AI & Agentic AI Readiness Check

A team upgrades from a smaller model to a much larger one and sees nicer phrasing, but factual errors still appear in answers about niche regulations. What is the best conclusion?

• AScale can improve fluency and pattern coverage, but it does not guarantee factual accuracy on specific domains
• BThe deployment likely failed because larger models should not make factual mistakes on text tasks .
• CThe issue is probably temperature only, since larger models mainly fail from randomness
• DIf the model sounds more certain after the upgrade, it is likely verifying answers internally

A product manager says, "The model answered in a very confident tone, so it probably checked itself." Which responses are technically sound? Select all that apply.

• Confident wording is a generation style and does not prove external verification
• Ask the system to cite sources or show retrieved evidence if verification matters
• Confidence usually means the model selected a fluent continuation, not that it validated the claim
• A polished explanation can still be wrong even when it sounds internally consistent

Assertion: Post-training methods like instruction tuning can make a model appear more helpful without reliably adding new factual knowledge. Reason: These methods mainly reshape response behavior and preferences rather than replaying broad world learning from scratch.

• 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 text that sounds plausible but is unsupported or fabricated, the usual term is ___.

Match each observed behavior to the most likely interpretation.

A support chatbot fails on company-specific policy questions. What is usually the best first move before considering full fine-tuning?

• AAdd retrieval over trusted policy documents and test grounded prompting
• BTrain a larger model immediately because scale fixes domain gaps fastest
• CLower the context window so the model is forced to answer more directly
• DCollect only thumbs-up user ratings and wait for behavior to self-correct

Which observations are evidence that prompting changes behavior but does not necessarily change the model's underlying knowledge? Select all that apply.

• Adding a step-by-step template improves answer structure on the same task
• The model answers correctly only when given relevant reference text at runtime
• A one-time system prompt permanently fixes future mistakes across deployments .
• Different phrasings of the same request lead to different quality levels

Arrange these interventions from fastest to try to slowest to try when a general-purpose model underperforms on an internal workflow.

Assertion: Benchmark gains on public leaderboards do not guarantee equivalent gains on your business workflow. Reason: Real tasks often depend on domain context, process constraints, and failure costs that generic benchmarks underweight.

• 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

If a model's output varies across runs because tokens are sampled rather than always chosen deterministically, that behavior is due to ___.

A CTO asks whether AI outputs should be treated like database lookups. Which replies are appropriate? Select all that apply.

• Generated answers are synthesized predictions, not guaranteed record retrievals
• If exactness matters, pair generation with validation or source lookup
• A low-temperature setting makes outputs more consistent, but it does not make the model a reliable source of record
• Repeated agreement across runs is useful but still not proof of correctness

Match each intervention to the misconception it helps correct.

A colleague says, "The model answered a question about an event from yesterday, so it must have learned it live from the internet." What is the best response?

• AIt may have been given fresh context at runtime, which is different from the base model learning online
• BThat is expected because deployed models continuously retrain on current web pages by default
• CThe answer may have come from a connected search or retrieval step, not from the model updating itself .
• DRecent facts mainly appear when the context window is large enough to trigger self-update

A generated image of a factory floor looks realistic, but a safety engineer notices impossible reflections and tool positions. What is the best takeaway?

• AVisual realism is not evidence of physical accuracy or scene consistency
• BImage models mainly fail because export resolution was too low
• CIf most viewers accept it, the geometry is accurate enough for operations
• DPhotorealistic outputs do not guarantee that spatial relationships or object placement are correct

Your team wants to use a multimodal model to inspect uploaded product photos for possible defects. Which cautions are justified? Select all that apply.

• The model may miss subtle defects even if it describes obvious features well
• A realistic visual explanation does not prove the identified defect is real
• If the caption is coherent, the visual encoder has likely captured all critical details
• Lighting, angle, and image quality can strongly affect detection reliability

Match each use case to the capability it most directly depends on.

