AB-730 AI Business Professional Study Guide¶

Independent AI-assisted resource — HUMAN REVIEW PENDING. The October 28, 2026 refresh maps all 41 detailed objectives in the October 20, 2026 English outline. See the current deep-review evidence and sources-and-objectives record. Microsoft's official blueprint is authoritative. Live product labs and human review remain pending; public-source review does not guarantee correctness.

Study guide at a glance

Start here

Report an issue with AB-730

Blueprint checked2026-10-28

Review stateReview required

Guide-only active reading1 hr 15 min–1 hr 45 min

Prerequisites: Basic Microsoft 365 business-user experience in Outlook, Word, Teams, PowerPoint and Excel plus hands-on use of Microsoft 365 Copilot prompts, conversations, agents, Researcher, Analyst, Pages and Notebooks; no coding or app-development experience is required.

Exam essentialsRead the objective map, key distinctions, and readiness checklist.

Deep understandingWork through every domain, including decisions, failure modes, and related items.

Hands-on labsBegin with Useful blog reading with an exercise, then use the remaining labs to produce evidence you can inspect and explain.

Official objective domains

Understand generative AI fundamentals25–30%
Manage prompts and chats by using Microsoft 365 Copilot20–25%
Manage business content and collaboration by using Microsoft 365 Copilot20–25%
Drive business outcomes by using agents20–25%

The active-reading range uses 90–130 words per minute for technical material. Labs, troubleshooting, note-taking, spaced review, and prerequisite gaps add time; it is not a total preparation promise.

Current English baseline: Skills measured as of October 20, 2026. The former July 22 snapshot is preserved with the historical reviews.
Blueprint status (checked October 28, 2026): The October 20 date has passed. The four-domain outline below is current English scope; no later revision is announced on the checked study guide. Localized exams can lag English updates, so verify the outline for your appointment language.
Lifecycle: The AI Business Professional credential and 45-minute exam are active; the credential page currently lists 13 languages. Recheck the revision date for your exam language.
Official source: AB-730 study guide

Source-review blockers — October 28: Current objective coverage is reconciled, but Microsoft documentation disagrees about connector writes, scheduling prerequisites, meeting capture and Cowork model-control scope. The bounded guidance below cites the more specific contracts and preserves the conflicting statements. The source gate remains blocked pending resolution; tenant behavior and the agent-history filter still need verification.

How to use this guide¶

AB-730 is a beginner business-user exam, but useful preparation is not a list of Copilot buttons. For every task, trace:

the business outcome, audience, decision, and definition of a good result;
the right Microsoft 365 app, chat, agent, page, notebook, Researcher, or Analyst experience;
the goal, context, sources, constraints, and expected output in the prompt;
what organizational or web data may be used and what the user is allowed to access;
the risk of fabrication, prompt injection, oversharing, bias, copyright, or over-reliance;
the verification and human judgment required before the output is used;
whether the reusable prompt, conversation, page, notebook, or agent should be saved, shared, scheduled, or deleted.

Practice with realistic, nonsensitive material in Word, Excel, PowerPoint, Outlook, Teams, and Microsoft 365 Copilot. Build a small agent from a template if your tenant permits it. The role improves work with AI; it does not build AI applications or write code.

About related items: A Related item: callout adds prerequisite, operational, architectural, or adjacent context that makes the current topic easier to understand. It is useful supporting knowledge, not a claim that the item appears verbatim in the published exam objectives.

Current scope and learning route¶

The October 20 blueprint has four current domains. Use these weights to plan study; product announcements do not add objectives:

Current study area (paraphrased)	Weight	Learning focus
Generative AI foundations	25–30%	Permission-aware context, responsible use, choosing chat or delegated work
Copilot prompts and chats	20–25%	Model choice, prompt refinement, memory/instructions, chat and notebook management
Business content and collaboration	20–25%	Connectors, Work IQ, cross-app output, visualizations, meeting preparation and recap
Business outcomes with agents	20–25%	Prebuilt/custom-agent choices, Cowork skills, task supervision and scheduling

The earlier custom-agent construction group is removed. The choice between prebuilt and custom remains, while the new domain requires applying agents, filtering their chats, and supervising Cowork. Sections 2–4 cover the added model, notebook, connector, visual, meeting, skill, and automation objectives. The retained custom-agent construction exercise is Related item: optional practical depth, not a replacement for selecting and applying prebuilt agents in the current outline.

Context, delegation, and checking the result¶

Work IQ combines organizational information, context, and tools under permission-aware governance. A useful study question is which authorized sources support a business conclusion. Specify the reporting period, people, files, and expected deliverable; verify that the cited evidence actually supports the answer. A persuasive summary does not establish that every relevant source was available.

Copilot Cowork performs tasks across Microsoft 365 and loads skills for particular kinds of work. Study the difference between generating an answer and authorizing actions that affect files, messages, or calendars. Review the proposed work, inspect progress, and verify the resulting artifacts. VERIFY CURRENT: Account type, organizational controls, preview features, and approval settings affect available behavior. The overview distinguishes generally available work/school access from personal-account preview; that does not establish eligibility for every feature.

The Cowork operating guide explains skills, session controls and schedules. Section 4 distinguishes a single action, a session-wide approval and automation preauthorization. Record scope, timing, expected output and how to stop recurring work; inspect a run before relying on it.

Original practice extension¶

Use a disposable work area and synthetic project files. Ask for a weekly project brief with unresolved decisions, evidence links, a small chart, and a draft follow-up. Define who may receive the output before starting. Compare an ordinary chat response with a delegated Cowork task; explain which parts need source retrieval, a document skill, and an action approval. Inspect the proposed recipients and generated numbers before allowing any communication. If scheduling is available, create a test recurrence, verify its settings, and disable it after the exercise. If unavailable, document the intended setup and the access limitation.

