20 Solution Design & Architecture Practice Questions & Answers

1. 1. An insurance company tells the architect: "We want to use AI to handle claims." Before choosing any architecture, what should the architect do first?

   • A.Define the specific decision or output the system must produce, who consumes it, the daily volume, the latency tolerance, the accuracy bar, and the cost ceiling — and only then choose an architecture that fits.✓ Answer
   • B.Build an autonomous multi-agent prototype to show what is possible and let the requirements emerge from it.
   • C.Benchmark several model tiers on a sample of claims and keep whichever performs best.
   • D.Load every historical claim document into a vector store so retrieval is ready before design begins.

   Discovery precedes design. "Handle claims" is a business wish, not a requirement: until the job to be done, its consumer, the volume, the latency tolerance, the accuracy bar and the cost ceiling are known, there is no basis on which one architecture can be justified over another, and any tier, index or prototype chosen first has to be re-justified later. Every downstream trade-off decision is measured against these constraints.

   Source: Anthropic Claude certification (Architect – Professional), Domain 2 (Solution Design & Architecture): "translate business problems into Claude-based AI solutions"; Domain 5: "conduct structured discovery and requirement gathering"Report a problem with this question

2. 2. A retailer routes about 40,000 inbound support emails per day into 12 well-defined categories. The taxonomy is fixed, a large labeled history exists, and each email needs exactly one label. What architecture fits?

   • A.A workflow that makes twelve model calls per email, one yes/no question per category.
   • B.A multi-agent system with one specialist agent per category plus a coordinator.
   • C.An autonomous agent per email that can browse the CRM and decide the category on its own.
   • D.A single augmented model call per email that returns the category as structured output.✓ Answer

   The default is the simplest tier that meets the requirement. This task is a single, fully specifiable step — fixed taxonomy, one label, labeled examples available — so one augmented call with an enforced output schema satisfies it. Agentic and multi-agent designs add latency, cost and non-determinism with no accuracy gain on a closed classification, and twelve calls per email multiplies token spend roughly twelvefold at 40,000 emails a day.

   Source: Anthropic Claude certification (Architect – Professional), Domain 2: "select appropriate architectural patterns (workflow, agentic, augmented LLM)"Report a problem with this question

3. 3. An accounts-payable team wants to validate supplier invoices: the line items must sum to the stated total, tax must fall within an allowed range, and the vendor must exist in the ERP. Invoices arrive as scanned PDFs. What should the architect design?

   • A.A single model call that reads the invoice and performs the arithmetic using chain-of-thought reasoning.
   • B.An agent with a calculator tool and an ERP lookup tool that decides how to validate each invoice.
   • C.Deterministic code for the arithmetic check, the tax-range check and the ERP vendor lookup, with a model call used only to extract fields from the unstructured PDF.✓ Answer
   • D.A retrieval pipeline over previously approved invoices so the model can judge whether the totals look normal.

   Recognizing where an LLM does not belong is part of solution design. Summation, range comparison and a reference-data lookup have exact, verifiable answers that deterministic code produces correctly every time, cheaply and auditably; a model introduces non-determinism into checks that finance must be able to reproduce. The model earns its place only at the one step that is genuinely a language problem — turning an unstructured scanned document into typed fields.

   Source: Anthropic Claude certification (Architect – Professional), Domain 2: "apply decomposition techniques" and "translate business problems into Claude-based AI solutions"Report a problem with this question

4. 4. A contract-review pipeline has five stages: extract text from the PDF, identify clauses, rate each clause against a written risk playbook, route the contract to a reviewer queue by risk band, and start an SLA timer. Which decomposition is correct?

   • A.Use a model for queue routing and the SLA timer, because deciding urgency requires judgment.
   • B.Hand the whole document to one autonomous agent and let it carry out all five stages.
   • C.Make every stage a model call so the pipeline behaves consistently from end to end.
   • D.Use model calls for clause identification and risk rating, and deterministic code for text extraction, queue routing and the SLA timer.✓ Answer

   Decomposition assigns each stage to the cheapest mechanism that can perform it correctly. Clause identification and playbook-based rating are language-understanding and judgment tasks that need a model; extraction, routing by a risk band that has already been computed, and timers are mechanical rules that code executes deterministically. Making rule-based stages model calls adds latency, cost and non-determinism to steps that must be reproducible and auditable.

