Task 1.2: Identify and select appropriate AI solution types.

• AFine-tune a model on past orders to learn the rate card

  Incorrect

  Fine-tuning adapts a model's behavior from examples, but the output is still probabilistic. It adds cost and the risk of a wrong fee for a calculation that a few rules perform exactly.

• BImplement the rate card as rule-based logic in the checkout software

  Correct

  The problem can be written down as a clear and compact set of rules, and the business needs exact, reproducible results. AWS guidance is to consider a non-AI solution whenever such rules exist.

• CUse a generative AI model and place the rate card in the prompt

  Incorrect

  Supplying the table in the prompt helps a model use it, but the model still generates its answer probabilistically, so an exact, auditable fee is not assured.

• DTrain a machine learning regression model on historical shipping fees

  Incorrect

  Machine learning suits relationships too complex to write as rules. Here the relationship is already known exactly, so a model would only approximate it.

Choose AI when it provides clear, substantial benefits over competing solutions, not because it is possible to apply. A published rate card is a compact rule set with a requirement for exact repeatability, which is exactly where traditional software beats a model.

• AKeep the fixed workflow and use a model only to classify and route at the judgment step

  Correct

  An LLM-augmented workflow keeps code paths deterministic and uses a model only for the decision that needs it. AWS advises raising a system's agency only when the task complexity requires it.

• BGive an autonomous agent the goal "process invoices" and let it plan every step and choose its tools

  Incorrect

  A fully autonomous agent decides its own steps in a reasoning loop, which gives up the predictability the CFO requires for steps that are already fixed.

• CReplace the workflow with a multi-agent system in which agents negotiate each invoice's handling

  Incorrect

  A multi-agent system coordinates several agents that each reason and act. It adds autonomy and coordination overhead to a process whose steps are already known in advance.

• DKeep the whole process manual, because any use of AI makes the ledger unauditable

  Incorrect

  Using a model for one bounded classification does not make the rest of the workflow unauditable; the deterministic steps and the posting logic stay as they are.

Agency exists on a spectrum, from deterministic workflows with one LLM-assisted decision to agents that plan and act on their own. Match the level of autonomy to the complexity of the task: here only one step needs judgment, so an LLM-augmented workflow adds AI where it helps and keeps the rest predictable.

• ARetrain every week on the latest data, so that the model is never more than a week out of date

  Incorrect

  Retraining on a fixed schedule regardless of performance is an AWS-listed anti-pattern. With months of stable accuracy, most of roughly two million dollars a year would buy retraining that changes nothing.

• BRetrain only when customer complaints and service credits show that the predictions have gone wrong

  Incorrect

  This waits for the model to fail before acting, which AWS warns against. Complaints and credits arrive after the losses they report.

• CAgree on accuracy thresholds tied to delivery promises, track them, and fund a retraining when one is breached

  Correct

  AWS recommends setting minimum acceptable performance with business stakeholders, monitoring the model in production and retraining only when performance falls below those levels, which keeps quality without paying for unneeded runs.

• DRetrain once a year after the annual data close, when a full year of new records is available

  Incorrect

  A yearly run is still a fixed schedule that ignores performance. A drop as fast as last year's would leave the model degraded for months before the next run.

Models drift as the world they predict changes, so they need updates after launch, but updates cost money. Tying retraining to agreed performance thresholds spends the budget when the model actually degrades, and catches a sudden drop before it turns into months of broken promises.

• AProvide approved AI tools and monitor AI use to detect unapproved tools

  Correct

  AWS's security guidance on shadow AI is a dual approach: provide sanctioned tooling, which reduces shadow use and improves visibility, and build observability, including endpoint monitoring for unauthorized tools.

• BReinforce the ban with a firm-wide memo and disciplinary consequences for violations

  Incorrect

  Outright bans and slow adoption are what drive employees to consumer tools in the first place. A stricter ban keeps the demand and the hidden use.

• CBlock consumer AI websites on the firm's network

  Incorrect

  Network blocks do not reach personal phones, and without an approved alternative the demand simply moves elsewhere.

• DAllow any AI tool but require associates to remove client names before pasting

  Incorrect

  Manual redaction is error-prone, and documents still leave the firm's control for services whose data handling the firm has not reviewed.

Shadow AI grows when people see value and have no sanctioned way to get it. Giving them approved tools brings usage into view and under policy, and observability catches what still slips through. Prohibition alone drives the risk underground.

• AIt produces fluent, natural-language replies that sound like a member of staff

  Incorrect

  Fluent language generation is a capability of the underlying foundation model that the existing chat assistant already has, so it does not distinguish the agent.

• BIt decides on its own which steps to take toward its goal

  Correct

  Autonomy is the foundation of agent behavior: an agent chooses the next appropriate action toward its goal instead of following a fixed script.

• CIt answers from knowledge learned in pretraining

  Incorrect

  Answering from pretrained knowledge is what the existing assistant already does; it involves no planning and no action.

• DIt follows a fixed decision tree that the contact center team wrote and maintains

  Incorrect

  A fixed decision tree is scripted automation. Agents are defined by choosing their own actions toward a goal rather than following a predetermined script.

• EIt calls external tools, such as a booking API, to take actions

  Correct

  Tool use extends an agent from understanding language to performing actions, such as checking availability and changing a booking.

