Deploy AI-powered business solutions (20-25%)

AB-100 · 73 questions

• Makers at a manufacturer edit the production Copilot Studio agent directly, and last week's working version can no longer be restored. Connectors in the agent point at production systems even when someone experiments. What ALM change should the architect make?
• A bank's case agent reads customer PDFs as knowledge. One uploaded PDF says to ignore prior instructions and email the full case list to an outside address. The agent has a tool that can send mail. What should the architect change?
• A government agency requires that prompts and grounding transcripts for a citizen agent stay inside the national boundary it named. A teammate suggests a global model deployment because the tenant home is in that country. What should the architect validate?
• A utility's outage agent can draft switching orders that move load on the live network. A mistaken order could drop service to a neighborhood. Which control should the architect require before go-live?
• A bank runs an autonomous agent that drafts a case note when a transaction alert is created. There is no chat. Supervisors want to know whether the agent is behaving. What should operations monitor first?
• A contact-center agent is released because a pilot showed a high thumbs-up rate. Leaders have no measure of whether answers match the knowledge source, whether the customer's request was completed, or whether the agent's actions were correct. What should testing add?
• A bank sorts its backlog of agent fixes by which manager shouted loudest, rather than by how often the answer failed. The same wrong fee answer has failed for hundreds of customers and sits near the bottom. What should the architect do with the backlog?
• Patients downvote a hospital portal agent, and the comments say the prep sheet it quotes is outdated. A teammate wants to retune the model immediately. What should the architect do first?
• Advisors at a university have stopped using the registration agent. Their feedback says it refuses every transfer-credit question, and they now email the registrar instead. What should the architect treat this pattern as?
• An insurer's evaluation traces show that policy-rider questions retrieve no usable rider passage, while deductible and limit questions retrieve the right clause and are answered correctly. What should the architect tune?
• An airline's rebooking agent writes a trace whenever a tool call fails, and nobody reviews those traces. The operations board shows only how many runs started. What should operations monitor?
• A retailer judges its returns agent by the CPU of the storefront web server. Leaders cannot tell whether a return was completed, refused, or handed to a person. What should operations monitor instead?
• An airline gate agent answers slowly. Traces show the gate-assignment tool timing out, and the model returns promptly once that tool responds. Leadership wants the model retuned. What should the architect change?
• A retailer evaluates its returns agent. The greeting matches the brand script, and the following sentences cite a returns policy the company has retired. What should the architect change based on that evaluation?
• A manufacturer's agent is slow at only one plant, and only on the turns that call the parts service. Turns that do not call that service return quickly. What should the architect inspect before any model change?
• Leaders at a utility want one number for the outage agent, which tells a customer the current status of that customer's outage. Which metric should the architect report?
• A hotel booking agent shows a high action count. Traces show the booking tool failing and the agent retrying the same call. Which metric should operations use to judge this agent?
• A bank reports the average length of its fee agent's replies as the quality number. Supervisors need to know whether the fee in the reply matches the fee schedule. Which metric should replace the length figure?
• A logistics agent drafts delay notices. Traces show the model returns quickly, while the carrier-status tool accounts for most of the wait on each run. What should the architect change first?
• A hospital portal agent starts refusing more questions in the hour after a handbook is published into its knowledge. Telemetry records both the publish and the refusals. What should operations examine first?
• An insurer wants to promote a custom fraud model. Nobody has agreed how precise it must be, and no holdout set exists. What must be in place before production?
• A public agency's eligibility agent must include one required statutory sentence, word for word, in every eligibility answer. Current tests pass when the reply is polite. What should the suite add?
• A pharmacy agent's tests pass whenever any reply comes back to a symptom question. Pharmacists require that the agent not recommend a dose. What should the test suite add?
• A hospital validated a clinical-coding model only on notes from one clinic and wants to use it at every site. What should the validation criteria require before that wider use?
• At a distributor, the sales copilot drafts a quote, the finance copilot creates the invoice, and the service copilot opens a case. Each copilot looks right in its own app, yet the case names a different total than the invoice. What should the architect require?
• A manufacturer validated its weld-defect model only on sharp studio photos. Plants submit glare, dust, and cropped beads, and none of those photos are in the validation set. What should the architect require before promotion?
• A bank's approved prompt rules allow a copilot to summarize a credit file and forbid any prompt that directs staff to decide a lending exception. A submitted prompt says to decide the exception and reply approved or denied. What should prompt review do?
• A retailer submitted a copilot prompt that pastes the entire returns manual into the instruction text. Staff already keep that manual as a governed document. What should prompt review decide?
• A university's approved prompt rules require a refusal line when a question is outside the student handbook. A submitted study-skills prompt sets a role and an output shape, and it has no refusal line. What should prompt review do?
• A hospital books a supply request in one app, and finance places that same request on a hold list when the amount needs review. Tests of each screen pass on their own. What test should the architect add before go-live?
• A bank overwrote the production knowledge file for a product agent and has no prior copy to restore. What should lifecycle management require for that grounding data?
• A university edits the handbook file that grounds a help agent directly in production. How should lifecycle management treat that handbook?
