Plan, integrate and measure a gen AI solution

Study guide for Generative AI Leader, Unit 4 · Topic 1. This is the topic's lecture in reading form — every slide's teaching, figures and worked examples, in order — followed by the official Google Cloud pages its claims rest on.

What the exam guide asks, quoted. Recognizing the different types of gen AI solutions (e.g., text generation, image generation, code generation, personalized user needs). Identifying the key factors that influence gen AI needs (e.g., business requirements, technical constraints). Describing how to choose the right gen AI solution for a specific business need. Identifying the steps to integrate gen AI into an organization. Identifying techniques to measure the impact of gen AI initiatives.

This hive's learning objectives for the topic:

1. Recognize text, image, code and personalization solution types and the business need each serves.
2. Weigh business requirements and technical constraints when choosing a gen AI solution.
3. Sequence the steps to integrate gen AI into an organization.
4. Choose metrics that measure the impact of a gen AI initiative.

From use case to measured impact

Name the outcome, choose the solution, change the work, prove the value

What kind of solution? text, images, code or personalization. Which approach? generative, traditional, pre-trained — or no AI. How does it land? people, process and a safe release. Did it work? measures agreed before the build.

This unit moves from technology to business strategy, and this first topic follows one initiative from idea to evidence. Google's guidance starts with a warning that sets the tone: generative and traditional artificial intelligence solutions are powerful tools, but they should always support your business goals, and they shouldn't exist in isolation. So the topic runs in four steps. First, recognise what kind of generative solution a need calls for — text, images, code or personalization. Second, weigh the requirements and constraints that decide whether generative AI is the right approach at all, or whether a traditional model, a pre-trained service, or no AI would serve better. Third, sequence the work of integrating the solution into the organization — the users, the processes that must change, the people who must be involved, and a release that can be rolled back. Fourth, measure the impact, against measures the business agreed before anything was built. Leaders are assessed on that discipline far more than on any model name.

Kinds of gen AI solution, and the need each serves

Match the output the business needs to the kind of solution

• Text: summaries and drafts that speed up reading and writing
• Images and video: new or edited visual content from text prompts
• Code: writing, explaining and documenting software
• Personalization: recommendations shaped by each user's behaviour
• Assistants: conversational help for support and online sales

Worked example (synthetic). A retailer has three asks: product photos in new colours, weekly digests of customer reviews, and suggestions tailored to each shopper. That is an image solution, a text solution and a personalization solution — three kinds, not one.

Start with the distinction Google draws between two families. Traditional artificial intelligence models excel at learning from existing data to classify information or predict future outcomes based on historical patterns. Generative models expand those capabilities to create summaries, uncover hidden correlations, or generate new content — text, images or videos — that reflect the patterns in their training data. Within generative solutions, the guide names text, image, code and personalized-experience solutions, and Google's own list of use cases maps onto them. Text: summarizing documents, articles, customer feedback and reports to help with more informed, data-driven decisions. Images and video: creating new images from text descriptions, modifying images with text prompts, and generating short videos from scripts. Code: generating code and assisting developers in writing, explaining and documenting it. Personalization: analyzing user behaviour, preferences, reviews and past interactions to provide personalized content recommendations. And conversational assistants for customer support and online sales draw on all of these. The skill being tested is matching the business need — the output somebody actually wants — to the kind of solution that produces it.

Solution type, business need, and what it produces

Name the need first; the solution type follows from the output

Worked example (synthetic). A bank's compliance team drowns in policy updates. The need is reading time, so the solution type is text — summaries — not image generation, however impressive a demo of it looked.

As a table, the mapping reads from the middle column outward. A team overwhelmed by reading or writing needs a text solution — Google lists summarizing documents, articles, customer feedback and reports. A team that needs visual content at volume needs an image or video solution — creating images from text descriptions, editing them with text prompts, or generating short videos. A development team that needs to move faster needs a code solution, which Google describes as generating code and helping developers write, explain and document it. And a business that needs each customer to see something relevant needs personalization — recommendations based on behaviour, preferences, reviews and past interactions. The order matters: start from the need. A scenario that begins with a striking demo of one solution type and a business problem of another kind is testing whether you will be distracted.

Requirements and constraints decide the approach

Gen AI is one option; the requirement decides, not the trend

• Start from a measurable business goal, then ask whether AI is the right approach
• Constraints: security, privacy, industry compliance, country regulation
• Predictive work on structured data usually suits traditional AI
• No training data, or speed to market first: generative AI with prompts
• A pre-trained AI API that meets the need wins over building

Worked example (synthetic). An insurer wants to forecast claim volumes from ten years of tables. The work is predictive and the data is structured, so traditional AI fits; a generative model adds nothing the forecast needs.