Readiness: explain why a fluent answer may still lack evidence; justify chat versus a prebuilt agent versus Cowork; distinguish a reusable skill from a scheduled task; and identify where a person checks the plan, actions, and final result.

Objective map¶

Published domain	Weight	Central question
Understand generative AI fundamentals	25–30%	Can you select an appropriate Copilot experience and use it responsibly with permitted context?
Manage prompts and chats by using Microsoft 365 Copilot	20–25%	Can you select a model, refine reusable prompts, manage memory/instructions, chats and notebooks?
Manage business content and collaboration by using Microsoft 365 Copilot	20–25%	Can you use authorized context, connectors, apps and meeting tools to produce verified business content?
Drive business outcomes by using agents	20–25%	Can you select and apply agents, supervise Cowork skills, and manage recurring work?

1. Understand generative AI fundamentals¶

Understand what generation does—and does not prove¶

Generative AI predicts useful output from instructions and context. It can draft, summarize, reorganize, compare, brainstorm, translate, explain, extract, and analyze. Fluent output is not evidence that the result is true, complete, current, authorized, unbiased, or suitable for the decision. Treat the response as a proposed work product whose verification depends on its use.

The model has learned patterns from training. The current conversation, referenced files, web results, app context, organizational content, and configured instructions can ground a response. Grounding makes a response more relevant and checkable; it does not guarantee truth. If the source is incomplete, outdated, ambiguous, or inaccessible, the output can inherit the problem.

Ask whether the task is:

creative or exploratory: ideas, variants, outlines, tone changes;
transformative: summarize, rewrite, compare, translate, extract;
analytical: trends, explanations, formulas, charts, risks, recommendations;
research-oriented: gather and synthesize multiple sources with citations;
action-oriented: an agent uses knowledge or capabilities for a repeatable task.

The higher the consequence, the stronger the source, verification, approval, and record needed. A brainstorm can tolerate uncertainty; financial reporting, personnel decisions, customer commitments, safety instructions, and legal claims cannot.

Understand context and organizational protection¶

Microsoft 365 Copilot can use the prompt, conversation, current app, attached or referenced material, permitted Microsoft 365 work content, and sometimes web content. Microsoft Graph connects the experience to organizational data the signed-in user can access. Copilot does not grant a user new permissions, but existing oversharing can make too-broad content discoverable and usable. “Copilot respects permissions” is not the same as “all existing permissions are correct.”

Context changes the response:

In Word, the current document and referenced files support drafting and revision.
In Outlook, a message or thread supplies communication context.
In Teams, meeting transcript/chat and channel context support summaries and follow-up.
In PowerPoint, a prompt or source document can seed a presentation.
In Excel, a structured table and explicit analytical question support formulas, insights, and visuals.
In Microsoft 365 Copilot Chat, work or web grounding and selected sources shape a cross-app response.

Check which context is active, whether it is complete and current, and whether every intended recipient may access the resulting content. A generated summary can reproduce sensitive details from its source even when the summary itself looks harmless.

Related item: Data hygiene precedes AI hygiene. Clear ownership, sensible sharing, sensitivity labels, retention, and accurate source documents improve both conventional search and Copilot results.

Distinguish chat, agents, Researcher, Analyst, pages, and notebooks¶

Experience	Strong fit	Important boundary
Copilot Chat	Flexible one-off or iterative prompting across work/web context	Conversation and available grounding vary; verify sources and permissions
Copilot Cowork	Delegate a deliverable spanning several actions or apps	Inspect action authority, skills, usage budget, progress and final artifacts
App Copilot	Work grounded in the active Word, Excel, PowerPoint, Outlook, or Teams context	Each app exposes different capabilities; do not assume feature parity
Agent Store agent	A discoverable prebuilt or organization-published agent already fits the task	Review publisher, knowledge, capabilities, sharing, and approved use
Custom agent	A repeatable, bounded task needs tailored instructions, knowledge, suggested prompts, or capabilities	Build only when an existing agent does not fit; test and share deliberately
Researcher	Multistep research and synthesis across allowed sources with an evidence trail	Inspect citations, source quality, omissions, and recency
Analyst	Deeper data reasoning, calculations, analysis, and visual exploration	Validate input shape, assumptions, computations, and business interpretation
Copilot Page	Editable, shareable canvas that turns a response into collaborative content	Sharing the page can broaden the audience; verify content and access
Notebook	Persistent project context that collects conversations and sources	Curate source scope, remove stale material, and understand who can access it

A chat responds inside a conversation. An agent has a more durable purpose, instructions, knowledge, capabilities, and sharing boundary. Create an agent when a repeated task needs controlled reuse; do not create one merely to rename a prompt. VERIFY CURRENT: agent availability, Agent Store content, Researcher/Analyst capabilities, Pages/Notebooks behavior, licensing, regions, sharing, and supported apps change frequently.

Connect work context without assuming complete access¶

Copilot connectors bring external systems into the context available to Copilot. Synced connectors index content; federated connectors fetch it live without building that index. Both depend on source permissions and organizational configuration. A live fetch still might omit a source or return incomplete data.

VERIFY CURRENT — October 28 documentation discrepancy: The general overview still calls federation read-only and describes writes as forthcoming. The dedicated federated-connector reference now documents create/update/delete tools where the connector and Copilot experience support them; Researcher remains read-only. A source connection is not approval for every action. Inspect the individual tool, source identity, parameters and destination. Its approval can cover one call, the conversation, or future calls until reset; those permissions are tool-specific. Confirm actual availability and existing consent before using a write-capable connector.