Assertion: A model that can describe an image well may still fail at counting objects in that image. Reason: Descriptive fluency and precise visual reasoning are not the same capability.

• 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 generated video looks believable but violates basic real-world mechanics, it shows high visual plausibility but weak ___.

A team is comparing a text-only model and a multimodal model for customer support. Which tasks genuinely favor the multimodal approach? Select all that apply.

• Interpreting a screenshot of an error dialog sent by a user
• Answering questions from a policy PDF after text extraction alone .
• Checking whether a product photo matches the claimed damaged part
• Rewriting a plain text refund email into a friendlier tone

Marketing wants to use an image generator for new ad concepts. What is the most appropriate operational assumption?

• AIt can speed up concept exploration, but human review is still needed for brand, legal, and factual issues
• BGenerated images are safer than stock photos because they contain no inherited risks
• CIf the prompt names the product correctly, visual details will stay on-brand automatically .
• DOnce a style prompt is approved, later outputs should remain consistent without checks

Arrange a sensible pipeline for answering a user's question about a screenshot they uploaded.

Match each failure mode to the most likely multimodal weakness.

Which claims about multimodal models are too strong and should be challenged? Select all that apply.

• If the model can caption an image, it can probably do precise measurement from it
• Adding image input automatically makes text answers more trustworthy
• A video summary can omit important events even when it sounds complete .
• Generated audio that sounds natural may still contain content errors

A system that turns spoken customer calls into written text relies first on automatic speech ___.

Assertion: For some document workflows, OCR plus a strong text model can outperform a more complex end-to-end multimodal setup. Reason: If the key task depends mainly on extracted text, the added multimodal stack may increase complexity without adding much value.

• 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 user uploads a chart image and asks for a trend summary. The model gives a polished answer but swaps the rising and falling series. What should you infer?

• AThe system may understand the request language better than the visual data details
• BThe model likely has no vision component at all because it produced the wrong trend
• CCharts are safe because data visualizations cannot trigger multimodal errors
• DThe model may be fluent at summarizing charts while still misreading legends or series mappings

A team plans to paste customer support transcripts into a public AI assistant after removing names. Which concerns still remain? Select all that apply.

• Other identifiers or context may still allow re-identification
• Contractual or policy limits may prohibit sending the data externally anyway
• Removing names guarantees the content is no longer sensitive .
• The transcripts may contain secrets or regulated details beyond personal names

An employee wants to use AI to draft performance feedback about a direct report. What is the strongest concern?

• AThe system may amplify bias or unsupported claims in a high-impact people decision
• BPerformance feedback is safe because it is internal rather than customer-facing
• CIf the tone sounds balanced, the underlying judgments are probably objective
• DBias matters only if the model was trained on employment data from the same company

Match each practice to the risk it primarily addresses.

Trying to remove or mask personal identifiers from data before wider analysis is called ___.

Which uses of AI most clearly need a human in the loop because the stakes or reversibility are high? Select all that apply.

• Approving a loan denial notice to a customer
• Publishing a mass layoff communication drafted by the system
• Generating several internal slogan ideas for a workshop .
• Automatically changing tax settings in payroll records

Assertion: A prompt injection attack can reach a model through retrieved documents, not only through direct user chat. Reason: The model may treat external text from tools or retrieval as instructions unless the system is designed to isolate and validate it.

• 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 compliance team asks whether an internal AI assistant should log prompts and outputs. What is the best answer?

• AUsually yes, with appropriate controls, because logs support investigation, auditing, and improvement
• BNo, because any logging automatically violates privacy law in all jurisdictions
• COnly successful interactions should be logged so failure patterns do not bias analysis
• DLogging is useful whether the model is open-source or vendor-hosted, as long as governance is in place

An AI tool proposes citations in a research memo. Which checks are appropriate before relying on them? Select all that apply.

• Verify that the cited source actually exists and supports the claim
• Assume citations are valid if the formatting matches your style guide .
• Check whether quoted numbers or passages are present in the source
• Notice whether several citations come from the same hallucinated journal or author

Match each scenario to the best control choice.

