Solution Design & Architecture

Question 1 · 1.1

An insurer’s claims director asks for “an autonomous AI agent to handle claims end to end”. Discovery shows the main delay is adjusters retyping data from scanned repair estimates into the claims system; decisions themselves are quick.

What should the architect propose first?

• AAn autonomous agent that reads the estimate, decides the claim and pays it.
• BFine-tune a model on past claims before designing anything else.
• CClaude extraction of estimate fields to a schema, validated in code, with adjusters confirming.
• DA multi-agent system with separate agents for intake, fraud and payment.

Question 2 · 1.1

A retailer wants to reclassify 200,000 product listings into a new category tree that changes every season. It has almost no labelled examples, and listings are written in five languages.

Which factors in this scenario favour Claude over training a traditional classifier? (Select 2.)

• AThe category tree changes every season.
• BThere are 200,000 listings to process.
• CThere are almost no labelled examples.
• DThe output is a category label.
• EThe retailer already uses a cloud provider.

Question 3 · 1.1

A public-sector benefits agency plans a Claude assistant that answers caseworkers’ questions about eligibility policy. The project charter states the goal as “give caseworkers better answers, faster”.

Which success criteria best support the design?

• AMost caseworkers rate the assistant as helpful in a launch survey.
• BEvery question is answered by the most capable Claude model available.
• CThe assistant never gives a wrong or incomplete policy answer.
• D95% match expert answers, each cited to a policy clause, in under 10 seconds.

Question 4 · 1.1

A business process involves calculating loan repayment schedules from a fixed formula, then writing a plain-language explanation for each customer. How should the work be split?

• AClaude does both steps, since it can do arithmetic and write explanations.
• BCode computes the schedule; Claude writes the explanation from the computed figures.
• CAn agent decides at run time which steps need a calculation tool.
• DCode does both steps using a fixed template, with no model involved.

Question 5 · 1.2

A bank’s trade-finance team uses Claude to extract fields from letters of credit and write them straight into a booking system. About 2% of bookings fail because dates arrive in mixed formats and some amounts include currency words.

Which change most directly fixes the failure?

• AMove to a larger Claude model so the extraction is more accurate.
• BConstrain output to a typed schema and validate values in code before booking.
• CAdd “always use ISO dates” in capitals at the top of the system prompt.
• DSend every extracted letter of credit to a human for full re-keying.

Question 6 · 1.2

A hospital network’s Claude assistant drafts replies to patient portal messages for nurses to approve. Six months after launch, nurses say drafts feel “worse”, but the team cannot say what changed or when.

What should the architecture have included? (Select 2.)

• ALogging each draft with the nurse’s final version and any rejection reason.
• BA larger context window so more of each patient’s history fits in.
• CA regression eval set, grown from flagged drafts and run before every change.
• DRemoving the nurse approval step to collect more unfiltered outputs.
• ESwitching model every quarter to the newest release.

Question 7 · 1.2

A public-sector agency wants Claude to summarise the 30,000 consultation responses it receives after each policy proposal. Analysts read the summaries the following week.

Which delivery design best fits?

• AReal-time streaming calls so analysts see summaries appear as they are generated.
• BAn agent that decides for itself which responses to summarise and when.
• CSubmit responses through the Message Batches API and load results for analysts.
• DProcess responses one by one on a web server as each one arrives.

Question 8 · 1.2

In an end-to-end Claude design, which responsibility belongs in deterministic code or a system of record rather than in the model?

• ADeciding which of fourteen teams should own an incoming complaint.
• BSummarising a clinician’s notes into a discharge letter draft.
• CDrafting a polite explanation of a declined card payment.
• DCalculating the refund due under a fixed tariff after a flight delay.

Question 9 · 1.3

A retailer receives supplier invoices in varied layouts. Every invoice goes through the same steps: extract fields, match to a purchase order in the ERP, flag mismatches, and post matched invoices.

Which architecture is most appropriate?

• AAn autonomous agent with ERP tools that decides how to process each invoice.
• BA multi-agent system with an agent per step coordinating through a supervisor.
• CA prompt-chaining workflow: Claude extracts, code matches and gates, Claude explains mismatches.
• DA single call that returns the posted journal entry directly.

Question 10 · 1.3

An internal IT team wants help resolving employee incidents. Tickets range from password resets to intermittent network faults whose cause takes several rounds of log searching, config checks and tests to find.

For the hard, open-ended incidents, which pattern fits best?

• AAn agent with read-only diagnostic tools, a turn limit and engineer approval before any change.
• BA fixed prompt chain that always checks logs, then config, then network.
• CA single augmented call that reads the ticket and proposes a fix.
• DA parallel vote of five calls that each guess the root cause from the ticket.

Question 11 · 1.3

Which situations are signals for the evaluator-optimizer workflow? (Select 2.)

• AThere are clear criteria a second call can check a draft against.
• BThe number of steps cannot be predicted in advance.
• CHuman feedback on drafts has been shown to improve them measurably.
• DInputs fall into distinct categories that need different handling.
• ELatency must be as low as possible for a live customer.

Question 12 · 1.3

A bank’s architecture review board rejects a proposal for an agent that answers customers’ questions about fee schedules. The questions are varied, but every answer comes from one published fee document.

What is the strongest basis for the board’s decision?

• AAgents cannot be used in regulated industries such as banking.
• BAgents always cost more than multi-agent systems, so neither should be used.
• CThe Agent SDK only supports coding tasks, not customer-facing ones.
• DA single retrieval-grounded call meets the need at lower cost and risk.

Question 13 · 1.4

A pharmaceutical company’s regulatory team wants a system that, for a new drug candidate, surveys approval precedents across several agencies and many therapeutic areas. A single agent’s runs are thorough on the first agency and shallow on the rest, as its context fills.

