7.5 Exam Day Quick Reference
Key Takeaways
- Logistics: 700/1000 to pass, ~100 minutes, about 40-60 questions, ~$165 USD; the AI-102 retires June 30, 2026 (11:59 PM CST).
- Current six domains and weights: Plan/manage (20-25%), Generative AI (15-20%), Agentic solutions (5-10%), Computer Vision (10-15%), NLP (15-20%), Knowledge Mining & Information Extraction (15-20%).
- Deprecated-to-current names: LUIS -> CLU, QnA Maker -> Custom Question Answering, Form Recognizer -> Document Intelligence, Cognitive Search -> AI Search; new Foundry branding (Microsoft Foundry, Content Understanding) is now in the objectives.
- Security hierarchy (best to worst): managed identity > Entra ID service principal > keys in Key Vault > keys in env vars > keys in config files > hardcoded keys (never).
- Content Safety uses a 0-7 scale with four categories (hate/fairness, sexual, violence, self-harm); Azure OpenAI requires its own dedicated resource and cannot be added to a multi-service Azure AI Services resource.
Quick Answer: Pass at 700/1000 in ~100 minutes. The exam was redesigned around Microsoft Foundry and now spans six domains. Lead with purpose-built services over generative ones, prefer managed identity, remember the 0-7 Content Safety scale, and note the exam retires June 30, 2026.
Exam Logistics (verified)
| Item | Value |
|---|---|
| Passing score | 700 / 1000 |
| Duration | ~100 minutes |
| Questions | ~40-60 (Microsoft does not publish a fixed count) |
| Cost | ~$165 USD (varies by region) |
| Vendor | Microsoft (delivered via Pearson VUE / online proctor) |
| Retirement | June 30, 2026, 11:59 PM CST |
| Renewal | Free online assessment on Microsoft Learn before expiry |
Current Domain Weights (Dec 2025 / Foundry redesign)
| Domain | Weight |
|---|---|
| Plan and manage an Azure AI solution | 20-25% |
| Implement generative AI solutions | 15-20% |
| Implement an agentic solution | 5-10% |
| Implement computer vision solutions | 10-15% |
| Implement natural language processing solutions | 15-20% |
| Implement knowledge mining and information extraction | 15-20% |
The December 23, 2025 redesign added the agentic domain (Foundry Agent Service, Microsoft Agent Framework, multi-agent orchestration) and reweighted the rest. If a practice resource still shows a "content moderation" domain or six-vs-five weightings from the old blueprint, it is stale.
Service Selection One-Liners
- Text from images -> Vision Read (OCR)
- Fields from forms/invoices -> Document Intelligence
- Sentiment / NER / PII / key phrases -> Language
- Custom intents + entities -> CLU
- FAQ answers -> Custom Question Answering
- Generate / summarize / chat -> Azure OpenAI
- Image generation -> Azure OpenAI (DALL-E)
- Search + AI enrichment -> AI Search
- Text/image moderation -> Content Safety
- Identity (1:1 / 1:N) -> Face (Limited Access)
- Video insights -> Video Indexer
- People counting -> Vision Spatial Analysis
- Multimodal extraction -> Content Understanding
Security Hierarchy (best to worst)
- Managed identity — no secrets in code, auto-rotated tokens
- Entra ID service principal — token-based, rotatable
- Keys in Key Vault — centralized, audited
- Keys in environment variables — better than files
- Keys in config files — source-exposure risk
- Hardcoded keys — never acceptable
Critical Numbers
| Item | Value |
|---|---|
| Passing score | 700 / 1000 |
| Duration | ~100 minutes |
| Content Safety scale | 0-7 (text returns 0/2/4/6) |
| Content Safety categories | 4 (hate/fairness, sexual, violence, self-harm) |
| 1 token | ~0.75 English words |
| GPT-4o context | 128K tokens |
| text-embedding-3-large dims | 3,072 |
| Composed model components | up to 100 |
| Fine-tuning minimum examples | ~10 (more recommended) |
Top Exam Traps
| Trap | Correct answer |
|---|---|
| Use LUIS for intents | Use CLU (LUIS deprecated) |
| Use QnA Maker | Use Custom Question Answering |
| API keys for production auth | Use managed identity |
| Public endpoints in production | Use private endpoints |
| Disable all content filters | Cannot fully disable; tune per category |
| Content Safety scale is 0-10 | It is 0-7 |
| Add Azure OpenAI to a multi-service resource | OpenAI needs a dedicated resource |
| Face emotion recognition | Retired |
| Translator on the standard endpoint | Uses api.cognitive.microsofttranslator.com |
| Adjust temperature AND top_p together | Tune one, not both |
| Fine-tune for fresh knowledge | Use RAG for current facts |
On the Exam: When two services could both work, pick the purpose-built one (cheaper, deterministic) unless the requirement is open-ended generation. When in doubt on auth, pick managed identity; when in doubt on knowledge freshness, pick RAG over fine-tuning.
Question Formats and Time Budgeting
With roughly 100 minutes for an expected 40-60 items, you have a little under two minutes per question on average — but the formats are uneven. Standard multiple-choice and multiple-response items are fast; spend your saved time on the slower ones. Drag-and-drop ordering items ask you to sequence steps (for example, the RAG pipeline order or the document-processing chain) — these reward the order-of-operations facts in section 7.4. Case studies bundle several questions behind one scenario; read the requirements and constraints carefully because a single phrase like "least cost" or "no secrets in code" decides multiple answers.
Some exams include a lab or repeated-answer-choice block. Flag anything uncertain and return to it; an unanswered question scores the same as a wrong one, so never leave a blank.
High-Frequency Fact Recall
These facts appear often enough to memorize verbatim. Azure OpenAI needs a dedicated resource (kind OpenAI) and cannot live inside a multi-service Azure AI Services resource. Content Safety is 0-7 with four categories. Translator uses api.cognitive.microsofttranslator.com and needs the region header with a multi-service key. The standard key header is Ocp-Apim-Subscription-Key, but Azure OpenAI key auth uses api-key. Face identification and Custom Neural Voice are Limited Access; Face emotion/age/gender are retired. A long-running operation returns 202 with Operation-Location.
429 means rate-limited — back off, do not rotate the key. Adjust temperature or top_p, not both. Use RAG for fresh facts, fine-tuning for style.
Final Decision Heuristics
When a question stalls you, apply these tie-breakers in order. (1) Match the verb to the purpose-built service before reaching for a generative one. (2) For authentication, climb the security hierarchy toward managed identity. (3) For networking, prefer private endpoints over public access. (4) For safety, never "disable" — tune thresholds, add blocklists, or insert human review. (5) For knowledge freshness, choose RAG; for response shape, consider fine-tuning. (6) For order-of-operations items, put safety screening on both the input (Prompt Shields) and the output (content filters) and place retrieval before generation.
(7) When two answers look correct, re-read the stem for the constraint phrase — "least effort", "lowest cost", "highest security", "offline" — that singles one out. These heuristics resolve the large majority of AI-102 scenario questions without any guesswork.
A production Azure AI application stores its API keys in environment variables, and a security review flags this. What is the recommended improvement?
According to the current Microsoft Foundry redesign of AI-102, which domain was newly added and carries the smallest weight?
Which statement about Azure OpenAI on the AI-102 exam is TRUE?
You must build one bot that answers HR FAQs and also processes structured leave-request actions. Which architecture is correct?
You've completed this section
Continue exploring other exams