Task 2.2: Measure and demonstrate AI business value.

• AA larger set of technical metrics, so that the sponsor sees the full picture

  Incorrect

  More technical metrics deepen the engineering view but still leave the gap between technical performance and business value.

• BA framework that traces each technical metric to a measurable business goal

  Correct

  AWS recommends a structured framework such as Objectives, Goals, Strategies and Measures (OGSM) so that every technical measurement is traceable to a meaningful business outcome, creating a shared language.

• CA single overall satisfaction survey of the assistant's users

  Incorrect

  User satisfaction is one useful measure, but on its own it does not connect the engineering metrics to the business goals and budget.

• DA comparison of the assistant's benchmark scores with competitors' published scores

  Incorrect

  Benchmark comparisons describe model capability in general, not the business result this assistant delivers for the insurer.

Technical metrics disconnected from business impact are a common failure point. Linking objectives to quantified goals, the strategies that reach them and the measures that track them lets executives and engineers talk about the same results.

• AThree months after deployment, once the agents have settled into a stable pattern

  Incorrect

  After deployment there is no clean record of the old process to compare against, so savings can only be estimated, not measured.

• BBefore deployment, including hidden costs such as errors and rework

  Correct

  AWS's economics guidance says to establish an accurate baseline of current process costs, including hidden expenses, failure rates and opportunity costs, before comparing against agentic alternatives.

• CBefore deployment, counting only direct labor hours because other costs are hard to quantify

  Incorrect

  Leaving out errors, rework and lost business understates the current cost and distorts the return calculation in either direction.

• DOnly at the one-year review, using industry averages

  Incorrect

  Industry averages describe other organizations. The CFO's question is about this firm's process, which needs its own measured baseline.

You cannot show what AI changed without knowing where you started. A baseline captured before deployment, covering labor, quality, delays and lost opportunities, is the foundation for realistic ROI projections and for measuring actual benefits afterward.

• AA live dashboard that tracks cost per interaction against value delivered

  Correct

  AWS guidance treats ROI as a dynamic KPI tracked on a dashboard, combining financial metrics such as cost per interaction with value metrics such as hours saved, revenue lift and satisfaction.

• BAn annual recalculation using the same method as the launch business case

  Incorrect

  A yearly snapshot misses changes in token consumption, adoption and model prices as they happen, which is why AWS says ROI cannot be a static calculation.

• CNo further ROI work, because the launch calculation already justified the project

  Incorrect

  Costs and value both change after launch, so the original calculation quickly stops describing the system that is actually running.

• DA report of monthly token spend, because cost is the only part that changes

  Incorrect

  Value changes too: adoption, user behavior and model quality all move. Tracking only spend shows half of the return.

The cost of a generative AI application is driven by consumption and scaling, and its value by adoption and quality, so both move after launch. Tracking ROI continuously turns it from a historical justification into an operational signal that guides further investment.

• AHigher customer satisfaction scores after support conversations

  Correct

  Customer satisfaction is a non-financial benefit. It matters to the business but does not appear directly as a cost or revenue figure.

• BLower contact center labor cost per resolved case

  Incorrect

  A reduction in cost per case is a tangible, financial benefit that can be measured directly.

• CAdditional revenue from upgrade offers accepted during conversations

  Incorrect

  Revenue from accepted offers is a tangible benefit that appears in the financial results.

• DImproved employee experience for agents who handle fewer repetitive questions

  Correct

  Employee experience and the productivity it supports are non-financial benefits, valued but not booked directly as revenue or cost.

• ELower infrastructure spend after retiring the old interactive voice system

  Incorrect

  Avoided infrastructure spend is a tangible cost reduction with a clear financial value.

A complete business case counts both financial benefits, such as lower costs and higher revenue, and non-financial ones, such as better customer and employee experience. Intangible benefits still need defined measures, for example satisfaction scores, so that they can be tracked.

• AAWS Cost Explorer

  Incorrect

  Cost Explorer analyzes costs and usage that already exist and forecasts from that history. With no resources deployed, there is nothing for it to analyze.

• BAWS Budgets

  Incorrect

  Budgets tracks actual and forecasted spending against thresholds you set and sends alerts. It monitors spending; it does not build an estimate for an architecture that does not exist yet.

• CAWS Pricing Calculator

  Correct

  Pricing Calculator is a free planning tool for modeling solutions before building them and creating cost estimates that can be shared and exported.

• DAWS Marketplace

  Incorrect

  AWS Marketplace is where you find and buy third-party software and services. It does not estimate the cost of AWS resources you plan to use.

Use Pricing Calculator to plan and estimate before you build, Cost Explorer to analyze and forecast what you are already spending, and Budgets to be alerted when spending approaches a threshold.

• AInstance-based hosting of the model on an endpoint that runs around the clock

  Incorrect

  Self-hosted models on SageMaker AI or EC2 endpoints are billed for uptime, so an always-on endpoint costs the same on days with no requests.

• BConsumption-based, per-token inference on a managed model service

  Correct

  Managed models charge for consumption measured in input and output tokens, so cost follows usage and idle days cost nothing for inference.

• CSeat-based subscriptions for every editor who might use the feature

  Incorrect

  Per-user subscriptions price access for named people. They do not track the volume of summarization requests a product feature generates.

