Skip to content

AI-901 deep review — September 28, 2026

The AI-901 guide was read in full and mapped to 29 detailed objectives across seven groups. It now contains six worked examples, eight labs, 36 explained original checks and two Microsoft blog exercises. Twenty-five offline assertions validate the arithmetic and a local invoice validator. All five Python fragments compile; only the local validator was executed. No cloud SDK, model, audio, extraction or paid lesson was executed or watched. Independent human review remains pending.

Blueprint and learning resources

The official study guide retains the April 15, 2026 skills baseline. The 29 objectives, accepted objective hash and status hash are unchanged; there is no future announcement. AI-900's June 30 retirement remains a historical prerequisite distinction, not a reason to reuse its implementation syllabus.

The credential page lists 13 languages and a sign-in Practice Assessment. The retrieved page did not establish exam duration. Its empty training widgets conflict with directly available resources, so the guide links the one-day course and two seven-module paths: concepts and implementation. The earlier 3h51 and 5h37 path totals were not reproduced. Replacement study budgets are explicitly editorial estimates. The course also lists 13 languages; course and exam metadata are separate observations.

A new Pluralsight implementation course is publicly listed by Clint Bonnett, updated September 16, 2026, at 1h59. Its six-module outline targets the second domain; it is not a complete two-domain path. Direct O'Reilly/Udemy retrieval was blocked. Indexed public outlines show Anand Rao Nednur at 4h4, April 2026; Christopher Nett at 6h20, 12 sections/49 lectures, June 2026; and Kuljot Singh Bakshi at 6h54, 11 sections/47 lectures, July 2026. These observations establish public catalog metadata, not current SDK accuracy, paid lesson quality or completeness.

Whizlabs exposed only a title shell directly and in the indexed product result. Earlier 63-video/three-quiz counts are now marked unverified. A bounded search did not find an exact standalone MeasureUp listing; this does not prove absence. The Microsoft assessment returned only a sign-in shell; no questions were read. Unvetted search results marketing real exam questions were not admitted.

Practical changes

  • Replace the client placeholder with current project/Responses and named-agent examples. Explain deployment names, Entra permissions, resource creation versus invocation, conversation state, SDK compatibility and syntax-only validation.
  • Keep text-analysis item errors visible. Distinguish Speech recognition success, no match and cancellation; retain the one-shot duration boundary. Separate a composed Speech pipeline from direct multimodal voice input.
  • Separate preview voice prompt agents from generally available adjacent voice integrations. Explain streaming/interruption measurement with an original sequential latency exercise.
  • Scope the Image Analysis API retirement to September 25, 2028 and its specific API/container deployments. Teach separate vision input and generation contracts.
  • Distinguish Content Understanding GA asynchronous analysis from preview synchronous Read/Layout and agentic analysis. Explain model aliases, resource defaults, per-request overrides and the Read/Layout no-model exception.
  • Add field confidence/grounding opt-in, normalized-value handling and a narrow original money validator. Arithmetic consistency never establishes source correctness, authenticity or payment authorization.
  • Add context-budget, subgroup, word-error-rate, latency and batch-outcome calculations. Keep failed and unfinished work in the denominator.

Primary implementation sources are linked beside claims in the guide. The operation receipt identifies which sections were read. Long SDK examples, other-language samples, all linked procedures and entire retained broad references were not exhaustively audited.

Blog intake

Content Understanding's August update by Peyton Fraser, Krishnakumar Muthukrishnan and Joe Filcik (August 12) informs an original API/version/evidence decision worksheet. The main announcement was read; benchmark percentages were not adopted as expected learner outcomes.

Foundry's July/August roundup by Nick Brady (September 9) informs an original voice interruption/control event log. Only the introduction, summary and hosted-agent/Voice Live sections were reviewed; the full roundup and linked samples were not audited. Current feature documentation controls preview and API claims.

Evidence and recurring review

The operation receipt is ADLC_Docs/operations/2026-09-28-ai-901-deep-review.json. It records objective mapping, fetch outcomes, earlier records, applied findings, offline checks and local validation. The historical source review is preserved; unchanged accepted objective/status snapshots are retained.

October 12 follow-ups cover client/voice/CU feature changes and learning-catalog uncertainty. The Microsoft review tracker reports actual program coverage. Scheduled automation detects changes and queues review; it does not replace semantic review, execute paid/cloud labs or provide an independent human sign-off.