AB-650 deep review — September 28, 2026

The AB-650 guide was read in full and mapped to all 58 detailed objectives across ten groups. Six synthetic worked examples, two additional labs (eleven total), 48 answered checks and a useful Agent 365 blog exercise now develop administrative decisions and their evidence.

This is same-context AI research and repair. Twelve offline arithmetic, set and policy-selection assertions passed. No tenant, mailbox, agent, billing policy, provider opt-in, tool registration, network probe or restore was executed. Paid question banks were not accessed. Independent human review remains pending. The guide remains review-required because second-year Backup restore frequency is not resolved by current official documentation.

Exam and resource baseline

The official blueprint retains the July 27, 2026 page date, with no separate skills-effective date or announced replacement objective set. Both objective and status hashes are unchanged; accepted snapshots were preserved. The credential and exam page still describe beta, English-only availability and no Practice Assessment. The exam page exposes no current instructor-led course. No fixed duration or GA date was verified.

Three Learn paths expose 17 modules: 5 + 5 + 7. Their former combined duration of 23 hours 8 minutes is now marked historical because the fetched current pages do not expose duration totals. The 35–55-hour study budget is an editorial estimate, not a provider runtime. Videos, channel catalogs, repositories and signed-in partner events were not treated as completed coursework.

Targeted searches found exact Udemy listings from Dean Ellerby and Hamdy Khaled. Their public search descriptions advertise six tests/360 questions and five tests/300 questions respectively, August updates and documentation-linked explanations. Both direct requests were access-blocked. These are catalog leads; content quality, originality, current coverage, runtime and seller certification claims are unverified. No exact current Pluralsight, O'Reilly, Whizlabs or MeasureUp offering was verified in the bounded search; this is not evidence that none exists.

Objective coverage

Every detailed bullet has a hashed mapping in the operation record.

Group	Objectives	Reviewed teaching and additions
Tenant configuration	8	Domains, profile, licensing, Backup setup/scope/recovery, network insights and health notifications
Workloads	5	Shared mailboxes, channel membership, Copilot meeting policy, SharePoint access/discovery
Entra identities	7	Users, guests, groups, PIM, administrative units, contacts and scoped bulk operations
Authentication/access	5	Methods, password protection, SSPR, troubleshooting and risk-policy migration
Defender	4	Threat policies, alert investigation, response and authorized simulations
Purview	5	DLP, classification/labels, retention, alerts and DSPM
Copilot	9	Readiness, oversharing, web/search/app controls, Cowork, AI providers and connectors
Agents	8	Identity lifecycle, access, ownership, settings, registry, requests, deployment and tools
Agent governance	3	Activity, sensitive content and protection-gap assessment
AI monitoring	4	Spending scopes, workload adoption, measurement and service health

Material updates

• Recovery and operations: distinguish full and granular restore points and permission outcomes; static versus dynamic scope; unsupported mailbox types; preview Full Workload Backup; and shortening-window deletion. Connect network location evidence to health notifications without inferring user entitlement.
• Collaboration: clarify shared-mailbox sign-in/licensing, shared-channel identity and residual file permissions, and meeting defaults versus enforced transcript requirements. Temporary audio processing does not imply no retained compliance evidence.
• Identity: preserve separate human-risk policies and the October 1 migration deadline. Distinguish delegated, agent-identity, agent-user and API-key access.
• Cowork: spending-policy scope grants access. Discovery, provider availability and asynchronous credit limits are separate controls. Policy precedence and carried consumption have their own worked example.
• Agent administration: correct broad-sharing assumptions, custom ZIP upload stages, publisher-specific block effects, owner management and firewall failure diagnostics. Explain the separate one-time template action for existing instances.
• Protection: explicit sharing and encrypted-file rights, generated-label limits, and narrower DLP interaction support complement broader audit visibility. Six examples require learners to calculate or reconcile evidence, not memorize a portal.
• Tool support: BYO MCP remains preview with named supported clients; republishing and deletion limitations affect the lifecycle exercise. Registration, approval, permission consent and successful invocation are separate evidence points.

The September Copilot release notes were reviewed for the latest September 23 entry, including license requests and gradual rollout; the full historical release archive was not audited. The May Agent 365 announcement supports product-level GA while individual integrations retain their own status. Current product procedures determine the guide's implementation details.

Blog intake

Alex Fleck's August 6, 2026 InsideTrack article, Implementing Agent 365: How we're governing and managing AI agents at Microsoft, was accepted for a bounded operations worksheet. Its public main article was read; linked videos, internal systems and performance claims were not reproduced.

The exercise turns an inventory into explicit identity keys, owner/sponsor roles, cross-team handoffs and dated actions. The original numerical example reconciles duplicate rows, missing platform IDs and active-agent telemetry. Microsoft’s internal scale figures are not used as a promised outcome. The article's historical Cowork Frontier description is qualified by current work/school GA guidance.

Contradictions and open evidence

1. Backup frequency — open: the policy procedure and September update allow two-year recovery windows, but the restore table stops at 365 days and the overview describes 52 weeks. No second-year cadence is inferred. This is the same unresolved source gap already recorded for MS-102.
2. Templates — reconciled by operation and date: the July template FAQ excludes already approved agents. September's agent settings explicitly adds a one-time bulk action for existing identity-backed instances, excluding blueprints and AI teammates. Ordinary editing remains nonretroactive.
3. Provider scope — reconciled by newer dedicated procedure: the September 8 Cowork access page says model settings are tenant-wide only. September 14 Cowork administration and September 18 provider settings explicitly support user/group scope. The guide uses the newer dedicated provider procedure and still separates provider availability from Cowork access.
4. Granular restore — reconciled by current procedure: the overview's coming-soon file-version wording is not used to deny the granular file/folder restore flow confirmed in the current restore procedure and September GA notice.

Follow-ups cover the October 1 risk-policy retirement, October 5 Backup frequency gap, October 12 Cowork/template/tool documentation and October 15 beta/catalog status. These events preserve unresolved evidence without continually displacing first reviews.

Verification boundary

The twelve offline assertions check supplied recovery clocks, spending-policy selection and carried consumption, template eligibility/backlog, simplified encrypted file permission gates, inventory intersections, telemetry/adoption denominators, synthetic cost per validated outcome and module/historical-runtime arithmetic. They are not tenant integration tests or proof of product enforcement.

The operation ledger is ADLC_Docs/operations/2026-09-28-ab-650-deep-review.json. It records fetch timestamps, hashes, link health, the original source-validation record, objective mapping, findings, blog limits and local validation results. The shared learning catalog and Microsoft review tracker are updated from the completed guide and receipt.