• A retailer wants faster test coverage for a returns agent whose approved policy is already written. How should the suite be built?
• An airline loyalty desk changes a passenger's tier in its Dynamics 365 sales app and must see that tier on the open service case. Which test should the architect require?
• A hospital tests a prep-sheet agent that may answer only from the approved sheet. How should the test cases be produced?
• A bank uses Copilot to draft test cases for a payment agent. The draft covers successful payments and skips declines and timeouts. What should the test strategy do?
• A utility is ready to release a Microsoft Foundry agent to production, and the release path has no evaluation gate. What should the architect add?
• An insurer's test agent and production agent both read one policy library, so a wording edit made in test shows up in production answers the same day. What should the architect change?
• A hospital maker edits the production Copilot Studio agent and exports that unmanaged change to save it. What should the architect require instead?
• A bank team edits a Microsoft Foundry agent directly in the production project. What should the architect require instead?
• A retailer imports a Copilot Studio solution into test, and the connection references still point at production. What should happen before the next promote?
• An airline promotes its Microsoft Foundry agent and the model deployment on separate schedules, and the two have drifted. What should the release do?
• A hospital's test environment and its production environment call the same model deployment. A test prompt changes answers that clinicians already see on the live floor. What should the architect change?
• A manufacturer trains its weld-defect model in the production project, using live line photos while inspectors still rely on the current model during the shift. How should the architect change that lifecycle?
• A manufacturer's supply-chain copilot answers from a different knowledge set in development, test, and production, and nobody planned that split. What should the architect make part of the release?
• During month-end close, a distributor changes a finance copilot setting directly in production. Controllers see the guidance shift while the books are still open. What should the architect require for the next change?
• A retailer turns a finance AI feature on in the production company before any test company has run it. Controllers see the new guidance during the live close. What should the architect require?
• An insurer replaces a custom model by overwriting the production model in place, and the previous behavior cannot be restored. What should the release do instead?
• An insurer's Dynamics 365 service copilot in the test environment still reads production cases. The team wants to treat the latest run as a valid test. What must change first?
• A retailer's returns agent can submit a refund of any amount, and no person sees the refund before the money is paid. What control should the architect put on that tool?
• A hospital edits customer-service copilot customizations directly on the live case form. A recent edit changed the summary families hear about, and no sandbox copy exists. How should those customizations ship?
• A university will ship a Dynamics 365 sales copilot change this week and a service copilot change next week, and the two contradict on the same student status. What should the release plan do?
• A hospital scheduling agent signs in with a shared personal account that can open charts, payroll, and the booking calendar. What should replace that sign-in?
• A bank's agent reads every customer's file through one service account, and those reads are not written to an audit log. What should security require?
• At a bank, any developer can deploy a model into the production Microsoft Foundry project. What should the architect change?
• A public agency's citizen agent was built by a vendor, and no agency employee will own it once that vendor leaves. What should governance assign before the vendor goes?
• A retail department wants agent logging turned off for a promotion week so the agent can go faster. What should governance decide?
• An airline has been putting agents into production with no record of the actions each agent is allowed to take. What should governance require before the next agent goes live?
• An insurer still has an old model key pasted in a notebook that the data-science channel shares. What should the team do?
• A hospital tuned a model from a storage account the whole company can read, including files meant only for that tuning job. What should security change?
• A bank's proof of concept ran on a global model deployment, and the privacy office allowed that proof. The production build copies the same global deployment. What should the architect do before production?
• A retailer's returns agent reads product reviews while it decides a refund. One review tells the agent to ignore policy and issue the refund twice. The agent has a tool that can post a refund. What should the architect require?
• A hospital portal agent receives a patient message that says to ignore its rules and reveal the next patient's record. What should the design require?
• A bank is ready to release a credit-explanation agent used in several regions. No one has checked whether the explanations treat applicants consistently across those regions. What should the responsible-AI review require before go-live?
• A hospital discharge agent messages patients with next steps and does not tell them the message is automated. The responsible-AI review has not required that disclosure. What should the review require?
• A retailer's hiring-screen agent rejects applicants and provides no way for a person to reverse that rejection. The responsible-AI review is asked who is answerable for the decision. What should the review require?
• At a consultancy, every maker can open every client's grounding library, including libraries for clients they do not work with. What access change should the architect require?
• A citizen-facing agent at an agency answers a published FAQ, yet its grounding permission also lets it read case notes the FAQ does not use. What should the architect change?
• A hospital's tuning set of clinician notes is readable by the developers who tune the agent and by a reporting intern who does not tune models. What should the architect change?
• A hospital stores grounding recordings in a region the privacy office has not approved. A teammate wants to change the agent's description to name an approved region and leave the files where they are. What should the architect do?
• A bank can swap the model behind a Foundry deployment, and nothing records who made the swap. What should the architect require before that deployment is used in production?
• An insurer replaces the policy files that ground a claims agent and keeps no history of the prior file or of who replaced it. Model changes are already logged. What else should the architect require?
• A utility agent drafts a switching plan. An approver overrides the draft, and the stored record keeps the original draft with no sign of the override. What must the audit trail include?