For a missing customer ticket, check the source directly, your access, connector enablement/authentication, query scope, and indexing delay when applicable. Record a missing source as a limitation. Do not turn “no ticket found” into “the customer has no problems.” Name account, time period and system, request citations, and reconcile conflicting records with their owners.

Work IQ supplies business context; it is not permission to see every record. Its API billing is separate from the included experience described for licensed Copilot users. Do not infer entitlement from a button or an API announcement. The September article's September 30 preview start and staged October rollout are historical announcements, not confirmation of availability in your environment; see the dated blog exercise below.

Worked example: derive a business conclusion across sources¶

A synthetic renewal brief has three permitted sources: a signed contract for 90,000, an older presentation proposing 120,000, and a support system containing an unresolved migration issue. Work IQ and connectors can help join this context, but they do not decide which commercial promise is authoritative. Ask for a table of claim, source, date, contradiction and owner. Use the signed amount for the contractual fact, keep the proposal as history, and ask the account owner how the unresolved issue changes renewal risk. If support access is missing, report that gap; do not conclude that the customer has no open issues. Deliver a cited draft with questions for the owner, rather than updating the account or emailing a commitment.

Apply responsible AI and data protection¶

Common user-level risks include:

fabrication: confident but unsupported facts, citations, calculations, or events;
prompt injection: untrusted text tries to override instructions or obtain data/actions;
over-reliance: a person accepts the result because it is polished or fast;
oversharing: sensitive source or generated content reaches an unintended audience;
bias and exclusion: source patterns or framing produce unfair or incomplete results;
privacy/copyright risk: personal, licensed, confidential, or third-party material is used inappropriately;
stale context: old policies, files, conversations, or memory shape the result;
purpose drift: a draft or analysis is reused for a higher-stakes decision than intended.

Use a verification plan proportionate to the task:

inspect cited or referenced sources and confirm they support each material claim;
compare critical facts, dates, names, totals, quotations, and policy statements with an authoritative source;
recalculate important numbers and test assumptions;
ask what is missing, uncertain, conflicting, or based on inference;
check audience, tone, accessibility, bias, confidentiality, and sensitivity labels;
obtain a qualified human review before consequential use;
preserve source and approval evidence where the business process requires it.

Sensitivity labels, DLP, permissions, and organizational policy can restrict which content is used, returned, copied, or shared. A restricted result is not a prompt failure to bypass. Remove unauthorized sources, request legitimate access, or use an approved alternative. Never paste confidential information into an unapproved consumer AI tool merely because an enterprise response is constrained.

2. Manage prompts and chats by using Microsoft 365 Copilot¶

Create a prompt with a testable contract¶

A useful prompt specifies:

Goal: the action and business outcome—draft, compare, summarize, analyze, recommend, or plan.
Context: audience, situation, background, constraints, and decision that follows.
Sources: the files, messages, meeting, data, or web sources it may use; state whether it must avoid outside knowledge.
Expectations: output format, length, tone, language, fields, assumptions, citations, quality checks, and what to do when evidence is missing.

Example structure: “Create a one-page management summary for regional directors from the attached approved report. Cover performance versus target, three material risks, owners, and next actions in a table. Cite the report page or section for every number. Do not infer missing values; list them under Questions.”

Begin with the minimum relevant context. Adding every available file can introduce conflicts, stale material, sensitive content, and distraction. Choose authoritative, current, permitted sources. Name the desired viewpoint and audience without asking the model to impersonate a person or conceal uncertainty.

Iterate deliberately. First inspect whether the response followed the task and sources. Then correct missing context, ambiguous instructions, unsupported claims, format, tone, or length. Ask the model to identify assumptions and uncertainties, but verify those statements independently. A longer prompt is not automatically better; clarity and relevant evidence matter more.

Related item: A prompt library becomes business process documentation. Give important reusable prompts an owner, purpose, approved sources, example input/output, validation checklist, version date, and retirement trigger.

Choose the available model for the task¶

The Copilot model selector offers Auto, Quick response and Think deeper: select for routine speed or more involved reasoning, then compare outputs against the same evidence. Model names and availability change by app, license and policy. Longer reasoning does not grant source access or prove correctness.

Researcher model choice has its own conditions: Claude needs administrator enablement, and the newer Auto/Model Council experiences require Frontier access. A second model's agreement is not independent confirmation when both use the same faulty source. Check citations and resolve disagreements in the source material. In Excel, model choice is session-specific and editing can change the workbook directly; use a disposable copy and review changes/version history.

Context engineering means curating the evidence: project, people, dates, source versions, permitted systems and intended audience. Prompt engineering defines the request; context engineering makes its evidence relevant and available. Use the same curated sources when comparing models so a different answer can be investigated rather than mistaken for independent confirmation.

VERIFY CURRENT: Cowork model selection is separate from Chat's selector. Auto chooses from enabled options; device model preferences and effort choices can persist. Higher effort can consume more credits and time without adding source access or proving correctness. Its model/access pages describe tenant-wide Anthropic controls, while the newer admin-governance page describes user/group scoping. Preserve that documentation conflict and check the applicable setting with the administrator; do not assume one universal control scope or freeze a model list.

Save a prompt when the task repeats and the prompt is worth curating. Give it a meaningful title and remove case-specific sensitive content. Retest it with different representative inputs. Schedule a prompt only when the source availability, timing, recipient, review step, and output destination are appropriate for unattended recurrence. A scheduled prompt that generates a weekly report still needs ownership and exception handling.

Share a prompt when colleagues can legitimately use its instructions and sources. Explain prerequisites, expected references, limitations, and verification. Sharing the words of a prompt does not necessarily share source access, licensing, or identical context; recipients can receive different results. Avoid embedding secrets, personal data, customer details, or inaccessible file paths.