   Source: Anthropic Claude certification (Architect – Professional), Domain 2: "apply decomposition techniques"; "design end-to-end architectures (input → processing → output → feedback loops)"Report a problem with this question

5. 5. An architect is explaining to a skeptical engineering lead why an internal policy assistant should retrieve source documents at request time rather than rely on what the model already knows. What is the strongest reason?

   • A.Retrieval makes the system deterministic, so an identical question always produces identical wording.
   • B.Retrieval removes the possibility of hallucination entirely.
   • C.Organization-specific policies change over time and are not reliably present in the model's weights; retrieval supplies the current authoritative text at request time and lets the answer cite the passage it came from.✓ Answer
   • D.Retrieval reduces total token usage compared with relying on the model's own knowledge.

   Grounding works because the authoritative text is supplied at request time instead of recalled from weights that were fixed before the document existed and may since have been revised; it also makes each claim traceable to a citable passage, which is what an internal policy answer must be. Retrieval reduces hallucination without eliminating it, adds input tokens rather than saving them, and does not make a probabilistic system deterministic.

   Source: Anthropic Claude certification (Architect – Professional), Domain 1 (Integration): "design a RAG pipeline"; Domain 2: "translate business problems into Claude-based AI solutions"Report a problem with this question

6. 6. In a grounded question-answering system, the retrieval step sometimes returns no passage that actually addresses the user's question. How should the architecture handle this case?

   • A.Let the model answer from its general knowledge whenever retrieval comes back empty.
   • B.Return the top one hundred chunks so that the answer is somewhere in the context.
   • C.Lower the similarity threshold so that some passage is always returned.
   • D.Treat "insufficient evidence" as a first-class outcome: return an explicit no-answer with a defined escalation path, and log the query as a coverage gap.✓ Answer

   A grounded system must be able to say it has no basis to answer; forcing an answer converts a known coverage gap into a confident wrong answer, which is worse than no answer because the user cannot tell the difference. Loosening the threshold or flooding the context returns more irrelevant material, not more correct material, and the logged gap is the signal that tells you which documents to add.

   Source: Anthropic Claude certification (Architect – Professional), Domain 1: "apply retrieval strategies matched to data shape and query pattern"; Domain 4: failure modes of LLM systemsReport a problem with this question

7. 7. A nightly job classifies 200,000 documents and must finish before the morning batch run. During design review, the team is asked what happens when the model endpoint is slow or briefly unavailable. Which design is correct?

   • A.Convert the job to synchronous per-document requests so failures become visible sooner.
   • B.Fail the entire batch on the first error and page the on-call engineer.
   • C.Retry the failed request immediately in a tight loop until it succeeds.
   • D.Process through a durable queue with bounded retries and exponential backoff, a dead-letter path for repeatedly failing items, and a documented manual fallback, so the job resumes rather than restarting.✓ Answer

   Failure paths are first-class parts of the design, not exceptions. For non-interactive work, slowness and transient unavailability are expected conditions: a durable queue preserves progress so the job resumes at the failure point, bounded retries with backoff avoid amplifying an overloaded dependency, a dead-letter path isolates poison items so one bad document cannot stall 200,000, and the documented fallback keeps the business process running while engineering responds.

   Source: Anthropic Claude certification (Architect – Professional), Domain 2: "design end-to-end architectures"; Domain 4: "identify risks, limitations and failure modes of LLM systems"Report a problem with this question

8. 8. A model-assisted system drafts customer refund approvals ranging from a few dollars to amounts that would require a finance write-off. Where should human review sit?

   • A.Gate by consequence: auto-execute small, easily reversible refunds and require human approval for large, irreversible or disputed ones.✓ Answer
   • B.Skip human review and rely on dashboards and alerting to catch mistakes after the fact.
   • C.Have a human spot-check a fixed random 5% of refunds regardless of amount or reversibility.
   • D.Require a human to approve every drafted refund, regardless of amount.

   Human-in-the-loop should be placed by irreversibility and blast radius, so that review effort is spent where a wrong decision is expensive and hard to undo. Reviewing everything destroys the efficiency the system was built for, reviewing nothing leaves large irreversible errors unguarded, and a flat random sample is a monitoring technique that detects errors after they have already been executed rather than gating high-consequence actions before they happen.

   Source: Anthropic Claude certification (Architect – Professional), Domain 4 (Governance, Safety & Risk): "apply human-in-the-loop validation strategies"Report a problem with this question

9. 9. A business sponsor asks for a system that is maximally accurate, responds instantly, costs almost nothing per request, and runs with no human oversight. What should the architect do?

