminated everything above it. That is the discipline the exam is testing: complexity must be forced by a constraint, never chosen for its own sake.
The Errors Leaders Make
- Choosing the solution before defining the failure mode. Every downstream decision - review workflow, accuracy bar, escalation threshold - depends on it.
- Building for differentiation that does not exist. Very few enterprises differentiate on meeting summarization.
- Ignoring the sustaining cost. A custom system is a permanent staffing commitment for evaluation, re-embedding, model migration, and patching.
- Letting a proof of concept dictate architecture. A demo that impressed the steering committee is not evidence that the approach survives volume, permissions, and audit.
Strategic Leadership Guidance: Exam Tips and Common Pitfalls
[!TIP] Exam Tip: Scenario questions usually plant one hard constraint - a residency rule, a latency budget, a regulatory prohibition. Find it first: it eliminates options far faster than comparing features.
Pitfall 1: Recommending custom development when a licensed product meets the need. Almost always the wrong answer unless the scenario names a limitation the product cannot meet.
Pitfall 2: Treating self-hosting as the safe default for sensitive data. Managed Google Cloud services already offer residency, CMEK, VPC Service Controls, and commitments that enterprise data is not used to train foundation models. Self-hosting is justified by a specific constraint those controls cannot satisfy.
Pitfall 3: Omitting the success metric. Without one defined up front, a project cannot be honestly evaluated and will be defended on anecdote.
An insurer must run claims triage inside a jurisdiction whose regulator forbids the data from being processed by any externally hosted model. Which solution type is justified, and why?
A department requests a nine-month custom build so staff can summarize meetings and draft documents faster. What should a Generative AI Leader recommend first?
Which question does the exam guide's selection discipline treat as the one that must be answered before a solution type is chosen?