The second objective is weighing what the business needs against what it is constrained by. Google's process begins with the goal and then asks a blunt question: determine whether artificial intelligence or machine learning is the right approach at all, and decide whether the business expectation requires generative AI, other types of AI, or no AI. Constraints come early too — Google says to identify business constraints such as security and privacy aspects that must meet specific industry compliances or country regulatory requirements. Then the technical factors. If the use case is predictive, use traditional AI; Google notes traditional predictive AI is particularly effective with structured data. If a pre-trained model meets your requirements — services such as Document AI, Vision AI, the Natural Language API and the Video Intelligence API — use the pre-trained model. If there is no training dataset, or not enough to custom train, use generative models with prompt engineering, and if the priority is a faster go-to-market, use generative AI. Google also lists the factors a deeper analysis weighs: the data, anticipated outcomes, expected serving latency, model metrics and more.

A decision path for the approach

The use case and the data you have choose the branch

Figure: A top-down decision tree. The use case branches three ways. Predicting from structured history leads to traditional AI. Classifying or detecting asks whether a pre-trained API meets the need; if yes, use it; if no, ask whether there is enough training data. Without enough data, use gen AI with prompt engineering. With enough data, choose by priority: control leads to custom-training a model, speed leads to gen AI with prompts. Summarizing, generating, conversing or multimodal work leads to generative AI.

Worked example (synthetic). A logistics firm wants to read delivery-note photos. A pre-trained document service already extracts the fields it needs, so it uses that and skips both custom training and a generative build.

Google publishes a simplified decision tree for this choice, and the branches are worth knowing by heart. If the use case is predictive, use traditional AI. If it is classification or detection, first check whether a pre-trained model meets the requirement; if it does, use it. If it does not, check whether enough training data exists. With no dataset, or too little to custom train, use generative models with prompt engineering. With enough data, the question becomes priority: high control over training points to a custom-trained model, while a faster go-to-market points to generative AI. And if the use case is itself generative — summarization, content generation, advanced transcription — or needs input from several modalities like text, images, video or audio, use generative AI. Google adds that the tree is simplified: an in-depth analysis also weighs the data, anticipated outcomes, serving latency and model metrics.

Not either-or: traditional and generative together

One model finds the segments; the other writes for each

Figure. Two cards joined by an arrow. Left: traditional AI learns from history to classify and predict which customers form which segment. Right: generative AI creates marketing copy tailored to each segment. A band beneath says each does the part it is good at, and the business need needs both.

Worked example (synthetic). A telecom predicts which customers are likely to leave with a traditional model, then lets account managers question those predictions in plain language through a generative assistant.

A trap in this objective is treating the choice as exclusive. Google says plainly that traditional AI and generative AI aren't mutually exclusive, and gives combinations. Traditional predictive AI can identify customer segments; you can then use generative AI to generate personalized marketing content tailored to each identified segment. The division of labour follows the two families' strengths: traditional models excel at learning from existing data to classify information or predict outcomes from historical patterns, while generative models create new content. So when a scenario needs both a prediction and a piece of tailored content, the best answer often uses both, each for the part it does well.

Integrating gen AI into the organization

A model is not adopted until the work around it changes

• Define the business problem and the required outcome before any build
• Identify the end users and how they will interact with the solution
• Change the processes and back-end workflows the solution touches
• Involve leaders, product owners, domain experts and end users
• Release safely: detect issues early and roll back quickly

Worked example (synthetic). A support team adds a chatbot but keeps every escalation rule as before. Agents receive the same tickets twice, and nothing improves until the escalation workflow is redesigned around the bot.

The third objective is the sequence of integrating a solution into an organization, and Google's guidance is consistent: the technology is the middle of the story. First, before you start any artificial intelligence or machine learning project, you must have a clear understanding of the business problem to be solved and the required outcomes. Next, identify the end users and how they will interact with the application or service. Then the step most often skipped: identify the changes the business must make to existing processes or workflows, because these interactions might require back-end processes to be changed or reinvented to realize the benefits — workflow automation included. People matter throughout. Google notes that although data scientists and machine learning engineers commonly lead model selection, the input of business leaders, product owners, domain experts and end users matters too. And at release, use an approach that lets you detect issues early, validate performance, and roll back quickly when required, with monitoring of the deployed service's health and performance afterwards.

The integration sequence

Each step depends on the one before it

Figure: A left-to-right flowchart: business problem and outcome, measures of success agreed, right approach (gen AI, other AI, or none), users and how they interact, processes and workflows changed, safe release with rollback, monitor and measure impact. An arrow loops from monitoring back to the approach, labelled evidence feeds the next decision.

Worked example (synthetic). A legal team pilots contract summaries. Measures are agreed first (review hours per contract), the filing workflow is changed to route summaries to reviewers, and the release can be switched off in a day.

Drawn as a sequence, integration has an order you can defend. The business problem and the required outcome come first. Measures of success come next, because Google asks you to clarify how the business plans to measure success of the goals before the solution exists. Then the approach — generative AI, another kind of AI, or none. Then the users and how they will interact. Then the processes and workflows that must change, including back-end ones. Then a release that lets you detect issues early, validate performance and roll back quickly. And finally monitoring and measurement, whose evidence loops back into the next decision. Google's operational guidance adds the human frame around the whole loop: build trust and alignment with clearly defined roles, responsibilities and metrics for success.

