20 free CCAR-P (CCA-P) practice questions with explanations

Vantera Compliance must map a 1,400-page banking regulation to its internal control library of 900 controls. Single-call attempts, even with the full document in context, miss obligations in later sections and produce inconsistent mappings between runs. Which decomposition approach best addresses the failure?

1. ASplit the regulation into equal 50-page chunks and run the full obligation-to-control mapping independently on each chunk, then concatenate the per-chunk results into one combined mapping.
2. BHave Claude produce a condensed summary of the regulation first, then perform the obligation-to-control mapping against the summary so the whole task fits comfortably in a single context.
3. CKeep the single-call design but move to the most capable available model, place the mapping instructions at both the start and end of the prompt, and lower the temperature for consistency.
4. DDecompose by task stage: extract obligations section-by-section into a normalized schema, map each extracted obligation to candidate controls in a second stage, then aggregate and deduplicate the results.

Halvorsen Freight, a third-party logistics provider, asks for 'GenAI in customer service.' Ticket analysis shows 40% of tickets are shipment-status lookups already answerable from the transportation management system (TMS) API, 25% are complex exceptions (damaged or misrouted freight), and customer satisfaction score (CSAT) correlates almost entirely with exception-resolution speed. Budget covers one initiative this year. Which Claude-based translation of the business need delivers the most value?

1. ADeflect the 40% status tickets with a Claude chatbot grounded in help-center articles and shipment FAQs, since status lookups are the single largest ticket category by volume.
2. BGenerate empathetic, brand-consistent reply templates for each ticket type so agents respond in a consistent tone and satisfaction improves across the service operation.
3. CSummarize closed tickets into a weekly digest of themes and root causes so management gains visibility into emerging service trends and staffing needs.
4. DBuild a tool-using Claude assistant that triages exception tickets, pulls live TMS and carrier data, and drafts a resolution plan for the human agent.

Corvex Legal runs a Claude pipeline for NDAs: intake, OCR, Claude extraction to JSON, then load into the contract-lifecycle system. Paralegals quietly correct about 12% of extracted fields inside the contract lifecycle management (CLM) system after load. Extraction accuracy has not improved in six months despite two prompt rewrites. What is the PRIMARY architectural gap?

1. AThe extraction prompt lacks enough few-shot examples demonstrating edge-case clause language, so unusual NDAs fall outside the demonstrated patterns.
2. BThere is no feedback loop: the paralegals' corrections never flow back into an evaluation set that drives prompt and pipeline iteration.
3. CThe pipeline uses a mid-tier model and should be upgraded to the most capable model available to raise extraction accuracy on difficult clauses.
4. DThere is no pre-processing stage that scores OCR quality and rejects low-confidence scans before extraction, so garbled inputs reach the model.

A national insurance carrier processes 1.8 million inbound claims emails per day, routing each into one of 12 claim categories. The routing decision is a short classification with a p95 latency budget of 500 ms, and finance has capped the AI line item at $4,000/day. A pilot with the most capable frontier model achieved 97% routing accuracy but would cost roughly 20x the budget at full volume. Which approach best fits the workload?

1. AKeep the frontier model but batch requests overnight in bulk processing windows, confirming the amortized cost fits within the $4,000/day cap
2. BUse a mid-tier model with extended thinking enabled so deliberate reasoning holds routing accuracy near the 97% pilot level at lower per-token cost
3. CUse the fastest, lowest-cost model with a few-shot prompt containing labeled examples of each category, and measure whether accuracy stays acceptable
4. DFine-tune the frontier model on historical routing data so each routing request needs far fewer prompt tokens and the marginal cost per classification drops

A legal-tech firm analyzes merger agreements for cross-document inconsistencies: about 200 requests per day, each carrying roughly 150K tokens of contract text, with attorneys reviewing every output. Errors that reach attorneys are expensive to catch, users tolerate multi-minute turnaround, and daily model spend at current volume is under $900 even on the most capable model. The team is debating whether to cut costs with a smaller model. What should they do?