   • A.Choose the cheapest possible design now, since cost is the only constraint with a hard number attached to it.
   • B.Explain that these properties trade against one another, ask the sponsor to rank them, and record an explicit target for each — a percentile latency, an accuracy rate measured on a fixed test set, and a cost per thousand items — so the design can optimize the binding constraint.✓ Answer
   • C.Optimize accuracy first and treat latency and cost as tuning work after launch.
   • D.Agree to all four and plan to reach them through prompt engineering.

   Accuracy, latency, cost and autonomy trade against one another: raising one generally costs another, so no design satisfies all four simultaneously and an architect who promises otherwise will miss the SLA. The architect's job is to surface the trade-off, obtain a ranking, and convert each requirement into something measurable on a non-deterministic system — latency as a percentile rather than an average, accuracy as an evaluated rate on a fixed set rather than an absolute guarantee.

   Source: Anthropic Claude certification (Architect – Professional), Domain 2: "align solutions to business value pillars (efficiency, transformation, productivity, cost, performance SLAs)"; Domain 5: "manage stakeholder feedback loops and expectation alignment including SLAs"Report a problem with this question

10. 10. The output of a model-based step will be consumed automatically by a billing system that cannot tolerate malformed input. How should the architect design the boundary?

    • A.Add a sentence to the prompt asking the model to reply in JSON and parse whatever comes back.
    • B.Let the billing system parse the free-text response with regular expressions.
    • C.Define an explicit output schema, enforce structured output, validate every response against the schema before it reaches billing, and define what happens when validation fails.✓ Answer
    • D.Send the free-text response to a second model call that reformats it, and pass that result downstream.

    An interface to a downstream system needs an enforced contract, not a request. A prompt instruction is a preference that a non-deterministic model can violate, regex parsing of free text breaks silently the moment phrasing shifts, and a reformatting call adds a second non-deterministic step without adding any guarantee. Schema enforcement plus validation at the boundary, with an explicit failure route, is what makes the downstream system safe to automate.

    Source: Anthropic Claude certification (Architect – Professional), Domain 2: "design end-to-end architectures (input → processing → output → feedback loops)"; Domain 4: output filtering and structured-output enforcementReport a problem with this question

11. 11. An automated process issues refunds through a payment provider's API. Network timeouts occasionally leave the caller unsure whether the refund was created. What should the design include?

    • A.Add a human confirmation prompt before every retry so a person decides.
    • B.Have the caller generate an idempotency key from the business event and send it with every attempt, so a retried request is recognized as a duplicate and returns the original result instead of creating a second refund.✓ Answer
    • C.Disable retries entirely so a refund can never be issued twice.
    • D.Instruct the model in its prompt to remember whether it has already issued the refund.

    Any call that crosses a network can time out after the work has actually been done, so operations touching external systems must be idempotent by design. A caller-generated key derived from the business event lets the provider recognize the second attempt as the same operation and return the original outcome. Disabling retries trades duplicate refunds for silently lost ones, a model's recollection is not a durable deduplication mechanism, and a human confirmation on every retry is a compensating control that does not make the operation safe.

    Source: Anthropic Claude certification (Architect – Professional), Domain 2: "design end-to-end architectures"; Domain 4: failure modes and safe handling of irreversible external actionsReport a problem with this question

12. 12. A design proposal is being reviewed for a system expected to process a high daily volume of long documents. The business has stated a monthly cost ceiling. What belongs in the design review?

    • A.Model expected volume × tokens per request (input and output) × unit cost for each route, plus peak concurrency against throughput limits, and compare the result with the stated cost ceiling before committing to the design.✓ Answer
    • B.Assume per-request cost is negligible and spend the review on accuracy instead.
    • C.Use the least expensive option on every route so that cost cannot become a problem.
    • D.Build the system first and measure real cost in production, since pre-build estimates are unreliable.

    Cost and capacity are design inputs, not post-launch discoveries: token volume is driven by architectural choices such as how much context is retrieved, how many stages call the model and how often, so those choices must be priced while they are still cheap to change. Peak concurrency must also be checked against throughput limits, because a design that meets the cost ceiling on average can still fail its latency SLA at peak.

    Source: Anthropic Claude certification (Architect – Professional), Domain 2: "align solutions to business value pillars (… cost, performance SLAs)"Report a problem with this question

13. 13. An organization plans to replace a manual triage process that 30 specialists perform today. The model-based system passed offline evaluation. What is the right rollout design?

