Compare files without hiding conflicts, publish checked SharePoint pages, test agents at their limits, and update every artifact when a source changes.
What you'll learn
Pick the right OneDrive or SharePoint move for a file task
Test a SharePoint agent against conflicting and out-of-scope questions
Review Planner tasks, Forms questions, and video answers as drafts
Propagate a source change through every dependent artifact
An onboarding coordinator opens a remote-work policy and a manager
checklist. One says security training must finish before remote
access. The other says by Friday of week one. Both documents are in
use, and they disagree. Once the work spans more files, a SharePoint
site, a plan, and a feedback form, Copilot's job is to help move it
along without smoothing over a fact, conflict, or source.
Decide what you're asking Copilot to do
Before writing the prompt, identify the source, operation, output, and
check:
Source scope
Operation
Artifact
How you verify
One file
Summarize
Structured summary
Match each claim to the file
Two related files
Compare
Difference table
Check both sides, keep conflicts
A set of files
Ask about a topic
Bounded answer
Confirm each answer is in scope
One PDF
Review
Page-linked findings
Match each finding to a page
SharePoint content
Create or rewrite a page
A publication
Check facts, links, dates, placeholders
Selected knowledge
Configure an agent
A reusable Q&A
Test in and out of the boundary
Source scope defines the evidence. The operation says what should
happen to it. The artifact is what you'll review or share. The last
column tells you how to check the result, before confident wording is
mistaken for a correct answer.
A qualifying plan isn't the same as an enabled feature
As of mid-2026 the Microsoft 365 Copilot add-on is listed at $30 per user per month for enterprise, paid yearly, on top of a qualifying base plan. Having a qualifying plan lets your organization buy Copilot. Whether every capability here is switched on for your account is a separate, unconfirmed question.
OneDrive is useful when the evidence is in files
OneDrive Copilot can summarize a file, compare two files, answer a
question across several, and review a PDF with page references.
Comparison and PDF review are especially useful because both expose
where a claim came from.
When you compare documents, keep their disagreements. Ask for one
column per document and a "Missing or conflicting information" column.
Tell Copilot not to decide which document wins. For the coordinator,
"before access" and "by Friday" belong in the same row. The conflict
is the finding, and the content owner needs to resolve it.
For a PDF, request deadlines, attachments, approval steps, and exceptions
with a page number for each. Open the cited pages. If you can't find the
support there, don't keep the finding.
Maintain an evidence log as the work moves between artifacts. Record
the claim, source, and status. Use Supported when the source backs the
claim, Conflict when in-scope sources disagree, and Missing when
none of them answers. A reviewer can then inspect the reasoning without
reading every file from the beginning.
Compare two files and keep the conflict· other
Bad example
Compare these two onboarding files and summarize the right process.
Good example
Compare these two onboarding files. Return a table with columns Topic, Document A, Document B, Practical impact, and Missing or conflicting information. Write "Not found" when a document is silent, and do not choose which document is authoritative. Preserve every disagreement.
Why this works: A side-by-side table where each conflict (like "before access" versus "by Friday") stays visible for a human to resolve.
SharePoint pages publish. Agents answer questions
A SharePoint page gives readers one curated, reviewed explanation.
A SharePoint agent answers different questions from the knowledge
you've selected. Use a page for a stable message. Use an agent when
people need a range of answers within a defined subject.
Review a page like any publication. Check facts, links, owners, dates,
and placeholder text. When Copilot drafts it, ask for missing links and
dates to appear as "To be confirmed" so they remain visible during review.
Test an agent differently. Its boundary matters as much as its answers:
Test
Question type
Correct behavior
In scope
Answerable from a source
Answers and names the source
Conflict
Two sources disagree
Reports both, picks no winner
Missing
No source has the answer
Says the sources don't answer it
Out of scope
Unrelated to its knowledge
Declines to give a figure
Boundary
Only partly covered
Answers the covered part, flags the rest
Ask the onboarding agent, "What's the travel reimbursement limit?" The
selected sources don't cover it, so the agent should say that. A $500
answer is a failed boundary test. If a test fails, change either one source
or one instruction, then rerun the full matrix. Changing both leaves you
unable to tell what fixed the behavior.