Copilot Chat scheduled prompts require an eligible license and the applicable organizational controls. VERIFY CURRENT — documentation discrepancy: that user-help page still names Optional connected experiences, while the administrator reference names Connected Experiences since August 2026. Check the documented admin control and the actual runtime; hiding scheduling does not necessarily cancel existing schedules. The user help lists ten schedules, while Cowork has a separate automation surface listing 25. Inspect each recurrence, run history, permissions and timezone; disable the test definition when finished. These are experience-specific limits, not a universal Copilot quota.

Manage conversation history and notebooks¶

Use the Chats list and its More menu to revisit, rename, move to a notebook, or delete a conversation. Name it so purpose and sensitivity are recognizable. Delete a chat when it is erroneous, obsolete, unnecessarily sensitive, or no longer required—subject to organizational retention and eDiscovery policies. Deleting from the user's visible history is not a promise that every compliance record disappears.

Start a new chat when prior context would confuse or contaminate the next task. Keep one conversation when iterative context is intentional. Before continuing an old conversation, inspect its sources, assumptions, audience, and date.

Add a conversation to a notebook when it belongs to a durable project context with curated sources and repeated work. A notebook is not a dumping ground for every related chat. Remove stale or contradictory material, organize it around the project outcome, and review access before adding sensitive content.

Notebook references define the grounding scope; unrelated organizational content is not automatically included. Adding a site or folder can widen that scope beyond one selected file. The current page distinguishes up to the first 300 grounded references for Copilot users from 50 for Copilot Chat users. A notebook with more references is not evidence that every item influenced the answer.

Sharing can change file access. Notebook sharing invites recipients to linked files when you have permission to share them, subject to source/plan restrictions; otherwise they need to request access. Review every reference before sharing, including sources that did not appear in the latest answer.

Worked example: the summary is harmless but its references are not¶

Alice's notebook contains a public team plan and a synthetic compensation worksheet. Bob needs only the plan. Alice can technically share both files, but the business purpose permits only the plan. Sending the notebook can invite Bob to both references even if the displayed summary mentions no compensation. Create a separate notebook with the plan alone, check its reference list and recipient access, then share that scoped copy. Removing sensitive words from the summary does not remove the referenced file.

Separate memory, history and instructions¶

Memory controls distinguish saved details, inferences from chat/work content, and explicit preferences. Deleting one saved memory does not necessarily remove other personalization sources. Review the relevant controls and custom instructions. The history control documents a 30-day deletion period; switching it back on within that window can restore inferred memories.

A temporary chat avoids using/updating personalization and is absent from visible history, but organizational retention can still apply. This is not a way around records policy. Keep durable formatting preferences separate from changing facts such as this month's target or a customer's approval status.

The current blueprint requires choosing prebuilt versus custom agents, not the former standalone custom-agent construction group. This optional exercise helps explain that choice.

Search the Agent Store before building. An existing agent is preferable when its publisher, task, knowledge, capabilities, permissions, and support fit. Create a new agent when the task is distinct, repeated, and bounded enough to justify tailored behavior.

Templates accelerate creation but still require review. Define:

a precise name, purpose, intended audience, and tasks it must refuse or escalate;
instructions describing workflow, tone, output, boundaries, verification, and uncertainty;
current, authoritative, minimum-necessary knowledge that intended users may access;
capabilities required for the task and no unnecessary action surface;
suggested prompts that teach users safe, valuable starting interactions;
owner, reviewers, test set, support route, and retirement/update triggers.

Test expected, vague, incomplete, sensitive, out-of-scope, conflicting-source, and malicious prompts. Confirm the agent does not reveal inaccessible knowledge or imply authority it lacks. Share first with a small team, communicate limitations, collect failures, and expand only after the evidence supports it. Sharing an agent does not automatically grant users access to every knowledge source.

3. Manage business content and collaboration by using Microsoft 365 Copilot¶

Draft a new document from a prompt¶

Specify document type, audience, outcome, source, structure, tone, length, and required action. Ask for placeholders rather than invented facts. A first draft should expose assumptions and questions. Review organization, factual accuracy, policy alignment, confidentiality, accessibility, and whether the content actually helps the audience decide or act.

For communications, distinguish internal update, customer email, executive brief, proposal, policy, and persuasive message. Tone is not the only difference: authority, evidence, approval, disclosures, and retention vary. Do not let Copilot send a consequential message merely because it created a plausible draft.

Generate from an existing document and create management summaries¶

When transforming an existing document, define whether to preserve facts, terminology, voice, citations, or layout. Ask the response to separate what came from the source from suggestions. Compare the output with the complete document, especially limitations, footnotes, exceptions, tables, and appendices.

A management summary should identify purpose, current state, material results, risks, decisions, owners, deadlines, and missing evidence at an appropriate level. It should not erase dissent, uncertainty, or a low-frequency high-impact risk. Verify every figure and named commitment. If two source sections conflict, report the conflict rather than picking the more convenient one.

Researcher can help collect and synthesize allowed sources for a richer draft. Inspect its citations, publisher authority, dates, source diversity, and whether the conclusion exceeds the evidence. Analyst can explore a data set and produce calculations or visuals; validate schema, filters, units, missing values, outliers, formulas, and the business meaning of correlation or trend.

Generate visuals that answer the business question¶

Choose a chart for comparison, trend or composition; specify units, dates, denominators, filters and how to treat missing values. Generate a concept image only when an illustration is appropriate, and check labels, accessibility, rights and whether it could be mistaken for evidence. A polished image of a chart is not a substitute for a chart tied to verified data.

Worked example: gross growth is not net growth¶

A synthetic workbook has gross revenue of 120 in Q1 and 150 in Q2, with returns of 10 and 30 respectively, all in the same units. Gross growth is (150 - 120) / 120 = 25%. Net revenue is 110 then 120, so net growth is (120 - 110) / 110 ≈ 9.09%. A slide claiming “net revenue grew 25%” is wrong even if its chart exactly matches gross revenue. Name the measure, retain formulas, reconcile the brief and deck, and show missing periods as missing rather than zero.