    • A.Shadow-run the system on live traffic without acting on its outputs, compare its decisions with the human ones on pre-agreed metrics, then pilot on a narrow slice with a defined rollback path.✓ Answer
    • B.Launch to 100% of traffic behind a feature flag so it can be switched off if complaints arrive.
    • C.Set a full cut-over date and train the team in advance so everyone is ready.
    • D.Extend offline evaluation until the scores are high enough, then launch to everyone at once.

    Offline evaluation shows performance on curated data; shadow-running measures the system against the incumbent process on the real traffic distribution at zero customer risk, which is the only way to discover the inputs the eval set never contained. Piloting a narrow slice afterwards limits blast radius while the failure modes are still being learned, and a rollback path must exist before, not after, exposure. A feature flag at full traffic is a switch, not an incremental rollout.

    Source: Anthropic Claude certification (Architect – Professional), Domain 5 (Stakeholder Communication & Lifecycle Management): "support lifecycle phases (discovery, design, handoff, monitoring, iteration)"Report a problem with this question

14. 14. A proof of concept has been running for several weeks. Which condition best justifies committing to production investment?

    • A.The prototype ran for a week without throwing any errors.
    • B.The proof of concept met success criteria that were agreed before it began, on a dataset representative of real traffic including edge cases, and the remaining unknowns are engineering and operations rather than feasibility.✓ Answer
    • C.The demonstration impressed the executive sponsor, who has asked for a production date.
    • D.The model handled the three hardest examples the team could think of.

    A proof of concept exists to retire feasibility risk against criteria fixed in advance; criteria invented afterwards can always be made to fit whatever the prototype happened to do. Representative data including edge cases is what distinguishes a result that will hold at production distribution from a demo. Sponsor enthusiasm, an absence of crashes and a handful of self-selected hard examples are not evidence about the traffic the system will actually see.

    Source: Anthropic Claude certification (Architect – Professional), Domain 5: "support lifecycle phases (discovery, design, handoff, monitoring, iteration)"; Domain 3: evaluation datasets built from representative and adversarial casesReport a problem with this question

15. 15. A nightly management report is produced from five steps that are always the same and always run in the same order: pull metrics, detect anomalies against thresholds, summarize each anomaly, assemble the narrative, and email the result. What should the architect design?

    • A.A multi-agent system with a planner that decides the steps each night.
    • B.A deterministic workflow in which code orchestrates the five steps and calls the model only for the language work, so each step can be tested, retried and monitored independently.✓ Answer
    • C.A single autonomous agent given every tool and told to produce the report.
    • D.One very large model call that receives all the raw data and produces the entire report at once.

    When the steps and their order are known in advance, code-controlled orchestration is the correct tier: it is predictable, individually testable, cheaper, and each stage can be retried without redoing the rest. An agent's value is deciding what to do when the task cannot be specified up front, which is not the case here, and a single monolithic call gives up per-step observability and makes any failure a total failure.

    Source: Anthropic Claude certification (Architect – Professional), Domain 2: "select appropriate architectural patterns (workflow, agentic, augmented LLM)"Report a problem with this question

16. 16. A design uses a coordinator that delegates three genuinely independent research tracks to subagents working in parallel. What must the architecture account for?

    • A.Each delegated brief must stand alone — objective, constraints, the inputs required and the expected output format — because a subagent does not share the coordinator's conversation, and results must be written to shared storage to be visible to the others.✓ Answer
    • B.Subagents should call back to the coordinator whenever they discover they are missing context.
    • C.Every subagent should receive the full tool set so that no track can be blocked.
    • D.Subagents automatically inherit the coordinator's conversation history, so a short instruction is enough.

    Delegation does not carry conversation state: a subagent starts from the brief it is given, so anything it needs — objective, constraints, inputs, output format — must be stated explicitly or placed in shared storage, and its findings are invisible to siblings unless written there. Relying on inherited context produces silent gaps, round-tripping for missing context destroys the parallelism the design was chosen for, and handing every subagent every tool is capability bloat that degrades tool selection and widens the security surface.

    Source: Anthropic Claude certification (Architect – Professional), Domain 2: "design multi-agent systems and orchestration strategies"Report a problem with this question

17. 17. An architecture diagram shows input → processing → output for a document-classification service, with monitoring attached. A reviewer says a required element of the end-to-end design is missing. What is it?

    • A.A second output format so that more downstream systems can consume the result.
    • B.A feedback path that captures outcome signal — human corrections, downstream acceptance or rejection, and telemetry — and routes it back into the evaluation set and the prompt and retrieval iteration cycle.✓ Answer
    • C.A larger model on the processing step to raise baseline quality.
    • D.A caching layer placed in front of the processing step.

