Practice Exams:

Microsoft AI-103: Online Evaluation of AI Systems

AI & Machine Learning

A model can pass a carefully curated prelaunch test set and still fail after real users introduce unfamiliar questions, document types or tool workflows. Online evaluation helps teams observe whether the running application produces useful, grounded and safe outcomes under actual conditions. But production interactions contain private content, variable authorization and delayed business results, so evaluation cannot mean sending every conversation to another model without controls. Microsoft Foundry supports evaluation of deployed interactions and traces, including capabilities with preview constraints. A credible online evaluation program samples responsibly, records configuration and connects automated judgments with human review and incident response.

On this page
  1. Decide which live outcomes matter
  2. Collect traces with explicit privacy controls
  3. Sample for representativeness rather than convenience
  4. Match the evaluator to observable evidence
  5. Connect observations to the correct subsystem
  6. Use online evaluation as a release and rollback signal
  7. Reconcile automated scores with human and business review

Decide which live outcomes matter

Define quality indicators that reflect the user’s job. A policy assistant needs source-grounded correctness and appropriate abstention, while a scheduling agent needs authorized tool calls and confirmation that a calendar change actually occurred. A summarization app may need coverage of critical facts without disclosure of restricted details. These are different from fluent wording or a high response rate. Write down severity categories and escalation rules so evaluations can identify release-blocking failures instead of reporting only a favorable aggregate. Online evaluation for Microsoft AI-103 should reward measurable user outcomes; a better aggregate score is not proof that a difficult real task improved.

Separate leading signals from verified outcomes. A model judge can flag possible unsupported assertions before a human reviews them; a later user correction or authoritative system readback can establish whether a business action succeeded. Do not label model-generated scores as unquestionable truth. Record uncertainty and evaluation limitations, especially when the judgment itself requires context a sampling system may not have captured.

Collect traces with explicit privacy controls

Useful traces can include the deployment/model version, prompt version, retrieval source IDs, tool decisions, tool results, authorization context and response timing. They need not indiscriminately preserve raw customer messages or full confidential documents. Determine which fields are necessary for the evaluation, apply redaction and access rules, and set retention periods. Tenant isolation matters when several organizations use the same agent or monitoring workspace. A trace system must not become a side channel for sharing otherwise protected content.

An insurance-service assistant might produce a convincing response yet cite a form that expired yesterday. A live trace should connect the user task category with the selected model, source document version, retrieval filter outcome and final citation, but it need not preserve the full customer’s medical history to be useful. Record a stable correlation ID and redacted source reference, then preserve restricted evidence in a governed location only when it is truly required for incident review. Build sampling cohorts across normal requests, escalations, difficult exceptions and denied-access cases. Otherwise the system may report excellent quality because easy informational questions dominate the observed traffic while costly or unsafe failures are underrepresented.

Separate a retrieval defect from a response defect when scoring. If the current form was never retrieved, a model-based judge should not grade the generator as though it had the necessary evidence. If the right form arrived but the answer omitted a critical exclusion, the generator deserves investigation. Have a subject-matter reviewer inspect a sample of high-severity cases and disagreements between automated judges. Use the resulting labels to update an offline regression dataset, then compare future releases against the same baseline while protecting personal data. Online evaluation helps only when its evidence is both representative and appropriate to the decisions an organization makes from it.

Use stable correlation IDs so an evaluation result points to the exact interaction and source conditions reviewed. Where the evaluation technology supports assessing existing traces, confirm the required evaluation telemetry conventions and preview status before relying on it in a production release process. If source records have been deleted or access revoked, evaluation storage and cached excerpts must follow the applicable data lifecycle. Security controls belong to the measurement pipeline as well as the application itself.

Sample for representativeness rather than convenience

Sampling only successful requests biases the results. Include user corrections, partial failures, timeouts, unsupported questions, long documents, multiple languages and authorized high-risk tool actions. Stratify by feature, user group and workload type when privacy and fairness rules permit it. High-severity failures may be rare, so a small random sample can miss them even if overall average quality looks stable. Add targeted audits of sensitive workflows while being honest that those samples estimate a different population.

Avoid evaluating only traffic that already triggered complaints; that creates a misleading negative baseline. Maintain both an unbiased sample and a focused incident sample with distinct labels. Track selection criteria and coverage over time. New product features may attract users whose questions were never represented in the training or test data, so sampling plans should be revisited when adoption changes rather than fixed permanently at launch.

