Question 100
A document assistant saves 600 analyst hours annually at $40 per hour. The annual subscription costs $9,000, and executives require at least $10,000 in annual net benefit. Ignoring one-time setup, what is the maximum annual human-review expense the project can support?
- $15,000, treating gross savings minus the target as the reviewer budget and omitting subscription expense.
- $14,000, treating total operating expense capacity as though all of it could be paid to reviewers.
- $5,000 per year.
- $24,000, treating all gross labor savings as available for review expense despite the net-benefit target.
Explanation
Correct Answer
C:Gross annual value is 600×$40 = $24,000. Required net benefit is $10,000 and subscription $9,000; remaining reviewer budget is $5,000.
Incorrect Answers
A:The $24,000 gross labor benefit must first fund the $9,000 subscription and leave $10,000 of required net benefit; $15,000 ignores licensing.
B:The total allowable operating spend is $14,000, but $9,000 is already committed to the subscription.
D:Using every dollar of savings for review expense leaves no room for licensing or the required retained net benefit.
Question 101
A custom agent costs $40,000 to build and $20,000 per operating year. A prebuilt option has $5,000 setup and $32,000 annual licensing. Assuming constant costs and no discounting, by the end of which operating year does the custom agent first become cheaper in cumulative TCO?
- Year 1, because the custom agent has the lower annual running cost.
- Year 2, because the first two years of annual license savings already recover the higher build cost.
- Year 4, because the prebuilt agent remains cheaper throughout the first three operating years.
- Year 3, when the cumulative annual difference first offsets the higher custom setup cost.
Explanation
Correct Answer
D:The custom option begins $35,000 behind on setup but saves $12,000 each operating year. The first whole year with 12,000×years exceeding $35,000 is year 3.
Incorrect Answers
A:In year 1 the custom total is $60,000 and the prebuilt total $37,000; the cheaper annual run rate does not erase the setup gap immediately.
B:After year 2, custom totals $80,000 and prebuilt $69,000, leaving custom $11,000 more expensive.
C:By year 3, custom totals $100,000 versus prebuilt $101,000; the crossover has already happened and is not delayed until year 4.
Question 102
A company needs an internal document assistant. It can add a supported extension to existing Microsoft 365 Copilot or build a new secure search application. What should drive the comparison?
- The proven permission, feature and maintenance tradeoffs plus total cost for both alternatives.
- Whether the extension uses a newer user interface than the custom portal, without testing permission integration or ownership cost.
- Choose custom development because vendor features are never updated.
- Assume extending Copilot includes no security work.
Explanation
Correct Answer
A:Extending can reuse identity and work context, but only validated capability and economics determine the right choice.
Incorrect Answers
B:Interface freshness is only a preference; security scope and total lifecycle expense determine whether a parallel app is justified.
C:Unverified assumptions about vendor evolution cannot justify build costs.
D:Extensions still require access control and evaluation.
Question 103
A tuned classification model improves accuracy 1% but requires expensive labeling and monthly retraining. How should a business architect decide?
- Declare the largest model best without measuring downstream effect.
- Compare training loss alone and ignore real claims processing outcomes.
- Treat annotation cost as zero because it is not a cloud invoice.
- Compare the value of corrected cases with annotation, training, deployment and monitoring expenses.
Explanation
Correct Answer
D:A marginal accuracy gain is valuable only if its risk-adjusted business benefit exceeds its lifecycle costs.
Incorrect Answers
A:Model size does not establish economic value.
B:Training loss may not reflect deployment error distribution.
C:Human labeling consumes budget even when not billed by a cloud provider.
Question 104
An enterprise router sends 80% of easy AI requests to a small model and difficult requests to an expert model. What must its routing policy validate besides cost?
- Route requests by a static estimate of complexity, but never evaluate whether the lower-cost route meets error and safety limits.
- Quality and risk thresholds so simple-path errors are escalated rather than silently accepted.
- Use customer sentiment toward compact-model replies as the routing threshold without checking task errors or high-risk outcomes.
- Set a fixed target of 80% compact-model traffic and call routing successful even if high-risk requests are incorrectly downgraded.
Explanation
Correct Answer
B:Cost savings are invalid if routing makes sensitive or hard requests fail accepted criteria.
Incorrect Answers
A:A routing rule cannot establish safety without outcome checks and escalation for requests misclassified as easy.
C:A pleasant reply can be materially incorrect; route selection must consider consequence and demonstrated answer validity.
D:The intended traffic share measures cost allocation, not safe classification or correctness; the policy must detect escalations and meet workload-specific quality limits.
Question 105
A contract-analysis service sends a typical 3,000 input tokens and 500 output tokens per request. Finance modeled only 500 tokens. Which error must be corrected?
- Budget both input and output token usage, plus model-specific rates and any retrieval overhead.
- Price only output tokens because input text is always free.
- Assume all future documents have exactly the same token profile as one sample.
- Estimate inference spend using request count and typical output size, omitting the large retrieved input context.
Explanation
Correct Answer
A:Cost models need the actual charged token dimensions and workload variability.
Incorrect Answers
B:Input may represent a substantial share of consumption and charges.