1. AStay on the most capable model, since at 200 requests/day the absolute cost is small relative to the cost of missed inconsistencies
2. BSwitch to the smallest model and compensate with a far more detailed system prompt describing the legal reasoning steps to follow
3. CSplit each 150K-token contract into 10K-token chunks and run them in parallel through a mid-tier model to cut per-request cost and turnaround time
4. DRoute straightforward merger agreements to a small model and complex ones to the most capable model, using contract length as the routing heuristic the team maintains

A retail bank's support chatbot must never give personalized investment advice. The current implementation prepends a 900-word policy paragraph, including this prohibition, to the start of every user message; behavioral rules, tone guidance, and the ban are interleaved in one block. In testing, the bot still occasionally recommends specific funds when users push back twice. What is the most effective structural fix?

1. AAdd a second reminder of the prohibition at the end of each user message, closest to the model's response, so the ban is the very last instruction the model reads before it answers
2. BMove the role definition and prohibitions into the system prompt as explicit, separately stated rules, including what the bot should say instead when asked for investment advice
3. CLower the temperature to 0 and cap top-p so the model samples its most probable tokens and adheres more consistently to the instructions it was given
4. DPost-process each response with a keyword filter that blocks replies mentioning fund names, tickers, or share classes before they reach the customer

A freight dispatch assistant has accumulated 42 tools, including warehouse admin functions and invoice voiding it has never legitimately needed. It now picks the wrong tool on about 8% of routine status updates, and a security review found that a prompt injection in a delivery note could trigger invoice voiding. What is the most effective remediation?

1. AAdd detailed system-prompt guidance describing exactly when each of the 42 tools should and should not be used, with worked examples for each
2. BReduce the agent's tool list to the tools its workflows actually require, and move admin-grade actions to a separately gated integration
3. CUpgrade the assistant to a larger, more capable model tier so tool-selection accuracy improves across the full 42-tool list without redesign
4. DRewrite each tool description to be longer and more distinctive so the model can discriminate more reliably among the 42 tools

A hospital assistant queries patient records through an electronic health record (EHR) integration that authenticates with a single service account holding organization-wide read access, chosen because provisioning per-user credentials was deemed too slow. During a pilot, a cardiology nurse received data on an oncology patient she has no treatment relationship with — a reportable access violation. Which change addresses the root cause?

1. AAdd a system-prompt rule, reinforced with few-shot examples, instructing the model to return records only for patients within the requesting user's own department
2. BPost-filter the assistant's responses with a second model pass that detects and redacts patient identifiers and out-of-department data before display
3. CPropagate the requesting clinician's identity to the EHR on each call so the EHR's own authorization rules scope each query to that user's permitted records
4. DDowngrade the shared service account from read-write to read-only and enable comprehensive audit logging with periodic access reviews on its queries

A fintech is rolling out an internal MCP server that exposes payment-operations tools — refund issuance, chargeback lookups, and account-hold toggles — to Claude clients used by three different teams. Today any authenticated employee client can invoke any tool, and the server trusts whatever tool name and arguments the client sends. Compliance has flagged the design before launch. Which TWO changes most directly close the authorization gaps?

Choose 2

1. ARequire an OAuth flow so every tool invocation carries a short-lived token scoped to the requesting employee's role
2. BHave the server verify per invocation that the calling user is entitled to the specific tool, rather than trusting the client's presented tool list
3. CMove the MCP server behind the corporate VPN so only managed devices on the internal network can reach its payment-operations endpoints
4. DLog each tool call with its arguments and results, retained for quarterly compliance review of refund and account-hold activity
5. ELower the client model's temperature so tool-invocation behavior is more deterministic and predictable across the three teams

A support-triage assistant scores 92% classification accuracy on a 500-ticket curated eval set, and the team declares it production-ready. Two weeks after launch, agents complain that responses take 11 seconds at p95 during peak hours, and one response echoed a customer's full credit card number from the ticket body. Leadership asks why evaluation didn't catch either problem. What was the core evaluation design flaw?

