AI & Agentic AI Readiness Check

A team upgrades from a smaller model to a much larger one and notices outputs are more fluent, but factual errors still appear in specialized answers. What is the best conclusion?

• AThe larger model is probably under-trained and just needs more epochs
• BFluency and correctness are different; scale can improve performance without guaranteeing truth
• CIf the model sounds coherent, the remaining errors are likely just edge cases in prompting
• DLarger models mainly improve breadth and fluency, but that does not ensure domain-specific accuracy

A PM says, "The model answered confidently, so it must have checked itself." Which responses are technically sound? Select all that apply.

• Confident wording is a style pattern, not evidence of verification
• Self-consistency across multiple runs can help, but matching outputs still does not prove truth
• If temperature is set low, confident answers are more repeatable, not necessarily verified
• Confidence can still be misleading even after fine-tuning on company data

Assertion: Post-training methods such as instruction tuning can noticeably change how a model responds to users. Reason: Post-training replaces the model's pretraining knowledge with freshly curated facts.

• ABoth A and R are true and R is the correct explanation of A
• BBoth A and R are true but R is NOT the correct explanation of A
• CA is true but R is false
• DA is false but R is true

When a model produces a detailed but unsupported answer as if it were true, the common term is ___.

Match each team statement to the most accurate interpretation.

A support workflow needs reliable answers about constantly changing product policies. The team can either fine-tune on last quarter's documents or keep a base model and connect it to a live policy source. Which is the better first choice?

• AFine-tune first, because all domain knowledge should live inside model weights
• BUse the live policy source first, because frequently changing facts are better handled outside the model
• CFine-tune first, because retrieval mainly helps tone rather than factual accuracy
• DUse the base model alone first, because adding external context often introduces extra complexity without improving freshness

A team is disappointed that a model scores well on a public benchmark but struggles in their workflow. Which explanations are plausible? Select all that apply.

• The benchmark may measure a different skill mix than the workflow requires
• The workflow may depend on grounding in current enterprise data that the benchmark does not test
• A strong benchmark score guarantees production reliability, so the issue is probably user training
• The benchmark may hide failure modes like formatting, tool use, or long-context degradation

A model is underperforming on an internal Q&A task. Put the investigation steps in the most sensible order before deciding to fine-tune.

Based on the chart, which model should a team choose if they need the best factual reliability per dollar rather than the highest raw benchmark score?

• AModel Orion
• BModel Vega
• CModel Lyra
• DModel Nova

Two teams compare outputs at temperature 0 and say, "Now the model is deterministic, so we've solved the quality issue." What concerns still remain? Select all that apply.

• The same wrong answer can now appear more consistently
• Deterministic output can still reflect missing context or flawed reasoning
• Temperature 0 makes answers more repeatable, but not inherently more grounded
• Reduced variation can make hidden systematic errors easier to overlook

An evaluation that reflects the tasks, constraints, and data of your real workflow is more useful than relying only on a generic public ___.

Assertion: A model can improve on average across many tasks and still fail badly on a specific high-stakes task. Reason: Aggregate performance hides distributional weaknesses and domain-specific gaps.

• ABoth A and R are true and R is the correct explanation of A
• BBoth A and R are true but R is NOT the correct explanation of A
• CA is true but R is false
• DA is false but R is true

A stakeholder asks whether AI outputs should be treated like database query results. What is the best response?

• AYes; once prompted carefully, generative models return stored facts deterministically
• BMostly yes; errors usually indicate missing indexes in the model's memory
• CNo; generated answers are probabilistic outputs and should be validated when correctness matters
• DNo; because AI systems cannot answer factual questions at all

A generated product demo video looks highly realistic. What is the safest conclusion?

• AThe underlying motion and physical interactions are probably accurate because the video looks natural
• BVisual plausibility does not guarantee physical or factual accuracy
• CIf the model rendered reflections correctly, the rest of the scene is likely trustworthy
• DHigh realism can still coexist with errors in causality, timing, or physical behavior

A team wants to use a multimodal model to review equipment photos and maintenance notes together. Which statements are sound? Select all that apply.

• Combining image and text can improve usefulness when the task genuinely depends on both
• A convincing visual description can still miss fine-grained defects relevant to the workflow
• If the model can caption the image well, it is effectively doing expert inspection already
• Task-specific evaluation is needed because multimodal fluency can hide important failure modes

Match each capability claim to the most accurate caution.

