Task 4.1: Describe the Google Cloud-recommended steps to successfully implement a transformational gen AI solution.

• ABuy gen AI licenses for every employee and see what they do with them

  Incorrect

  Broad access can encourage exploration, but without a defined problem and measures there is nothing to judge the investment against.

• BTrain a custom foundation model on the company's shipment data

  Incorrect

  Training a model is a solution looking for a problem here, and the most expensive one available.

• CDefine the business problem and break the goals into measurable KPIs

  Correct

  Google Cloud's guidance is to start any AI project with a clear understanding of the business problem and required outcomes, broken down into measurable key performance indicators.

• DChoose the most capable model and design projects around it

  Incorrect

  Model choice follows from the use case and its constraints, not the other way around.

Strategic focus comes before technology: a defined problem with KPIs lets the team prioritize use cases and later prove value.

• ARecord the current average handling time to compare against

  Correct

  Impact can only be shown against a starting point. Measuring today's average handling time gives the business KPI a baseline that the pilot results can be compared with.

• BCount how many suggestions Agent Assist makes during the pilot

  Incorrect

  Suggestion counts show usage, not whether calls got shorter or customers were better served.

• CCollect public benchmark scores for the underlying model

  Incorrect

  Benchmarks describe a model in general, not this contact center's outcome.

• DTrack how many prompt versions the team tests during setup

  Incorrect

  Prompt iterations are an effort measure; they say nothing about business impact.

Measure business KPIs, not activity, and capture the baseline before the change so the difference can be attributed to the pilot.

• AGemini Enterprise on the technicians' phones

  Incorrect

  Gemini Enterprise is a cloud service; it needs a connection, which the technicians often lack.

• BBatch inference on Agent Platform overnight

  Incorrect

  Batch jobs run in the cloud with delayed results; technicians need help on site, with no network.

• CGemini Nano on the device, through the ML Kit GenAI APIs

  Correct

  The ML Kit GenAI APIs run Gemini Nano entirely on the device: data is processed locally and features keep working without a reliable internet connection.

• DA Gemini model called through Agent Platform from the app

  Incorrect

  Calling a cloud model needs connectivity and sends the notes off the device, breaking both constraints.

On-device AI keeps working without a network and keeps the notes on the phone, which no cloud service can do for these technicians.

• AUse an AI assistant to cut invoice processing time within nine months

  Incorrect

  It names the process and a deadline but no target, so nobody can measure whether it was achieved.

• BUse an AI assistant to cut invoice processing time by 20% over time

  Incorrect

  It has a measurable target but no deadline, so it is not time-bound and cannot be reviewed on a date.

• CRoll out an AI assistant to 20% of finance staff within nine months

  Incorrect

  It is measurable and time-bound, but it tracks rollout activity rather than a business outcome such as efficiency or cost.

• DUse an AI assistant to cut invoice processing time by 20% in nine months

  Correct

  Google Cloud recommends SMART objectives for AI projects; this one names the process, a measurable target tied to efficiency, a deadline and the means.

SMART objectives turn enthusiasm into something the business can measure, fund and review. A missing target, a missing deadline or an activity in place of an outcome each breaks the pattern.

• ABuild a custom multi-agent system with ADK first

  Incorrect

  A code-first build needs developers and time that the insurer does not have.

• BStart with Gemini Enterprise and expand later

  Correct

  When people, money and time are limited and the need is common knowledge work, a prebuilt offering such as Gemini Enterprise delivers search across connected sources, assistance and agents without a build project.

• CBuild a RAG app with RAG Engine on Agent Platform

  Incorrect

  RAG Engine is a framework for developing RAG applications; it would serve search, but it needs developers that the insurer does not have.

• DTurn on Gemini in Gmail and Docs and stop there

  Incorrect

  Gemini in Gmail and Docs helps with drafting inside those apps, but it does not give staff one search across the company's document systems.

Resources shape the solution: limited skills and time favor prebuilt offerings that cover the whole need, while specific needs and engineering capacity favor building.

• AAdoption, user satisfaction and content quality

  Correct

  Google Cloud's guidance lists adoption, satisfaction and content quality as the project-specific metrics for generative AI.

• BAccuracy and precision of the model's predictions

  Incorrect

  Accuracy and precision are the suggested project metrics for predictive AI, which classifies or forecasts rather than generates drafts.

• CRequest rate, error rate and latency of the endpoint

  Incorrect

  These availability metrics show whether the service is healthy, not whether the drafts are used, liked or any good.

• DRevenue growth and cost savings across the company

  Incorrect

  These are business value metrics for the program as a whole; on their own they do not show whether this particular tool is working.

Pick metrics that fit the kind of AI: generative AI is judged by use, satisfaction and output quality, alongside cost, operational health and business value.

• AThe analytics dashboard in Gemini Enterprise

  Correct

  Gemini Enterprise analytics gives insight into usage trends, search quality and end-user engagement in an interactive dashboard powered by Looker.

• BCloud Billing reports for the Gemini Enterprise project

  Incorrect

  Billing reports show what the service costs, not how employees use it.

• CModel Monitoring drift alerts

  Incorrect

  Model Monitoring tracks drift in models you deploy; it does not report on employee adoption.

• DThe Gen AI evaluation service

  Incorrect

  The evaluation service scores model outputs against metrics; it does not track usage across the company.

Measuring impact includes measuring use; Gemini Enterprise has adoption analytics built in.

• ASet one budget alert for all of the company's AI spending

  Incorrect

  A budget alert warns when total spend passes a limit, but it neither splits costs by project nor measures what each project delivered.

• BEstimate, track and assign owners for each project's costs and benefits

  Correct

  Google Cloud recommends FinOps principles: each AI project has estimated costs and ways to measure actual costs over its lifecycle, and its costs and benefits have assigned owners with clear accountability.

• CRun all projects on Provisioned Throughput so that the bill stays fixed

  Incorrect

  A fixed capacity commitment makes some spend predictable but still does not attribute costs or benefits to projects.

• DReview the company's total cloud bill with finance every quarter

  Incorrect

  A total bill shows what the company spent overall, not what each project cost or what value it returned.

Value has to be measured to be managed: estimate, track and assign ownership for both cost and benefit, project by project. Totals and limits alone cannot answer either question.