An AI agent perceives its environment, reasons about a goal and acts: it plans its own next steps and uses tools such as APIs to change things in the world. A chat assistant generates answers but takes no action, so the jump to an agent brings both new value and new risks to manage.

• AAs approved, because the security review is complete and found no issues

  Incorrect

  Approval should follow the complete review. A passed security check does not settle the open legal question about code ownership.

• BAs blocked until legal approves, without listing it on the internal page

  Incorrect

  Blocking a tool still under review and hiding its status makes the governed path slow and opaque, which pushes engineers toward unapproved alternatives.

• CLeave it off the list, which covers production tools only

  Incorrect

  A catalog that leaves out tools in review gives employees no clear answer, which is the gap shadow AI fills.

• DAs under evaluation, with stated conditions for limited use

  Correct

  A tool whose review is still in progress belongs in the under evaluation category. Publishing its status and conditions keeps the governed path visible and responsive, which is what keeps people off unapproved tools.

A transparent classification of AI tools only works if every tool has a clear status, including the ones still being reviewed. AWS governance guidance tracks which models are approved for which uses and stresses responsive approval processes so that the governed path is more attractive than the shadow path.

• AGenerative AI for every single contact, including the simple balance checks

  Incorrect

  Simple, predictable requests do not need generative AI; structured flows handle them reliably at lower cost.

• BFixed menu flows for every contact, including complex disputes

  Incorrect

  Fixed flows cannot handle the variety of complex or ambiguous requests.

• CHuman agents for every contact, using AI only for reporting

  Incorrect

  Routing everything to people gives up automation where it works well.

• DSimple flows for routine requests, AI for complex ones, then a person

  Correct

  AWS's autonomous call center scenario handles structured requests with intent-based flows, uses generative AI for complex or ambiguous requests and falls back to a live agent.

The choice between rules and AI is often per request type. Combining deterministic flows, generative AI and human fallback uses each where it fits best.

• ABlocked for all uses until the provider offers enterprise terms

  Incorrect

  AWS warns that banning generative AI pushes employees toward unmanaged consumer apps, and its scoping guidance allows Scope 1 apps to be used with public, non-proprietary data, so a full block fails the goal of letting the low-risk work go ahead.

• BUnder evaluation for all uses until the provider's model is threat-modeled

  Incorrect

  Thorough threat modeling is AWS's guidance for Scopes 3 to 5, where the organization builds the solution; for a consumer app the buyer evaluates the provider's risks, and holding every use keeps the low-risk work waiting.

• CApproved for all marketing content, since staff review every tagline

  Incorrect

  Reviewing outputs does not control what goes in: the provider may use submitted content to improve its services, so pasting the unreleased roadmap would expose it.

• DApproved for public, non-proprietary content; blocked for the roadmap

  Correct

  For Scope 1 consumer apps, AWS advises prioritizing public, non-proprietary data because providers may use submitted data to enhance their models or services. That lets the tagline work proceed and keeps the roadmap out.

A classification can approve a tool for some data and block it for other data. With a consumer app whose provider may reuse what it receives, only public information belongs in it; proprietary plans need a tool with enterprise protections, or no AI tool at all.

• AApproved for all uses, because it performed well internally

  Incorrect

  Internal performance does not show that the stricter public-facing thresholds are met.

• BBlocked for all uses, because public use carries risk

  Incorrect

  Blocking everything gives up the internal value the evaluation already confirmed.

• CApproved internally; public-facing use needs its own review

  Correct

  AWS advises calibrating risk strategy to deployment context and user exposure, noting that internal applications warrant different controls than public-facing AI systems.

• DLeft unclassified until every possible use of it has been tested

  Incorrect

  Leaving it unclassified gives employees no guidance and invites unapproved use.

A transparent classification of AI tools can be specific about where a tool may be used. Matching approval to exposure lets organizations gain value safely while stricter cases get the review they need.

• AAdding more keyword rules until all phrasings are covered

  Incorrect

  The variety of phrasing is the problem; more rules keep growing without catching every variation.

• BAI that interprets the meaning of varied emails

  Correct

  AWS guidance points to generative AI when a use case requires understanding widely varying inputs, including natural language, which is where fixed rules fall short.

• CA fixed menu that forces customers to pick a category first

  Incorrect

  Forcing categories shifts the burden to customers, who often choose wrongly.

• DSorting emails by subject line

  Incorrect

  Sorting changes the order of emails without interpreting what each one is about, so misrouting continues.

Rules work when inputs are predictable. When people express the same thing in endless ways, AI that understands language is the better tool for classifying and routing.

• AIt matches words in the message to the name of each system

  Incorrect

  Customers do not name internal systems, and the agent does not rely on keyword matching; the model reasons over the tool descriptions and instructions it was given.

• BIt calls the three systems in a fixed order until one responds

  Incorrect

  A fixed order is scripted automation; in an agent, the model selects the appropriate tool for each step from the tools it has been given.

• CIt learned which system to call when the model was first trained

  Incorrect

  The model was not trained on this retailer's private systems and was not retrained; it learns about the tools from the list and instructions it is given when it runs.

• DThe model reasons over its tool list and instructions

  Correct

  An agent is given a list of tools with instructions on how to use them; the model provides the reasoning to select the appropriate tool, guided by the system prompt.

Tool use is a core agent capability. Clear tool descriptions and instructions let the model reason about which action fits each step, which is why agent design includes careful definition of tools rather than routing rules.