When an image model creates a hand with the right overall look but the wrong number or arrangement of fingers, that is a failure of visual ___ rather than proof the whole system is useless.

Your team needs slide illustrations for internal workshops, not photorealistic advertising assets. Which setup is most sensible?

• AUse the most photorealistic model available, since realism is the same as usefulness
• BChoose a generator whose style control, speed, and editing workflow fit the task
• CFine-tune a huge image model immediately, because internal branding always requires custom weights
• DAvoid generative imagery entirely, because non-photorealistic outputs cannot be production quality

A colleague says, "If the model can transcribe and summarize calls, we no longer need to worry about missing action items." Which responses are justified? Select all that apply.

• Transcription errors can propagate into the summary and action list
• Even perfect transcription does not ensure the model inferred the right priorities
• Audio quality, accents, overlap, and domain jargon can affect results
• Once summarization works on five sample calls, the workflow is effectively validated

Assertion: A multimodal model that performs well on clear product photos may still fail on low-light warehouse images. Reason: Input distribution changes such as lighting, angle, and resolution can materially alter model behavior.

• ABoth A and R are true and R is the correct explanation of A
• BBoth A and R are true but R is NOT the correct explanation of A
• CA is true but R is false
• DA is false but R is true

Based on the chart, which system is the best choice for a receipt-processing workflow that values high field extraction accuracy but cannot tolerate long latency?

• AVisionFlow-A
• BVisionFlow-B
• CVisionFlow-C
• DVisionFlow-D

You are assessing whether a text-only model or a multimodal model is needed for an insurance intake workflow. Which signs suggest the multimodal route is genuinely necessary? Select all that apply.

• Claims depend on interpreting photos of damage alongside written descriptions
• The main challenge is rewriting customer emails into a standard tone
• Important evidence appears in scanned forms with handwritten annotations
• The workflow mostly involves retrieving policy clauses from a structured database

A generated chart that looks polished but contains invented values is visually plausible but not data-___.

Match each use case to the most important evaluation focus.

A designer says, "Since the image model can follow our prompt exactly, it probably understands our brand guidelines." What is the best reply?

• AYes; prompt adherence is strong evidence of stable conceptual understanding
• BMaybe partly, but repeated compliance on a few prompts does not guarantee robust understanding across cases
• CNo; image models ignore textual instructions unless they are fine-tuned
• DYes; because brand understanding is mostly the same as memorizing color palettes

Based on the chart, which conclusions are justified about the four speech-to-summary systems? Select all that apply.

• EchoNote has the best summary score but also the slowest end-to-end runtime
• PulseBrief is the fastest, but its lower transcript accuracy may limit trust in summaries
• VoxMap is the best balanced option if the team wants strong summary quality without the highest delay
• ClearThread can be ignored because its summary score is below EchoNote's

Assertion: Removing names from a document does not automatically make it safe to paste into a public AI tool. Reason: Other details can still identify people or expose sensitive business information.

• ABoth A and R are true and R is the correct explanation of A
• BBoth A and R are true but R is NOT the correct explanation of A
• CA is true but R is false
• DA is false but R is true

Your company is considering an external AI assistant for drafting customer responses. Which precautions are sensible? Select all that apply.

• Check whether sensitive prompts or outputs may be retained or used by the provider
• Set clear rules about what categories of data employees must not paste in
• Assume contractual compliance language makes all downstream use safe by default
• Require human review before sending externally visible high-impact messages

When a model follows malicious instructions hidden inside retrieved content rather than the user's actual goal, that is a prompt ___.

A manager says, "It's fine to let the model draft legal language without review because we trained it on our templates." What is the best response?

• AReasonable, because template training mostly eliminates legal risk
• BNot reasonable, because familiar style does not guarantee correctness or appropriateness in a specific case
• CReasonable, if the model uses low temperature and fixed formatting
• DNot reasonable only if the templates contained personal data

Match each risk scenario to the strongest primary control.

Which situations can create prompt-injection risk even if end users never type directly to the model? Select all that apply.

• A retrieval system passes web pages into the model context
• A tool returns text from an external API and the model treats it as instructions
• A static internal taxonomy is hard-coded and never updated from outside inputs
• An email-processing pipeline lets the model read message bodies before acting

Based on the chart, which deployment looks least appropriate for handling sensitive HR cases?

• AConfig A
• BConfig B
• CConfig C
• DConfig D

A safety measure that requires a person to approve high-impact outputs or actions before execution is a human-in-the-___ control.