Which design best addresses the problem?

• AA larger context window and an instruction to give every agency equal attention.
• BA lead agent that briefs one subagent per agency, each returning a condensed, sourced summary.
• CA chain of planner, searcher, summariser and reviewer agents passing work down a line.
• DA shared-state mesh where agents write findings to a common store with no coordinator.

Question 14 · 1.4

An orchestrator agent for a bank’s internal audit team occasionally spawns dozens of subagents on broad requests, and one run cost far more than budgeted. The prompt already says “use no more than five subagents”.

What is the most appropriate change?

• AStrengthen the prompt wording to insist on five subagents and warn about cost.
• BRemove the orchestrator and have one agent handle every audit request sequentially.
• CMove every subagent to the most capable model so fewer are needed per task.
• DSet enforced limits on concurrency, nesting depth and total spend for the query.

Question 15 · 1.4

Which situations are good reasons to introduce a second agent? (Select 2.)

• AA lookup returns thousands of tokens that the rest of the task does not need.
• BThe business process has four stages owned by four different departments.
• COutputs must be checked against a written set of criteria before release.
• DEach step needs to see every earlier decision in full to do its job.
• EThe stakeholder wants the system to sound more advanced in the demo.

Question 16 · 1.4

A logistics firm’s orchestrator delegates carrier-rate research to subagents. Results come back inconsistent: two subagents research the same carriers, and another returns rates in a different currency and format.

What should the architect change first?

• AAdd a message bus so subagents can see each other’s work as they go.
• BReplace the subagents with a single agent that researches every carrier in turn.
• CHave the lead write each subagent a brief with scope, exclusions and a fixed return format.
• DAsk each subagent to return its full transcript so the lead can reconcile the differences.

Question 17 · 1.5

A retailer’s customer-insight team feeds 12,000 free-text survey responses into a single request and asks for the top themes with counts. The counts are inconsistent between runs and later responses are under-represented.

Which decomposition is most appropriate?

• ATag each batch of responses with themes in separate calls, then count and roll up the tags in code.
• BKeep one request but instruct Claude to give equal weight to every response.
• CBuild a planner, analyst, statistician and writer agent that pass the survey along.
• DSample 200 responses at random and report their themes as the result.

Question 18 · 1.5

A bank’s credit team proposes a seven-step chain for small-business loan memos: read statements, summarise, extract ratios, compute ratios, assess risk, write the memo, check tone. Latency is high and the risk assessment often contradicts figures in the statements.

What is the best change to the decomposition?

• AAdd an eighth step that re-reads all earlier outputs and reconciles contradictions.
• BCollapse everything into one prompt so the model sees the statements throughout.
• CExtract figures to a schema, compute ratios in code, assess risk from statements and ratios.
• DRun all seven steps in parallel so that total latency drops.

Question 19 · 1.5

Which are good signals that a piece of a problem should be split out as its own step? (Select 2.)

• AIt needs different context from the rest of the task.
• BAn approval or audit point must see its output before work continues.
• CA different department performs that part of the work today.
• DIt is short and closely tied to the step before it.
• EThe project plan would look more thorough with more steps.

Question 20 · 1.5

An insurer asks an agent to rewrite 600 policy-wording templates to a new plain-language standard. After one long session the agent reports “all templates updated”, but spot checks show fewer than a hundred were changed.

What should the architect do?

• ARaise the session’s turn limit so the agent has room to finish all 600.
• BAsk the agent to double-check its work before claiming completion.
• CAssign the rewrite to a human team and use Claude only for spell-checking.
• DKeep a structured list of templates with a pass flag and a check; do one or a few per session.

Question 21 · 1.6

A wealth-management firm’s head of strategy sponsors a Claude project so that advisers can offer a personalised portfolio review to clients with smaller balances, a segment the firm has never served profitably.

Which primary success metric best reflects the value this sponsor is buying?

• AAverage minutes an adviser spends preparing each existing client review.
• BRevenue and retention from the newly served smaller-balance segment.
• CToken cost per portfolio review generated by Claude.
• DAdviser satisfaction with the drafting assistant in a survey.

Question 22 · 1.6

An architect’s business case for a claims-summary assistant cites an industry report that AI makes knowledge workers 80% faster and multiplies it by the claims team’s salary cost. The finance director rejects it.

What is the best way to strengthen the case?

• AReplace the industry figure with a more conservative published figure, such as 30%.
• BAdd a sensitivity table showing savings at 40%, 60% and 80% speed-ups.
• CReport the token cost to show the project is cheap regardless of benefit.
• DBaseline handling time and errors now, pilot, and measure the same, including review time.

Question 23 · 1.6

A telecom operator wants an agent-assist tool that suggests answers to call-centre staff during live calls. Operations requires suggestions to start appearing within about two seconds, and the tool must handle Monday-morning peaks without failing.

Which design choices best serve these SLAs? (Select 2.)

• AStream responses and use a smaller model with a capped output length.
• BSend suggestions through the Message Batches API to reduce cost.
• CSize rate limits for the peak, cache the stable knowledge prompt, and add a fallback.
• DAdd an instruction telling Claude to respond as quickly as possible.
• ERoute every call to the most capable model to avoid retries.

Question 24 · 1.6

A regional hospital group plans an overnight job that drafts discharge-summary coding suggestions for coders to review the next morning. About 8,000 summaries a night share the same long coding-guidelines prompt.

Which change most reduces running cost without harming the service commitment?

• ARun the job in real time throughout the day to spread the load.
• BSubmit through the Batch API with the shared guidelines as a cached prefix.
• CSplit each summary across a lead agent and four subagents.
• DRemove the coding guidelines from the prompt to shorten it.