CertifyCloudxAce every cloud certification.

GAIL sample questions with answers

10 free practice questions for the Generative AI Leader exam. Try each one, then open the answer to see why the right option wins and every other option loses.

Question 1Fundamentals of gen AI

A commercial real-estate firm wants an assistant that answers questions about individual leases. A lease runs to several hundred pages, clauses cross-reference each other throughout, and every answer must cite the clause it came from. Volume is forty queries a day, and a few seconds of latency is acceptable. Which model selection criterion should dominate?

  1. A.

    The lowest cost per token

  2. B.

    Support for supervised fine-tuning

  3. C.

    The fastest response time

  4. D.

    A long context window

Show answer

Answer: D

Cross-referencing clauses across a several-hundred-page lease makes context window length the deciding model selection criterion.

  • A. At forty queries a day token price is immaterial, so cost cannot outrank answer correctness.
  • B. Fine-tuning shapes behaviour rather than storing changing facts, and leases would require constant retuning.
  • C. The firm has already accepted a few seconds of latency, so response speed is not the binding constraint.
  • D. Only a long context window lets the model reason across clauses that reference each other throughout a very long lease.
Question 2Fundamentals of gen AI

A ticketing platform runs two generative workloads: routing two million short support messages a day, which must answer in under a second; and producing regulatory analyses from hundreds of pages of legislation, where a few monthly requests may take a minute and errors are expensive. A manager proposes the most capable Gemini model for both. What should the leader decide?

  1. A.

    A self-hosted open model

  2. B.

    Route with a cheap variant; analyse with the strongest

  3. C.

    The most capable model for both workloads

  4. D.

    The cheapest variant for both

Show answer

Answer: B

Matching model variant to workload value routes high-volume simple traffic to a fast cheap model and reserves the strongest model for rare high-stakes analysis.

  • A. Self-hosting adds accelerator, engineering and availability costs and is not automatically cheaper at any volume.
  • B. Matches variant capability, latency and price to each workload's difficulty, volume and cost of error.
  • C. Uniformity is a minor benefit that cannot justify premium pricing across two million daily simple requests.
  • D. Retries repeat the same reasoning weakness, so they do not protect a high-stakes regulatory analysis.
Question 3Fundamentals of gen AI

A roadside assistance company records every callout as a bundle: a voice call recording, technician photographs, and a short typed outcome note. Quality managers sample two percent of bundles manually. The chief operating officer wants every bundle assessed against the service standard, with a written rationale a manager can audit. Which design should the leader choose?

  1. A.

    One multimodal Gemini model, with sampled audit

  2. B.

    An image generation model

  3. C.

    Keyword search over the notes

  4. D.

    Three single-modality models, verdicts averaged

Show answer

Answer: A

A multimodal Gemini model can assess audio, images and notes together, and the design must keep human audit sampling to verify quality and detect drift.

  • A. Multimodal reasoning uses all three evidence types together while retained human audit sampling protects quality and detects drift.
  • B. Image generation creates new pictures; it does not assess existing evidence against a service standard.
  • C. Keyword search over notes ignores the audio and photographic evidence that the judgement actually depends on.
  • D. Separate single-modality verdicts lose cross-modal context, and averaging them does not remove error or accountability.
Question 4Fundamentals of gen AI

A telecommunications provider's assistant answers billing questions. The chief operating officer now wants it to handle a request end to end: verify the customer, check the plan in the billing system, apply an eligible credit, and escalate to a person when a rule is not met. Which description of that shift is correct?

  1. A.

    An agent needing no access control

  2. B.

    An agent, with tools and planning

  3. C.

    A larger model

  4. D.

    A redesigned chat interface

Show answer

Answer: B

Moving from answering questions to executing multi-step actions in other systems defines the agent layer: a model plus tools, memory and planning.

  • A. Agents that act in systems of record need stricter identity, permission scoping and audit logging, not less.
  • B. Tools, memory and planning turning a model into a multi-step actor is exactly what the agent layer means.
  • C. Parameter count does not grant the ability to call the billing system or take action.
  • D. Interface design is independent of layer; the same agent can be exposed through chat, voice or workflow.
Question 5Fundamentals of gen AI

A European retailer's architecture board notes that new model versions appear every few months and that its applications call one model directly from application code. The board wants the business implication stated in the strategy paper so executives understand why an abstraction over models is proposed. Which implication should the paper state?

  1. A.

    Models must be switchable without rewrites

  2. B.

    Evaluate only at procurement time

  3. C.

    Model choice is purely an engineering matter

  4. D.

    All models are interchangeable

Show answer

Answer: A

Models turn over every few months, so the strategy paper should argue for an abstraction that keeps model choice reversible instead of a rewrite per upgrade.