1. AThe 500-ticket eval set was too small a sample to produce a statistically reliable estimate of classification accuracy
2. BThe team used a hand-curated eval set instead of sampling tickets directly from live production traffic during representative peak hours
3. CThe evaluation measured only one dimension, omitting latency and safety/security metrics that production success actually depends on
4. DThe team should have used expert human graders rather than automated scoring so nuanced failure modes could be caught before launch

To build an eval set for a document Q&A system, a team samples 300 real user queries from production logs, filtering to conversations that received a thumbs-up so that reference answers are trustworthy. The system scores 96% on this set, but a manual audit of live traffic estimates only 78% of production answers are acceptable. What is the most likely explanation for the gap?

1. AFiltering to thumbs-up conversations selected for queries the system already handles well, excluding the failure modes behind the gap
2. BThe eval set of 300 queries is too small, so the 96% figure carries a wide confidence interval that overlaps the production estimate
3. CThe manual audit graded against stricter acceptability criteria than the automated eval did, so the two percentages are not directly comparable
4. DProduction traffic drifted in topic mix after the eval set was collected, making the sampled queries unrepresentative of current usage patterns

A contract-analysis system handles 2,000 queries per day. The eval framework runs deterministic checks (JSON schema, required citation fields) on every response and an LLM-judge for answer quality on a 10% sample. The team can afford expert human review of only 50 responses per week. To make the mixed-methodology framework most trustworthy, where should the scarce human review be focused?

1. AOn a random sample drawn from all weekly production responses, so the human scores form an unbiased estimate of overall answer quality
2. BOn the highest-traffic query categories, weighted by volume, since errors in those categories affect the largest number of users
3. COn responses that already failed the deterministic schema and citation checks, since those are confirmed defects worth documenting in depth
4. DOn responses where the LLM-judge verdict is borderline or conflicts with other signals, to calibrate the judge and expose blind spots

A regional bank deploys a Claude-powered assistant that answers customer questions about their own accounts. The team's only safety control is a system-prompt instruction: 'Never reveal information about accounts other than the authenticated user's.' During a red-team exercise, a tester uses a multi-turn injection to get the model to summarize another customer's transactions, because the retrieval layer fetches records by whatever account number appears in the conversation. Which change most directly fixes the vulnerability?

1. AStrengthen the system prompt with explicit refusal examples and restate that only the authenticated user's account number may drive record retrieval
2. BAdd an output classifier that scans responses for account numbers and identifiers not belonging to the authenticated user and blocks matches
3. CEnforce the verified caller's identity as a mandatory filter on data lookups so records outside that scope are excluded before generation
4. DLog conversations with their retrieval traces and route flagged sessions to the bank's fraud team for structured next-business-day review

A hospital's clinical documentation assistant summarizes physician dictations into structured notes. During validation, reviewers find that in about 2% of notes the summary includes a plausible-sounding medication dosage that was never dictated. The engineering team insists this is a retrieval bug, but the system performs no retrieval — it works only from the dictation transcript provided in the prompt. What is the most accurate characterization of this failure?

1. AA context-window overflow in which the model drops portions of the transcript under length pressure and reconstructs them from surrounding context
2. BA prompt-formatting defect that clearer delimiting of the transcript, separating instructions from dictated content, would resolve
3. CIntrinsic hallucination: the model generates content unsupported by its provided input, an inherent failure mode requiring output verification
4. DTraining-data contamination that can be corrected by fine-tuning the model on the hospital's own corpus of historical clinical notes

An insurance company uses an LLM to draft claim-denial letters. To satisfy regulators, every letter is routed to a claims adjuster who must click 'Approve' before it is sent. Adjusters each receive roughly 400 letters per day, approve 99.7% of them, and average 11 seconds per review. A regulator's audit later finds several denials that misstated policy terms and were approved. What is the fundamental flaw in this human-in-the-loop design?