C:One request does not describe the distribution of document sizes.
D:Output-only estimates ignore the 3,000 input tokens and any retrieval/context overhead, even when request count is accurate.
Question 106
A government department wants ROI measured as fewer incorrect benefit letters rather than time saved. Which primary KPI fits?
- Count of agent responses without checking recipient correctness.
- Use average generation-token cost per letter as the primary outcome measure without verifying legal accuracy or resulting appeals.
- Measure only claimant satisfaction with the letter’s explanation without reconciling its entitlement decision and downstream corrections.
- Verified reduction in wrong-letter rate and downstream appeal or remediation cost.
Explanation
Correct Answer
D:Error reduction can be monetized by avoided corrections, appeals and citizen harm.
Incorrect Answers
A:An output count can rise while quality declines.
B:Inference costs are part of operational value but do not show that incorrect benefit letters or remedial appeals declined.
C:A clear or courteous letter can still contain a wrong benefit determination; audited correctness and remediation are the decisive outcomes.
Question 107
An expense-review agent saves analyst time valued at $60 per hour and costs $400 per week to operate. Assume each saved minute replaces billable analyst time. What minimum number of saved analyst minutes per week is needed to break even before counting any additional risk or review costs?
- 1,000 minutes, which produces a net saving rather than the break-even threshold under these assumptions.
- 240 minutes, converting $400 as though it were $100 per hour of saved labor.
- 600 minutes, which values $600 of gross time and therefore leaves a surplus over the weekly expense.
- 400 minutes.
Explanation
Correct Answer
D:At $60 per hour, the unit value is $1 per minute. Saving 400 minutes generates exactly the $400 needed to offset weekly operating cost.
Incorrect Answers
A:At $1 per saved minute, 1,000 minutes gives $1,000 gross value and $600 net after the $400 cost; that exceeds break-even.
B:The specified analyst rate is $60/hour = $1/minute, not $100/hour; 240 minutes saves only $240.
C:600 minutes yields $600 of labor benefit, $200 more than required to cover the $400 operating expense.
Question 108
A custom agent costs $90,000 initially plus $18,000 annually; a prebuilt option costs $10,000 initially plus $35,000 annually. Over four years, prebuilt TCO is initially $12,000 lower. By how much would its annual license fee need to rise, with all other costs fixed, for the two four-year totals to tie?
- $3,000 more per year.
- $12,000 more per year, applying the full four-year cost difference each year.
- $17,000 more per year, confusing the current annual custom-versus-prebuilt maintenance gap with the required increment.
- $80,000 more per year, using the difference between initial setup prices as if it recurred.
Explanation
Correct Answer
A:Custom totals $90,000+4×$18,000 = $162,000. Prebuilt totals $10,000+4×$35,000 = $150,000. To close $12,000 over four years requires an added $3,000 annually.
Incorrect Answers
B:$12,000 is the entire four-year headroom, not an annual increase; applying it every year would add $48,000.
C:The existing annual cost difference is $17,000, but the total comparison already accounts for it and the $80,000 setup gap.
D:The $80,000 setup gap is a one-time difference. An annual sensitivity calculation must allocate the $12,000 remaining TCO difference over four periods.
Question 109
A legal records workflow has a supported prebuilt classifier with 94% quality and a custom prototype at 95%, but the custom version requires continual labeled data. What should govern selection?
- Treat the vendor classifier as unusable simply because its architecture is closed.
- Calculate TCO only from first-month hosting charges.
- Automatically choose custom because one percentage point is always worth any cost.
- Compare the marginal error reduction against model operations, governance and migration cost.
Explanation
Correct Answer
D:Incremental benefit may be small relative to added training and validation burden.
Incorrect Answers
A:Supported prebuilt performance may satisfy the contract.
B:Recurring lifecycle spending matters beyond one month.
C:Quality gains require cost and risk context.
Question 110
A policy-support agent gives outdated answers because it retrieves superseded documents. A vendor recommends training a custom language model on the same historic policies. What architectural comparison should the solution architect make before paying for custom training?
- Refresh and permission-filter the approved retrieval corpus, benchmark the grounded base model against task cases, and justify custom training only if a measured residual gap remains.
- Train a custom model on all archived and current policies without preserving effective dates, because model weights automatically select the legally binding revision.
- Select custom training based on a lower training loss rather than comparing factual support and total lifecycle cost on held-out policy questions.
- Increase training compute to compensate for stale or unauthorized documents entering the retrieval pipeline.
Explanation
Correct Answer
A:A retrieval and source-governance failure is not proof of insufficient model capability. Correct the evidence and compare representative accuracy, governance and ongoing costs before tuning.
Incorrect Answers
B:Embedding obsolete policy text into weights does not establish version precedence and may make source corrections harder to audit or revoke.
C:Training fit alone cannot show improved policy correctness on unseen queries; customization also adds labeling, deployment, monitoring and retraining costs.
D:More compute does not fix the grounding source selection, effectivity filters or access controls that determine which policy is authoritative.
Agentic AI Business Solutions Architect (AB-100) Practice Tests