Assertion: An AI summary can be unsafe to rely on even when every sentence in it is individually plausible. Reason: Safety depends on omission, context, and downstream use, not only sentence-level plausibility.

• ABoth A and R are true and R is the correct explanation of A
• BBoth A and R are true but R is NOT the correct explanation of A
• CA is true but R is false
• DA is false but R is true

A business unit wants an AI tool to generate hiring-screen summaries. Which evaluation practices are appropriate? Select all that apply.

• Test performance across varied candidate profiles and communication styles
• Focus only on average accuracy because subgroup analysis is usually too noisy
• Review whether the summary leaves out context that a recruiter would consider material
• Check whether the model's recommendations become more confident when temperature is lowered

For a new high-impact AI workflow, order these steps from most sensible earliest action to latest.

An employee says, "This note is safe to paste into a public model because it has no customer names." Which missing risk is most important to point out?

• AWithout names, there is no meaningful privacy risk left
• BOperational details, timelines, or account patterns can still expose sensitive information
• CPublic models only store final answers, not prompts, so exposure is unlikely
• DThe bigger risk is mainly that the model will write in the wrong tone

Based on the chart, which conclusions support deploying Model Harbor for internal-only drafting but not external auto-send? Select all that apply.

• Its hallucination rate is lower than DraftPro's but still high enough to warrant review
• Its privacy risk score is acceptable only because the use is internal
• Its human-review requirement is lower than SentinelWrite's, which may tempt overautomation
• Because Harbor beats one peer on quality, external auto-send is likely acceptable

Match each design choice to the primary problem it addresses in agentic systems.

A team is building an agent that can email vendors, update records, and cancel orders. Which capability most needs a human approval gate?

• ADrafting a status summary for internal review
• BLooking up inventory levels in a read-only system
• CCanceling an order that cannot be easily restored
• DReformatting supplier notes into a template

A product lead says, "Once we add retrieval, hallucinations are solved." Which responses are accurate? Select all that apply.

• Retrieval can reduce unsupported answers, but the model can still misread or overstate the retrieved content
• Grounding depends on retrieval quality, ranking, and how the model uses the context
• If the documents are authoritative, any answer drawn from them is automatically trustworthy
• Retrieved context can itself contain ambiguity, conflict, or malicious instructions

Connecting a model to external documents or databases so its answers rely on current source material is commonly called ___.

Assertion: An agent can fail safely on a planning step but fail unsafely on an execution step. Reason: The consequences of tool use depend not only on reasoning quality but also on permissions and action design.

• ABoth A and R are true and R is the correct explanation of A
• BBoth A and R are true but R is NOT the correct explanation of A
• CA is true but R is false
• DA is false but R is true

Based on the chart, which agent design is most appropriate for a finance workflow that values grounding and low action risk over maximum autonomy?

• AAgent Argo
• BAgent Beacon
• CAgent Crest
• DAgent Drift

Which design choices usually improve the safety of an agent that uses tools? Select all that apply.

• Use least-privilege tool access rather than broad default permissions
• Separate planning from execution so approval can happen before risky actions
• Let the agent rewrite system rules when it encounters blocked actions
• Prefer read-only tools where possible for early deployments

Match each symptom to the most likely agentic weakness.

Giving an agent only the minimum tool permissions needed for its task follows the principle of least ___.

An ops team wants an agent to autonomously process refund requests. Which rollout plan is most responsible?

• AGive it full refund authority immediately if offline tests looked strong
• BStart with draft recommendations and require human approval for actual refunds
• CLet it handle small refunds without logging, then add oversight later
• DFine-tune it on past tickets and remove retrieval to simplify the stack

Based on the chart, which statements about the retrieval setups are justified? Select all that apply.

• Index Pine has the highest document recall but not the best answer faithfulness
• Index Cedar likely offers the best practical balance for grounded Q&A
• Index Maple can be preferred only because its latency is lowest, regardless of quality
• Index Birch shows that better recall alone does not guarantee the most faithful answers

Assertion: A grounded agent can still provide a wrong answer even when the correct document is retrieved. Reason: The model may misinterpret, selectively quote, or combine retrieved evidence incorrectly.

• ABoth A and R are true and R is the correct explanation of A
• BBoth A and R are true but R is NOT the correct explanation of A
• CA is true but R is false
• DA is false but R is true

Based on the chart, which agent appears most vulnerable to costly runaway behavior if left unsupervised?

• ATaskPilot
• BLoopRunner
• CCaseGuide
• DFetchMate