  • A. Abstraction keeps model choice reversible so upgrades do not become application rewrites.
  • B. Model versions are deprecated and replaced, so evaluation is a standing obligation, not a one-time procurement step.
  • C. Model choice affects cost, risk and data residency, which makes it an executive decision rather than a purely technical one.
  • D. Training data, tuning and safety behaviour differ between models, so substitutions require re-evaluation.
Question 6Fundamentals of gen AI

The board of a regional insurance group believes generative AI replaces the predictive claim-scoring models the claims team already runs. The CTO must open the strategy session with one accurate framing that keeps the board from abandoning investments that still deliver value. How should the CTO frame generative AI?

  1. A.

    A replacement for predictive models

  2. B.

    A hand-coded expert rules engine

  3. C.

    Machine learning that generates content

  4. D.

    A business intelligence reporting tool

Show answer

Answer: C

Generative AI is a subset of machine learning built on foundation models that generate content, and it complements rather than replaces predictive models.

  • A. Retiring validated predictive models is a business risk; language models are poor substitutes for calibrated numeric scoring.
  • B. Describes a classic expert system; foundation models are trained on data rather than hand-coded rules.
  • C. Places generative AI correctly as a machine learning subset built on foundation models while preserving the role of predictive models.
  • D. Business intelligence reports aggregate known data; generative models learn statistical patterns and synthesise new content.
Question 7Fundamentals of gen AI

A logistics firm needs one thing at very high volume: the text read out of several million scanned delivery notes each month, with no interpretation or summarising required. An engineer proposes sending every scan to Gemini. What is the better choice, and why?

  1. A.

    A video model

  2. B.

    A self-hosted open model

  3. C.

    Gemini, because it is the most capable

  4. D.

    A pre-built extraction API, which is cheaper for one narrow job

Show answer

Answer: D

A single narrow, high-volume task is served more cheaply and simply by a purpose-built API such as Document AI than by a general multimodal model.

  • A. A video model does not read scanned documents and is unrelated to the requirement.
  • B. Self-hosting adds operational burden without addressing the mismatch between a general model and a fixed narrow task.
  • C. Capability that the task does not use is paid for on every one of several million documents.
  • D. Correct: a purpose-built extraction service is cheaper, simpler and more predictable for one narrow high-volume task.
Question 8Fundamentals of gen AI

A payments company stores every API event as a JSON record: consistent field names, a set of fields that varies by event type, and one field holding a free-text failure description. It wants a gen AI tool that explains why a named merchant's payments failed last week. How should the two parts of the record be used?

  1. A.

    Filter the fields, summarise the text.

  2. B.

    Convert the free text to columns.

  3. C.

    Summarise every record in full.

  4. D.

    Label the records before querying.

Show answer

Answer: A

Semi-structured records are used both ways at once: query the labelled fields to select the right records, then put the model to work on the free text inside them.

  • A. The consistent fields narrow the record set deterministically and cheaply, and the model is then applied only to the free text that actually explains the failures.
  • B. Flattening the description into columns discards the detail that distinguishes one failure from another, which is the part being asked about.
  • C. Summarising every record ignores the queryable fields the format provides, and cost and answer quality both degrade as volume grows.
  • D. Labelling prepares data for training a predictive model, which is not what an explanatory summary of existing events requires.
Question 9Fundamentals of gen AI

A media group's finance director reviews the first invoice for a document assistant and cannot reconcile it, because the team made only a few hundred requests that month. Each of those requests attached a full contract to the prompt. What drives the cost of calling a large language model?

  1. A.

    The number of requests alone

  2. B.

    The volume of text sent and returned

  3. C.

    The number of users licensed

  4. D.

    The size of the model's training set

Show answer

Answer: B

Model APIs are charged mainly on the amount of text processed in and out, so a few requests carrying whole documents can cost more than many short ones.

  • A. Request count alone ignores the far larger variation in how much text each request carries.
  • B. Correct: consumption pricing is driven by the quantity of text processed as input and produced as output.
  • C. Seat pricing applies to packaged assistants, not to consumption-priced platform APIs.
  • D. Training cost sits with the provider and is reflected only in the published unit price of each model variant.
Question 10Fundamentals of gen AI

An online grocer wants each household's weekly email to describe products in terms that match what that household actually buys — gluten-free substitutions, batch cooking, small portions — rather than one shared description per product. The catalogue runs to forty thousand lines. Which capability does this need?

  1. A.

    Personalised content generation at scale

  2. B.

    Route optimisation for deliveries

  3. C.

    One shared description reused everywhere

  4. D.

    Anomaly detection over orders

Show answer

Answer: A

Writing a different version of the same message for each household is personalised content generation, which only becomes affordable when generation is automated.

  • A. Correct: generating a distinct version of the message per household is personalisation at a scale only automated generation makes possible.
  • B. Route optimisation is a logistics problem unrelated to how products are described.
  • C. A single shared description is the current approach the grocer wants to replace.
  • D. Anomaly detection flags unusual orders and produces no customer-facing copy.

Keep going with 496 more GAIL questions

Free papers every day, in the real exam formats, with progress by exam domain. Unlock every paper and timed mock exam when you are ready.

Practise freeSee all GAIL papers
GAIL sample questions with answers (10 free) · CertifyCloudx