Who takes part, and what each contributes

Engineers lead the selection; four other groups shape it

Worked example (synthetic). A hospital's coding assistant is chosen by engineers alone. Clinical coders, its end users, cannot tell how to check its suggestions, and adoption stalls until they are brought in.

Integration is also a question of who is in the room. Google notes that data scientists and machine learning engineers commonly lead the model selection process, but it is important to consider the input of key stakeholders, and it names what each contributes. Business leaders and decision-makers approve the selection when it is aligned with the business priorities. Product owners might require influence or more control of the model behaviour to align it with the product priorities. Domain experts apply their domain expertise to improve model effectiveness. And end users might need to understand the output of the model and how to incorporate it into more informed decisions. A rollout that leaves any of these groups out tends to fail at that group's step — approval, product fit, accuracy, or adoption.

Measuring the impact of a gen AI initiative

Agree the measure before the build, then track it after launch

• Clarify how success will be measured before building
• Return on investment is a key measure of AI project success
• Financial: revenue gained or costs reduced
• Efficiency: faster time-to-market or issue resolution
• Experience: satisfaction scores and retention

Worked example (synthetic). A team reports its chatbot answered 40,000 questions. The board asks what changed: the agreed measures were support cost per quarter and first-contact resolution, and neither was tracked.

The last objective is measuring impact, and Google frames it from the first step of the project: clarify how the business plans to measure the success of the identified goals and objectives. It names return on investment, or ROI, as one of the key measures of AI project success, and says it can be measured through several kinds of metric. Direct financial gains: increased revenue or reduced costs. Operational efficiency: faster time-to-market or quicker issue resolution. Customer experience: increased satisfaction scores or improved retention. Notice what is not on the list — counts of activity, like how many questions a bot answered. Activity shows the solution was used; the measures Google names show whether the business goal moved. For a support chatbot, Google's own example measures include the percentage decrease in support operational costs over a period, the percentage of tickets resolved without human intervention, the average decrease in time-to-resolution, and the increase in customer satisfaction scores tied to chatbot use.

Three families of ROI evidence

Each family answers a different business question

Figure. A band labelled return on investment holds three pillars: direct financial gains (revenue up, costs down), operational efficiency (faster resolution) and customer experience (satisfaction, retention). A callout reads: set the measure before the build, then track it after launch.

Worked example (synthetic). A marketing team's image generator is judged on cost per campaign asset (financial), days from brief to launch (efficiency) and click-through on personalized ads (experience).

The three families Google names for measuring return on investment, or ROI, cover different questions, which is why a strong measurement plan usually draws on more than one. Direct financial gains ask whether revenue went up or costs went down. Operational efficiency asks whether things got faster — time-to-market, or how quickly issues are resolved. Customer experience asks whether customers are happier and staying — satisfaction scores and retention. The callout is the discipline that makes any of them meaningful: Google's process asks the business to clarify how success will be measured while the goal is being defined, so the baseline exists before launch and the comparison afterwards is honest.

Goal, then the measure that shows it moved

A support chatbot's goals, each with Google's example measure

Worked example (synthetic). The figures a team reports should be these, measured against a baseline taken before launch — not a count of conversations.

Google's worked example of a support chatbot shows how a goal becomes a measure. For support efficiency: the percentage decrease in customer support operational costs over a defined period, such as quarterly. For ticket resolution: the percentage of tickets resolved without human intervention. For speed: the average decrease in time-to-resolution for the inquiries the chatbot handles. For customer experience: the increase in customer satisfaction scores in surveys tied to chatbot usage. And the increase in first-contact resolution rate — issues solved in a single interaction. Each measure ties back to a stated goal, and each needs a baseline from before the chatbot existed. An exam answer that offers a vanity number — volume of chats, number of prompts — instead of one of these is the distractor.

What this topic actually tests

Four questions a leader answers before saying yes

What output does the business need? text, image, code or personalization. Is gen AI the right tool? or traditional, pre-trained, combined — or none. What must change around it? users, process, people, a safe release. How will we know? ROI measures agreed in advance.

Close on four questions a leader answers before approving a generative initiative. What output does the business need — text, images, code, or personalized experiences? Is generative artificial intelligence the right tool, given the requirements and constraints, or would traditional AI, a pre-trained service, a combination, or no AI serve better? What must change around it — the users' interactions, the processes and back-end workflows, the people involved, and a release that can be rolled back? And how will we know it worked — return on investment, or ROI, measured as financial gains, operational efficiency or customer experience, agreed before the build. The next topic turns to keeping such systems secure.

Official sources for this topic

• Evaluate and define your generative AI business use case — Google Cloud
• When to use generative AI or traditional AI — Google Cloud
• AI and ML perspective: Operational excellence — Cloud Architecture Center