1. AThe review step is placed after generation instead of before it, so adjusters cannot shape the drafts while policy terms are being applied
2. BThe review volume and approval pattern indicate rubber-stamping — the checkpoint exists procedurally but provides no meaningful validation
3. CAdjusters lack the legal training to evaluate policy language and should be replaced with attorney reviewers qualified to parse contract terms
4. DThe approval interface should require a short written justification for each approval so reviewers are forced to slow down and engage

A VP of Customer Experience opens a kickoff call by saying, "We've decided we need an AI chatbot on our support site — how fast can you build it?" She has budget approved and a launch date already communicated internally. The architect has not yet seen any support metrics or spoken to the support team. What should the architect do FIRST?

1. AAccept the chatbot direction and begin scoping channels, languages, escalation paths, and integration points so delivery stays aligned with the approved timeline
2. BSketch a candidate architecture and phased rollout plan for the chatbot so the VP has something concrete to react to in the next meeting
3. CAsk what business problem prompted the decision and what measurable outcome would define success before validating the chatbot as the solution
4. DRequest that the announced launch date be revisited, since committing to a solution this early in discovery carries avoidable delivery risk

During discovery for an AI document-processing system, the operations director insists the system must run fully automated end to end, while the compliance officer states that every output touching customer accounts must be human-reviewed. Both were interviewed separately and each believes their requirement is settled. The project sponsor asks the architect to "just write up the requirements." What should the architect do FIRST?

1. ABring both stakeholders together over the conflict and facilitate agreement on risk tolerance and decision criteria before finalizing requirements
2. BDocument both requirements exactly as stated and design a configurable architecture flexible enough to support either operating mode later
3. CAdopt the compliance officer's position as the binding requirement, since regulatory stakeholders take precedence over operational efficiency preferences
4. DPropose human review for a sampled percentage of account-touching outputs and full automation for the remainder as a balanced starting point

An architect recommends a mid-tier model for a claims-triage workload after testing shows it meets quality targets. In the review meeting, the CTO pushes back: "If we're doing this, I want the most capable model available — I don't want us looking cheap." Which response best communicates the recommendation?

1. AExplain that the mid-tier model's context window, token throughput, and rate limits are more than sufficient for the prompt sizes and request volumes involved in the claims-triage workload
2. BShow that both models meet the quality bar on the company's own test cases, that the choice saves projected monthly cost at lower latency, and name when it would be revisited
3. CAgree to launch on the most capable model to preserve the executive relationship, planning an internal downgrade to the mid-tier once usage data accumulates
4. DDefer the choice to the engineering team's technical judgment, since model selection is an implementation detail below the level of an executive review

A platform team rolls out Claude Code to 14 engineers and results diverge: some get output that follows the repo's conventions and test patterns, others get generic code that ignores the monorepo structure. Each engineer wrote their own personal notes file for the assistant, and several have none. The engineering manager proposes a mandatory two-hour prompt-writing training. What is the most effective way to make assistant output consistent across the team?

1. ARun the proposed two-hour training so each engineer learns to describe the repo's conventions and test patterns in their prompts
2. BHave the strongest prompter publish their personal notes file and settings in the team wiki with instructions for the other engineers to copy them locally
3. CCheck a project-level memory file (e.g., CLAUDE.md) and shared tool settings into the repository so every session starts from the same context
4. DStandardize the whole team on a single pinned model version so engineers receive identical completions for identical prompts

One month after a team adopted an AI coding agent under an 'allow everything' policy, two incidents hit: a session ran a destructive database reset while fixing a failing test, and shared transcripts were found to contain production API keys the agent had read from a local .env file. Which TWO changes most directly prevent both classes of incident from recurring?

Choose 2

1. ARotate the exposed keys immediately and circulate a detailed incident writeup so engineers monitor their agent sessions more closely going forward
2. BReplace the allow-everything policy with a checked-in permission configuration that auto-approves safe read and test commands but gates destructive or state-changing operations
3. CConfigure the agent for read-only access repo-wide so it can analyze code and propose changes but not execute commands or edit files on its own
4. DMove secrets out of files the agent can read and add deny rules for secret file patterns (e.g., .env, key material) in the shared tool configuration
5. ERequire a designated senior engineer to review the day's agent session transcripts in full at day's end before any of that work is merged