Move data and insights across Microsoft 365 apps¶

A useful workflow often crosses app boundaries:

summarize a Teams meeting and confirm decisions/owners;
turn approved decisions into an Outlook follow-up;
analyze a structured Excel table and validate totals;
create a Word management brief using the approved analysis;
build a PowerPoint narrative for a defined audience;
place working content in a Copilot Page for collaboration.

Each transition can lose source context, formatting, permissions, sensitivity labels, or nuance. Recheck audience and access at the destination, keep a link to the authoritative source, and distinguish live data from a copied snapshot. Do not move restricted data into a broadly shared page or deck just because Copilot makes the transfer easy.

Use Copilot for meetings and collaboration¶

Before a meeting, use allowed documents, email, and prior action items to develop an agenda and questions. During or after a meeting, Copilot can summarize discussion, decisions, disagreements, unanswered questions, owners, and follow-ups when transcript and policy settings support it. Participants should know recording/transcription practices, and a human should confirm consequential decisions.

Ask for evidence: “Which transcript passage supports this decision?” Verify names, due dates, and speaker attribution. Silence is not consent, and absence from a transcript is not proof that something did not occur. Handle sensitive HR, legal, customer, and security meetings according to policy.

Copilot Pages turns a response into an editable shared workspace. Use it to refine plans, briefs, research, or meeting outcomes with colleagues. Before sharing, remove irrelevant private context, verify facts, set the right audience, and assign ownership. Edits can make the page diverge from its generated sources, so retain provenance for important claims.

Distinguish the three meeting experiences¶

Experience	Useful task	Required check
Private Copilot catch-up	Ask what you missed or identify disagreement during a meeting	Available spoken context depends on meeting settings; private prompts differ from shared notes
Facilitator	Shared live notes, agenda tracking and meeting-chat assistance	License, meeting type, organizer/presenter controls, and participant visibility
Intelligent recap	Review notes, tasks and supported recording markers afterward	Transcript/recording, language, license and permissions determine available features

Real-time catch-up can work in the “Only during” mode without saved transcription. Its help describes transcript-based access to spoken discussion afterward; chat-only context is different. VERIFY CURRENT — meeting-memory exception: the recap reference separately documents AI-generated meeting memory in Teams TAP preview, subject to organizational settings. Copilot/Facilitator can use that memory without a saved transcript. Check whether the citation identifies memory or transcript; memory is not a verbatim record. Deleting a transcript or recording does not itself delete meeting memory. The recap-deletion operation is broader, permanent and not a learner cleanup shortcut. Do not infer that this preview is available in every tenant. Protected meetings can also restrict exporting responses.

Facilitator requires a Copilot license to add/enable it; internal participants can see shared updates. The scheduled online-meeting route excludes channel meetings, instant meetings and calls; the documented mobile in-person capture is a separate experience. Its Loop notes are editable, so they are not an immutable transcript. Task management by @mention is currently public preview. Automatically captured tasks require accepting sync; directly requested tasks can sync to Planner. Check task owners and dates before treating them as commitments.

Intelligent recap is available through Teams Premium or Copilot. AI summaries need a supported-language transcript and an event at least five minutes long; some recording markers need both recording and transcription. A copied recap link still requires access; explicitly sharing a recap can grant it. Review the audience and any draft summary email before sending.

Worked example: incomplete capture is incomplete evidence¶

A synthetic 40-minute meeting starts at 10:00, but transcription starts at 10:20. That transcript covers only the last 20 minutes. A conclusion based only on it cannot establish that no budget decision happened in the first half. If preview meeting memory is cited, identify its separate capture and generated-evidence limitations rather than assuming the transcript defines its entire scope. Record the coverage gap, ask the owner to confirm the decision against an approved record, and distinguish that confirmation from the generated recap.

Memory and instruction settings are covered in Section 2. Avoid assuming the same personalization or chat-deletion controls exist in every app: Excel's current help does not offer individual conversation deletion.

Related item: The value of Copilot is a better business outcome, not maximum generated volume. Measure time saved together with correction rate, decision quality, adoption, risk, accessibility, customer or employee impact, and rework.

4. Drive business outcomes by using agents¶

Select an agent and return to its work¶

For a supplier-research brief, open an approved Researcher agent from the agents experience available to you in Microsoft 365 Copilot, select permitted sources, and ask for a cited comparison with missing evidence listed. For a task started inside an app, check the particular agent's supported channel before selecting it; an agent available in chat is not automatically available in every app. Confirm its identity and source scope, run a small draft task, inspect the result, then reopen that agent's chat to refine it. If it is absent, ask the administrator to check deployment/access rather than substituting an unapproved agent.

Choose a prebuilt agent when its approved task and data scope fit; choose a custom agent when repeatable requirements need different instructions or sources. Researcher fits cited synthesis; Analyst fits deeper data exploration; neither replaces verification. Apply the agent from a supported Copilot surface and confirm its name, publisher and available sources. Agent availability depends on organizational deployment, access and supported channels.

The October blueprint explicitly requires filtering chats by agent. Practice finding a known agent conversation using the agent filter available in your chat-history experience, then reopen it and verify the agent and date before continuing. If the filter is absent in your rollout, record that limitation and use clearly named history as a fallback; the public history article reviewed here does not establish a universal filter location. Filtering history organizes prior work; it does not change permissions or convert an ordinary chat into that agent.