    An end-to-end LLM architecture is input → processing → output → feedback loop; without a path that captures how outputs actually fared, the system cannot improve after launch and quality drift goes unnoticed until users complain. Monitoring records what happened, but it is the feedback path that converts corrections and downstream outcomes into new evaluation cases and prompt or retrieval changes. The other options are optimizations, not missing structure.

    Source: Anthropic Claude certification (Architect – Professional), Domain 2: "design end-to-end architectures (input → processing → output → feedback loops)"Report a problem with this question

18. 18. One product surface is an interactive assistant where users wait for a reply; another enriches five million catalog records overnight. Both use the same underlying capability. How should the architect design them?

    • A.Make both paths asynchronous so that the architecture stays uniform.
    • B.Keep one design and simply raise concurrency limits at night.
    • C.Use one synchronous design for both and set a long timeout on the bulk job.
    • D.Design two distinct paths: the interactive one optimizes perceived and tail latency through streaming, tight context and route-appropriate processing, while the overnight one runs asynchronously in batches and optimizes throughput and cost per item.✓ Answer

    Latency tolerance is a requirement that drives architecture, so two surfaces with opposite tolerances warrant different designs even on shared capability. The interactive path is judged on tail and perceived latency, which favors streaming and minimal context; the bulk path has no waiting user, so it should be queued and batched to maximize throughput per unit cost. Forcing one shape on both either wastes money on the batch path or violates the interactive SLA.

    Source: Anthropic Claude certification (Architect – Professional), Domain 2: "align solutions to business value pillars (… cost, performance SLAs)"; "design end-to-end architectures"Report a problem with this question

19. 19. During design review of a customer-facing assistant, someone asks what happens when the model declines to answer a request or determines it is out of scope. What is the correct design decision?

    • A.Add prompt instructions telling the model never to refuse or say that it cannot answer.
    • B.Treat refusal or abstention as a defined output state with its own route: show the user a clear message, hand the case to a human or an alternate process, and log it so the pattern can be reviewed.✓ Answer
    • C.Retry with rephrased inputs until the model produces an answer.
    • D.Treat refusals as rare noise and let the existing error handler surface a generic failure message.

    Refusal and abstention are expected behaviors of the system, not errors, so they belong in the design as a named output state with a route that still serves the customer. Instructing the model never to refuse pressures it toward answering cases it should not, retrying with rephrasings is an attempt to defeat a safeguard and produces inconsistent behavior, and a generic error message strands the customer while hiding the signal that would tell the team what scope is missing.

    Source: Anthropic Claude certification (Architect – Professional), Domain 2: "design end-to-end architectures"; Domain 4: "identify risks, limitations and failure modes of LLM systems"Report a problem with this question

20. 20. A grounded assistant currently sends the fifty highest-scoring passages into context for every question. Answer quality is mediocre and cost is high. What is the correct design reasoning?

    • A.Always retrieve the maximum number of passages the context allows, so that the answer is guaranteed to be present somewhere.
    • B.Retrieve the smallest set that reliably answers the query pattern — using metadata filters and reranking — because loosely relevant passages dilute the signal and add cost and latency, and measure answer quality as a function of how many passages are retrieved.✓ Answer
    • C.The number of retrieved passages affects cost but not answer quality, so tune it purely against the budget.
    • D.Keep retrieving everything and add a system-prompt instruction telling the model to ignore irrelevant passages.

    Retrieval quality is not monotonic in quantity: retrieving too much is a distinct failure mode in which weakly related passages compete with the correct one and degrade the answer, on top of raising tokens, cost and latency. The fix is architectural — narrow the candidate set with metadata filters and reranking so the retrieved set matches the query pattern — and the right size is found empirically by measuring quality against passage count, not by adding an instruction that asks the model to compensate.

    Source: Anthropic Claude certification (Architect – Professional), Domain 1: "apply retrieval strategies matched to data shape and query pattern"; Domain 2: decomposition and end-to-end designReport a problem with this question

Practice questions based on the official Claude Certified Architect – Professional (CCAR-P) exam guide and Anthropic's public documentation. This is an independent study tool, not affiliated with or endorsed by Anthropic, and does not grant certification. The real exam is 63 questions, 120 minutes, passing at a scaled 720/1000, delivered via Pearson VUE ($175). Official certification page →