Match the evaluator to observable evidence

Use deterministic checks for schema compliance, forbidden tool names, missing citations and explicit policy violations. Use human reviewers or calibrated model-based evaluators for nuanced faithfulness, coverage and harmfulness judgments. A groundedness evaluator requires appropriate evidence; without it, the judgment may confuse plausible writing with factual support. Record the evaluator and rubric version along with its scores so teams can distinguish a product change from a change in measurement.

Spot-check automated labels on cases with independent expert decisions. Review disagreements, false alarms and missed severe errors. If an evaluator’s prompt or model changes, run it against a stable calibration set before comparing new scores with historical data. Automated evaluation should prioritize reviews and inform decisions, not silently assume authority over questions that require domain expertise or legal interpretation.

Connect observations to the correct subsystem

A poor online evaluation can arise from several stages. Search may return the wrong policy version, a prompt may omit uncertainty, a tool may fail after the model proposed a correct action, or an agent may repeat a side effect during retry. Keep a failure taxonomy that separates retrieval, generation, authorization, tool execution and user experience. Each category has a different owner and repair path. Replacing the model when the real issue is source deletion or API authentication can make the product less stable.

Combine trace evidence with operational metrics such as p95 completion time, throttling, cost per useful task and repeated user attempts. An answer that is technically correct after an unacceptable delay may still fail its product goal; a fast unsupported answer certainly does. Evaluate complete outcomes, not just the final model text. For side-effecting tools, confirm authoritative state rather than trusting an agent’s description of what it believes happened.

Use online evaluation as a release and rollback signal

During a canary release, compare the candidate and stable system on equivalent tasks and authorized user cohorts. Define stop conditions before seeing results, including severe factual, safety and tool errors. Report confidence intervals or uncertainty where sample sizes are small, and avoid promoting a candidate based on a few favorable conversations. Shadow evaluation can be useful for informational paths but must not cause duplicate business actions or violate privacy expectations.

Keep the rollback configuration ready, including prompt, model deployment, retrieval index and tool definitions. If online signals indicate a severe regression, pause traffic according to the approved incident process rather than waiting for a monthly dashboard review. Document the decision and add sanitized examples to future offline tests. Online evaluation should tighten the loop between real failures and controlled improvements, not become an unmanaged experiment on users.

Reconcile automated scores with human and business review

A strong online program has a clear workflow for adjudicating important cases. Independent domain reviewers should see enough source evidence to decide whether the answer was right, whether a refusal was justified, and whether a proposed action was safe. Store approvals and corrections with the exact interaction version. Avoid calling a second AI judge an independent human reviewer. If human signoff is required by the release policy, a green automated dashboard cannot substitute for it.

Use the findings to improve sources, filters, prompts, tool contracts or product UX as appropriate. Report trends by risk and user task, including cases where the application correctly declined to answer. Online evaluation earns trust when it can explain which real outcomes improved, which remain uncertain and what evidence supports promotion or rollback. The purpose is actionable operational quality, not ever-more elaborate scoring machinery.

Continue learning

Related guides

Designing AI Evaluation DatasetsBuild representative evaluation sets with failures, edge cases, and rubrics.Handling Hallucinations in Azure AIDetect unsupported claims and verify answers against trusted evidence.GenAIOps on AzureOperate generative AI with versioning, telemetry, evaluation, and rollback.AI App ObservabilityTrace prompts, retrieved evidence, tool actions, and answers across AI incidents.

Related Posts

• Think Smart, Build Smarter: Mastering AI-900 with Microsoft Azure

• Microsoft AI-300: Serving GenAI: Balancing Latency, Throughput, and Cost

• Microsoft AB-620: Give an AI Agent a Job Before More Tools

• Microsoft AB-100: Integration Design for AI-Powered Business Processes

• Microsoft AB-900: Conditional Access in an AI-Enabled Workplace

• Microsoft AI-900: Machine Learning, Generative AI, and Agents

• Microsoft AI-103: Capacity Planning for Azure AI Services

• Microsoft AI-103: Choosing Azure AI Deployment Types

• Microsoft AI-103: Operating GenAIOps on Azure

• Microsoft AI-103: Grounding Azure AI in Enterprise Data