When a model's output systematically disadvantages a group for reasons unrelated to the task, the issue is commonly described as ___.

A vendor says its model has safety filters, so your company does not need internal controls. Which responses are justified? Select all that apply.

• Vendor controls help, but local policy, monitoring, and approval rules are still needed
• Your risk depends on your use case, data, and connected tools, not only the base model
• If the provider blocks obvious abuse prompts, most important governance concerns are already covered
• Procurement can replace internal governance because the vendor owns the model behavior

Assertion: A low-risk pilot can still create serious exposure if it quietly gains access to production data and actions. Reason: Risk depends on actual permissions and impact pathways, not on whether the team labels the experiment a pilot.

• 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 manager asks for one rule of thumb for employee use of general-purpose AI tools. Which rule is strongest?

• ADo not paste anything you would not be comfortable sharing under your organization's approved data policy
• BPublic AI tools are fine for internal data if you avoid customer names
• CUse them freely for drafts because generated content never leaves your browser .
• DSensitive use is acceptable when the answer is only for personal productivity

An agent can search docs, call APIs, and send emails. Which added design choice most directly reduces the risk of harmful autonomous actions?

• ARequire approval gates for high-impact or irreversible actions
• BIncrease the model size so planning errors become rare
• CLet the agent retry failed actions until one succeeds
• DHide intermediate steps from users so the system stays focused .

A team says, "If we add retrieval, the agent will stop hallucinating." Which replies are accurate? Select all that apply.

• Retrieval can reduce unsupported answers, but the model can still misread or invent around the retrieved text
• Grounding helps only if the retrieved material is relevant and the agent uses it correctly
• Once citations are attached, hallucination is effectively eliminated .
• The retriever itself can surface stale, low-quality, or conflicting sources

Match each component to its main role in an agentic system.

Assertion: An agent that successfully completes many steps in a row may still be less reliable overall than a simpler single-shot assistant. Reason: Each extra tool call or decision point creates another opportunity for compounding error.

• 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

Supplying an LLM with relevant external documents at runtime to improve answer fidelity is commonly called ___.

Which signs suggest an agent should not be trusted to run fully autonomously yet? Select all that apply.

• It occasionally uses the wrong tool despite having the right permissions
• It completes demo tasks well but fails unpredictably on slightly varied real cases
• It explains plans clearly and produces readable logs .
• It cannot distinguish between reversible drafts and irreversible external actions

A grounded QA system retrieves the correct policy paragraph but still answers incorrectly. What is the most likely lesson?

• AGrounding retrieved context is helpful, but answer synthesis can still fail
• BThe retriever must have returned nothing useful if the final answer was wrong
• CEven with the right paragraph present, the model can still misread or overgeneralize from it
• DThe failure proves retrieval harms accuracy compared with no external context

Arrange the typical loop of a retrieval-using agent answering a business question.

Match each failure pattern to the most relevant mitigation.

You are designing an agent to update CRM records from customer emails. Which controls are prudent? Select all that apply.

• Show the record diff and require approval for high-risk field changes
• Constrain tool access so the agent can edit only allowed objects and fields
• Let the agent infer missing identities from message tone when IDs are absent .
• Keep an audit trail of retrieved evidence, proposed edits, and executed actions

An agent's ability to call software functions or external services is usually described as tool ___.

Assertion: Limiting an agent's available tools can improve overall reliability even if it reduces theoretical flexibility. Reason: A smaller action space can make planning simpler and lower the chance of harmful or irrelevant actions.

• 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 why an agent needs both retrieval and memory. What is the best answer?

• ARetrieval brings in relevant external knowledge, while memory preserves task state or prior context across steps
• BMemory stores all world facts permanently, so retrieval is needed only for backup .
• CRetrieval is for older information, while memory is for recent internet content
• DThey are interchangeable features and teams usually pick one for convenience