Choose a skill, then supervise the deliverable¶

A skill supplies reusable task instructions; a connector supplies access to a data service; a plugin can package both. These are different from a schedule that decides when work starts. Begin with an appropriate built-in skill. Use a custom skill for a repeated process with specific outputs, boundaries and review steps.

In Cowork Customize, inspect the skill's source and instructions, create or edit it, then test in a new conversation. Shared updates require Re-share. Only import trusted material. The operating guide describes automated skill evaluation, while the application card warns that user-authored skills are not Microsoft-validated: a quality score is not certification that your business process is correct. Test missing inputs, wrong recipients and malicious source instructions.

Delegate a concrete result: scope, sources, file destination, audience, deadline, and what needs review. Inspect the plan, progress, active skills and final artifacts. Not every internal operation has a visible step. Cowork controls distinguish pausing after the current step, immediate pause and cancel; none promises to undo completed work. Check existing outputs before retrying so an interrupted run does not duplicate an action.

Understand the approval you give¶

The action dialog can approve one action, all currently pending actions, or similar actions for the rest of the session. Recipient/domain-scoped options can narrow message approvals. Review and revoke session approvals in the Permissions panel. An existing Office file can be edited in place after approval, affecting collaborators; inspect version history rather than assuming every output is a separate draft.

Worked example: one approval can cover a later message¶

In a synthetic session, Alice authorizes messages only to Bob for the rest of that session. A later message to Bob may not prompt again. That is different from approving only the displayed first message; neither grants a new conversation the same approval. An automation may have separately preauthorized actions, so inspecting only the current conversation is insufficient. Review the actual approval scope and task definition before relying on a future pause.

Manage recurring work and access¶

In Cowork, inspect both schedule definitions and run history; verify timezone, recurrence, source scope, destination and action permissions. “Activate and run now” differs from activating for the next scheduled time. Pause or delete test schedules after verifying one result. Event-driven email/Teams tasks are related context, separate from time schedules.

Organizational governance says automated work runs as its creator and asks before sensitive actions by default, but users can preauthorize actions during setup. Do not assume every unattended run waits for a fresh approval. Cowork access comes through a spending policy selecting the service; a tiny credit allowance still grants access. Model enablement and discoverability are separate controls. Review expected usage and the approved service with your administrator, without treating a small budget as an access denial.

Check the execution surface¶

Cowork can misinterpret instructions, omit steps or use incomplete evidence. Its application card lists protected-file and cloud-file deletion limitations; permissions alone do not establish feature support. Explicit upload is different from unrestricted access to local files.

Local browser tasks use supported Edge on the user's device and its signed-in session. They require tenant enablement and are currently unavailable in the desktop/mobile app. Review the actual site account and let the user complete credentials, MFA or human-verification handoffs. Browser capability does not imply permission to bypass a blocked site or approve a form submission.

Integrated scenarios¶

Scenario A: Weekly executive status¶

A project manager references approved project files and meeting notes. A saved prompt requests progress versus milestones, decisions, top risks, owners, dates, and questions in a one-page format with citations. Analyst checks the structured tracker; the manager independently verifies totals. A scheduled run creates a draft, not an automatic executive publication. The manager reviews sensitivity, corrections, and changes before moving approved content to Word and PowerPoint.

Scenario B: Client proposal team¶

Researcher gathers approved internal case studies and current public client/industry sources. A prompt separates sourced facts, assumptions, and proposed language. The team creates a Word draft, uses a Copilot Page to collaborate, then builds a PowerPoint deck. Every claim, image right, price, capability, and commitment receives owner approval. Restricted internal references do not move into the client deliverable.

The HR team first checks Agent Store, then creates a bounded agent from a template because no approved agent fits. Knowledge contains current employee-facing policies, not confidential cases. Instructions cite the policy, state uncertainty, refuse personal determinations, and route exceptions to HR. Suggested prompts cover common questions. Tests include obsolete policy, conflicting sources, prompt injection, sensitive requests, and two user personas. Sharing starts with a pilot team.

Practical labs¶

Context comparison: Run one business task in Copilot Chat and two relevant Microsoft 365 apps; record which context and capabilities change the result.
Responsible-AI review: Seed ten benign inaccuracies, missing facts, sensitive fields, and malicious instructions into test material; build and apply a proportionate verification checklist.
Prompt clinic: Write and compare at least 12 prompts using goal, context, sources, and expectations; keep evidence of why revisions improved the result.
Prompt lifecycle: Save, rename, share with a test colleague, and—if available—schedule a nonsensitive prompt; document differences in access, context, results, and controls.
Conversation/notebook: Continue and restart the same task, rename and find chats, test deletion behavior, and curate selected work into a notebook without stale sources.
Related item — custom-agent pilot: Compare Agent Store options, create from a template, configure instructions/knowledge/capabilities/suggested prompts, test adversarial cases, and share narrowly.
Cross-app workflow: Move a validated analysis from Excel through a Word brief and PowerPoint deck, tracking source, permissions, labels, figures, and human approvals.
Meeting/collaboration: Use a synthetic meeting transcript to create decisions and actions, verify attribution, publish a sanitized Copilot Page, and review memory/instruction effects. Compare private catch-up, Facilitator notes and recap; retain an explicit missing-transcript example.
Context and sharing audit: Use synthetic files you own and permitted test accounts. Compare notebook reference scope and sharing before/after, a missing connector source, and the gross-versus-net chart example. Record source timestamps, recipients, observed access and corrected calculations; remove the test share afterward.
Supervised delegation: In an enabled test environment, use a built-in skill for a draft brief, then compare a narrowly scoped custom skill in a new session. Inspect single-action/session approval options without sending to real recipients; verify in-place file edits and recovery on a disposable copy. Run one test schedule, inspect its result and permissions, and disable it. Locate the agent's prior chat/filter where available. If access is unavailable, submit a tabletop plan with the limitation rather than claiming execution.
Current-scope evidence worksheet: Use the renewal example to record source authority, date, contradiction and a permission gap. Compare a read-only Researcher connector with a supported write tool and its existing approval scope. For a hypothetical Teams TAP meeting, distinguish transcript citations from meeting-memory citations and explain why deleting a transcript does not prove all generated meeting artifacts are gone. Record unavailable product access as tabletop work.