• DA long-term capacity commitment sized to the busiest expected day

  Incorrect

  A commitment sized to the peak of an unpredictable, mostly idle workload pays for capacity that sits unused on most days, the opposite of tracking usage.

Match the pricing structure to the usage pattern. Consumption-based pricing makes cost follow usage, which suits low or unpredictable volume; always-on instances and fixed commitments charge whether or not requests arrive, and seat-based pricing charges per person rather than per request.

• AAWS Cost Anomaly Detection monitoring Bedrock spend

  Incorrect

  Cost Anomaly Detection alerts on spend patterns that its machine learning models judge unusual, allowing for seasonality and natural growth; it does not compare spend with the amount the CFO approved.

• BAWS Budgets with a Bedrock cost budget and forecast alerts

  Correct

  AWS Budgets sets spending limits for services and can alert on forecasted spend before it accrues, as well as on actual spend, by email or Amazon SNS.

• CAWS Cost Explorer filtered to Bedrock, with its spending forecast

  Incorrect

  Cost Explorer shows costs and forecasts future spending, but someone has to open it; the forecasts are used to set budgets, which send the alerts.

• DAWS Pricing Calculator estimates of monthly Bedrock usage

  Incorrect

  Pricing Calculator estimates costs before building; it does not track the spending that is actually happening.

Cost control for AI includes early warning against an agreed amount. A budget scoped to the service, with alerts on forecasted spend, lets owners act before a variable consumption bill overruns what was approved.

• AAWS Pricing Calculator

  Incorrect

  Pricing Calculator estimates costs from assumptions before building; it does not analyze ten months of actual spend.

• BAWS Cost Anomaly Detection

  Incorrect

  Cost Anomaly Detection flags unusual spend and ranks the root causes of each anomaly; it does not show the overall trend by service or project the coming months.

• CAWS Cost Explorer

  Correct

  Cost Explorer lets you view and analyze costs and usage for up to the last 13 months and forecast likely spending for the coming months, based on past usage.

• DAWS Budgets

  Incorrect

  Budgets tracks spend against amounts you set and sends alerts; it is not the tool for analyzing ten months of trends by service.

Measuring and planning AI cost relies on visibility into actual spend. Cost Explorer shows trends and cost drivers from real usage and projects them forward, which is the evidence a budget decision needs.

• AIndex only the records from recent years, which staff search most

  Incorrect

  Dropping older records conflicts with the records policy, which requires every record to stay searchable; AWS's guidance cuts vector store cost by shrinking each vector instead.

• BSwitch to another embedding model whose default vectors are smaller

  Incorrect

  Another embedding model would need the security review that is not available this year, and the approved model already offers smaller output vector sizes of 512 and 256 besides its default of 1,024.

• CTrim the instructions in the prompt sent with each staff question

  Incorrect

  A shorter prompt lowers the token cost of inference, but the estimate shows that the vector store, not inference, is the largest cost line.

• DUse a smaller vector size of the approved model and retest retrieval

  Correct

  AWS recommends embedding with the smallest vector length the model supports to lower vector store costs, then testing retrieval quality, because smaller vectors can make retrieval less accurate.

In retrieval augmented generation, the vector store is a cost line of its own, and shorter vectors lower the costs attributed to it. Choosing a smaller vector size cuts that cost without dropping data or changing models, but it is a deliberate cost and quality trade-off, so retrieval must be retested before the change goes live.

• AMark only the user's new question for the guardrail to evaluate

  Correct

  AWS's cost guidance is to tag user input so that filtering applies only to user-provided content, leaving retrieved passages and conversation history unevaluated; with the Converse API, guardContent blocks limit the assessment to the marked text.

• BApply the guardrail only to the model's responses

  Incorrect

  Guardrails can screen both what users enter and what the model generates; checking responses only would stop screening user questions, which the trust and safety lead requires.

• CSwitch to a smaller, cheaper foundation model

  Incorrect

  Guardrail charges are based on the text units that each enabled filter evaluates, so a cheaper model lowers the model charges but leaves the guardrail charges as they are.

• DReplace the content filters and denied topics with free word filters

  Incorrect

  Word filters are free but match only exact words and phrases, while content filters detect harmful categories and prompt attacks, so the same policies would no longer screen user questions.

Safety controls have costs too, charged on the text they assess. Pointing the guardrail at the untrusted part of each request, the user's new input, keeps protection where it is needed and stops paying to screen trusted context again and again.

• AThe Flex tier for its discount

  Incorrect

  Flex suits workloads that can handle longer processing times, not an interruption-intolerant service.

• BThe Standard tier, with no reservation, for everyday workloads

  Incorrect

  Standard serves everyday workloads without reserved capacity.

• CBatch inference with overnight results

  Incorrect

  Batch processing is asynchronous and cannot serve live trading hours.

• DThe Reserved tier with reserved tokens-per-minute capacity

  Correct

  The Reserved tier reserves prioritized capacity for mission-critical applications that cannot tolerate downtime, with overflow to the Standard tier and a fixed price per 1,000 tokens per minute.

Capacity choices are a cost and continuity decision. For steady, mission-critical demand, reserving capacity trades a fixed commitment for reliable availability.