Use synthetic or approved nonsensitive data. Save prompts, expected outcomes, sources, output, errors, corrections, decisions, and what a human had to add. All eleven product labs remain unexecuted in this refresh.

Knowledge checks¶

Why is fluent output not evidence that a claim is true?
What kinds of context can change a Copilot response?
Why can Copilot-respected permissions still expose an oversharing problem?
When should work stay in an app Copilot instead of Copilot Chat?
How does a chat differ from an agent?
When are Researcher and Analyst appropriate, and what must be verified?
What is the purpose of a Copilot Page and a notebook?
When should you create an agent instead of using Agent Store?
What are fabrication, prompt injection, and over-reliance?
Which verification steps fit a high-consequence business decision?
Why should a user not bypass a data-protection restriction?
How can sensitivity and audience change during cross-app reuse?
What are the goal, context, sources, and expectations in an effective prompt?
Why can too many references make a prompt worse?
How should a prompt handle missing evidence?
What makes iterative prompting deliberate rather than random?
When is a prompt worth saving?
What must be true before scheduling a prompt?
Why can two people get different results from a shared prompt?
When should a new chat replace an old conversation?
What does deleting visible chat history not necessarily change?
What belongs in a curated notebook?
What must be checked before choosing an Agent Store agent?
Which fields define a bounded custom agent?
Why does sharing an agent not automatically share its knowledge?
Which adversarial tests should an agent pilot include?
What belongs in a good first-draft request?
How should a generated document distinguish source facts from suggestions?
What can a management summary accidentally hide?
How do you validate Analyst output?
How do you validate Researcher citations?
What can be lost when content moves between apps?
How should Copilot meeting decisions and attribution be verified?
What must be checked before sharing a Copilot Page?
What information is inappropriate for memory or persistent instructions?
Which outcome measures matter beyond the volume of generated content?
When should model choice change, and what can it never establish?
How do synced and federated connectors differ, and what does a missing result prove?
Why can sharing a notebook expose a reference absent from its latest summary?
Why is the example's net growth about 9.09% rather than 25%?
How do private catch-up, Facilitator and intelligent recap differ?
How are a skill, connector, plugin and schedule different?
Why might Cowork take a later action without displaying a new approval?
What must be checked before retrying or stopping delegated work?
How should a renewal recommendation handle a signed contract, an older proposal and a missing support source?
Can a federated connector always be treated as read-only, or always require a fresh approval?
Does deleting a Teams transcript prove Copilot cannot answer from meeting memory?
Does an agent available in Copilot chat necessarily work inside every Microsoft 365 app?

Answer checkpoints¶

Fluency reflects plausible generation; corroborating evidence establishes the claim.
Prompt, app, conversation, referenced material, authorized work/web sources and personalization.
Existing broad access can make inappropriate source content available without new privilege.
Use the app when its active artifact and editing tools fit; compare available context.
Chat supports an exchange; an agent applies a durable task purpose and configured capabilities.
Researcher supports synthesis; Analyst supports data reasoning. Verify sources, computations and interpretation.
Pages supports collaborative output; notebooks organize the references and conversations for focused work.
Build when approved prebuilt choices cannot meet a repeatable, bounded requirement.
Unsupported output, hostile instructions embedded in content, and accepting output without sufficient scrutiny.
Authoritative source checks, independent calculations, missing-evidence review and accountable human judgment.
The restriction expresses a data/access policy; use a legitimate authorized path.
Copied content can enter a different audience or lose protections; inspect the destination.
Outcome, situation/audience, allowed evidence, and the required format/quality/boundaries.
Irrelevant, stale or conflicting material can obscure the needed evidence.
Identify the gap and request the source; do not manufacture a value or commitment.
Change a specific defect and compare the revision against explicit acceptance criteria.
It repeats, has a clear owner and works on varied representative inputs.
Confirm entitlement, timing, sources, ownership, destination, review and stop behavior.
The prompt text does not reproduce another person's permissions or context.
When earlier context would distort the new task; separately check personalization settings.
Compliance retention and other personalization sources might remain.
Current, relevant, permitted references with a defined project purpose.
Publisher, task, data/action scope, availability, permissions and support.
Purpose, audience, instructions, knowledge, capabilities, prompts, tests and escalation.
Agent access and source access are separate checks; inspect the actual sharing mechanism.
Missing/conflicting/obsolete sources, injection, sensitive requests, wrong identities and out-of-scope actions.
Audience, purpose, evidence, structure, tone, constraints and placeholders for unknowns.
Mark supported facts, uncertainty and proposed additions separately; preserve provenance.
Exceptions, dissent, coverage gaps and important low-frequency risks.
Inspect schema, units, filters and missing values; independently reconcile key calculations.
Open the cited source, check date/authority and whether it supports the particular claim.
Source links, units, meaning, permissions, labels and live-versus-snapshot status.
Check captured evidence and ask accountable participants about uncertain attribution or decisions.
Content, audience, source permissions, protections and ownership of later edits.
Secrets and transient facts that should be rechecked for the current task.
Quality, correction/rework, time, decision usefulness, accessibility and business impact.
Match speed/reasoning and allowed models to the task; choice cannot grant data access or prove truth.
Synced content is indexed; federated content is fetched live. Missing results do not establish absence.
Sharing can invite recipients to linked files the sharer can share, including unused references.
Net values are 110 and 120 after returns; gross values use a different measure and denominator.
Private questions, shared live assistance, and post-meeting review have different visibility and capture requirements.
Instructions, data access, a package of capabilities, and a timing/trigger definition respectively.
A session-wide approval or separately preauthorized automation may already cover it.
Check completed actions, file versions, pending permissions, schedules and duplicates; cancel is not rollback.
Identify authoritative current facts, preserve the conflict and missing evidence, ask the accountable owner, and produce a cited draft without inventing a commitment.
No. Supported tools can modify source data, while Researcher remains read-only. Conversation or persistent tool approval can already cover a later action.
No. The documented Teams TAP preview can retain separate AI-generated memory under organizational settings; inspect the cited artifact and relevant lifecycle controls.
No. Verify that agent's supported channel, tenant deployment and user access; availability on one surface does not establish another.

Useful blog reading with an exercise¶

Introducing Copilot Home, Code and Autopilot — Jared Spataro, September 25, 2026. Classify a lookup, a delegated brief and recurring work by required context, actions and supervision. The article announces staged Frontier/private previews; it does not establish that Home, Code or Autopilot is available in your tenant or add those product names to AB-730's objectives.
Work IQ business and workplace intelligence — James Oleinik, September 25, 2026. Draw a renewal-evidence table linking opportunity, unresolved case, contract and customer message; identify the missing source or approval that would change the recommendation. The article described a September 30 preview start, October rollout and planned late-October Finance/Operations support. These dated announcements do not verify delivery or current tenant availability. Use a tabletop exercise until your environment has the supported feature.

Both public article texts were read. Marketing performance claims, videos, customer deployments and preview access were not independently validated. Use their scenarios to ask better questions, then verify feature behavior in current product documentation.

Final readiness checklist¶

Places to learn¶

This is a selective starting set, not a complete list and not a requirement to consume everything. Pick the explanation, hands-on practice, and assessment style that works for you, and map each resource to the blueprint effective for your appointment. Selected public course/path listings were rechecked October 28; unchanged older observations remain dated below. Public landing-page access does not establish access to paid lessons or assessment questions.

Resource	Access	Estimated time
Official self-paced path	Free public reading; product practice needs eligible access	4 modules; about 4–6 hours including notes (editorial estimate; no directly verified total)
Focused Work IQ foundations module	Free public reading	About 25–40 minutes for reading and a context worksheet (editorial estimate)
Focused Cowork module	Free public reading; eligible product access for practice	About 1–2 hours for selected units and a supervision worksheet (editorial estimate)
AB-730T00-A instructor-led course	Paid or partner-sponsored	1 day
Microsoft Practice Assessment	Free	45–75 minutes per attempt plus review
Microsoft exam-prep video	Free	Public page retrieved; runtime/video content not independently reviewed
Pluralsight AB-730 path	Paid	4 rounded hours / 3 courses (4h26 summed); path still in production
O'Reilly AB-730 Crash Course	Paid	About 3 hours of agenda plus breaks/exercises
MeasureUp AB-730	Paid	110 questions; allow 5–8 hours across remediation
Udemy Scott Duffy practice	Paid	100 questions advertised; about 3–6 hours for attempts/remediation (editorial estimate); paid questions uninspected
Partner Skilling Hub	Partner-restricted	Event-specific; verify signed-in start/end times

Primary route: Microsoft Learn AB-730T00 and the linked Enhance business workflows with AI path. It now lists four modules: content, data/visuals, meetings/collaboration, and Cowork. It is no longer the earlier six-module listing; use the full course and objective map to identify remaining fundamentals and prompt-management needs.
Targeted current-scope reading: Get ready to work with Microsoft Copilot introduces Work IQ, protection and prompting. Get more done with Copilot Cowork has ten units; selected how-it-works, guardrails and a 15-minute exercise entry were inspected. Use synthetic data and a plan/progress/result worksheet. The course's simple approval description must be read alongside product documentation on existing session approvals and preauthorization; the module is not proof that every action prompts again. No exercise launcher or tenant task was run.
Readiness: free official Practice Assessment and official AB-730 prep video. Use the assessment diagnostically; investigate every miss in first-party documentation.
Structured subscription path: Pluralsight AB-730 now lists three courses totaling 4h26: 1h16 fundamentals (August 22), 1h45 prompts/conversations (September 11) and 1h25 business-content training (October 6), plus a practice exam. The path remains in production and its three-domain description does not establish complete October 20 coverage.
Live guided preparation: O'Reilly AB-730 Crash Course by Renaldi Gondosubroto has a 180-minute teaching/exercise agenda plus breaks. It still describes three domains and custom-agent construction; verify October coverage, the actual event date and local start/end time before enrolling.
Paid assessment: MeasureUp AB-730, 110 questions released June 2026. Its generic overview text incorrectly mentions Azure AI workloads, while its detailed mapping follows the earlier three-domain AB-730 scope. Neither establishes coverage of the new October agent domain; verify revision coverage before purchase.
Additional assessment listing: Udemy AB-730 practice by Scott Duffy and Jordi Koenderink, public listing rechecked October 28 advertises 100 questions in four sets and an August 2026 update mapped to July 22. September automation was blocked; current browser-readable metadata does not establish October 20 coverage, question originality or paid-content quality. Recheck before buying.
Partner-restricted learning: Partner Skilling Hub; sign in to confirm current AB-730 delivery, exact start/end times, seats, and prerequisites.

No exact Whizlabs AB-730 product was independently verified in the earlier September 1 catalog review; this deep review did not establish a new product listing. Avoid unusually large banks, “real questions,” pass guarantees, or content that claims to mirror live questions. A small set of original scenarios plus hands-on Copilot use is more useful than memorizing hundreds of unsupported answers.