Draft a SharePoint page with visible gaps· sharepoint
Bad example
Make a complete SharePoint onboarding page for new managers.
Good example
Create a SharePoint page titled "New Manager Onboarding" for managers in their first 30 days. Include a first-week checklist, remote-access requirements, key contacts, source links, and unresolved questions. Mark every missing link, owner, or date "To be confirmed", and never fill a gap with a plausible guess.
Why this works: A reviewable page draft with no hidden placeholders: every gap is visibly labeled "To be confirmed" instead of disguised as finished content.
Instruct a bounded SharePoint agent· sharepoint
Bad example
Create a SharePoint agent that answers new-manager onboarding questions.
Good example
You are the New Manager Onboarding Assistant. Use only the selected onboarding sources. For each answer give the direct answer, the required action, the owner or contact, and the source. If the sources don't answer, say "The selected onboarding sources do not answer this." Report any source conflict without choosing a winner.
Why this works: An agent that answers inside its knowledge, names its sources, surfaces conflicts, and declines cleanly when a question falls outside the material.
Change one variable at a time
When an agent test fails, adjust either a source or an instruction, never both, then rerun every test. Otherwise a green result tells you nothing about which change earned it.
Plans, forms, and video answers all need review
In Planner, Copilot proposes a plan. Keep each task to one observable
action. "Approve the handbook and schedule orientation" needs two tasks.
Put prerequisites in order as well. Feedback collection follows the
workshop, so the plan should show that dependency. Owners and dates remain
unconfirmed until the responsible person accepts them.
Copilot can draft a survey in Forms, but you still own every question.
Check that each one measures a single idea, uses neutral language, and has
answerable options. "How useful and engaging was this excellent workshop?"
is leading, asks about two qualities, and gives no meaning to an unlabeled
1–5 scale. Ask "Overall, how useful was the workshop for your work?" and
label 1 "not useful" and 5 "extremely useful."
Video covers two separate workflows. A video created with Copilot is a
draft to check against the communication brief. Questions about an existing
video must be answered from that video. If you ask "what is the HR team assigned to do?"
and the video never mentions HR, the answer is "Not found."
Draft a Planner plan of proposals· other
Bad example
Create a four-week onboarding plan in Planner with owners and dates.
Good example
Create a four-week plan to launch an employee onboarding program for a small HR and IT team. Include preparing the handbook, configuring access, scheduling orientation, collecting feedback, and reviewing improvements. Use one observable action per task, sequence prerequisites before the work that depends on them, and keep any owners or dates as proposals to confirm.
Why this works: A sequenced draft plan with one action per task and prerequisites in order. Owners and dates are marked as proposals awaiting confirmation.
Keep the evidence intact as the work moves
A decision may begin in a Copilot Chat comparison, move to a
Page, inform a
Notebook synthesis, and appear later in a Word brief and a deck. Two
checks matter across that chain.
First, confirm picker eligibility. A source available in Chat may not
appear in a Notebook. Check the sources each surface actually offers. If
a file isn't available, copy the exact passage into a verified extract and
record where it came from. Don't claim Copilot used a source it never saw.
Second, follow change propagation. If a feedback count changes from
34 to 29, update the Page, synthesis, document, deck, and any narration
that cited it. Keep a change log of the affected artifacts so the stale
value doesn't survive on slide 3. Then revisit the conclusion. If 29 now
ranks second, the "most-selected resource" claim also needs to change.
Try it yourself
Run the agent boundary test
Use three fixtures. Policy: "Security training must finish before remote access. Equipment requests are due five business days before the first remote day. The equipment-request owner isn't specified." Checklist: "Complete security training by Friday of week one." FAQ PDF: "The manager sends the completion record to Security Desk, which enables access within one business day." Write the expected agent answer for each question below.
01
In scope: "Who enables remote access?" Expected: Security Desk, citing the FAQ PDF.
02
Conflict: "When is security training due?" Expected: report both "before remote access" (policy) and "by Friday of week one" (checklist), choosing no winner.
Hint: A conflict is reported, never resolved silently by the agent.
03
Missing: "Who owns the equipment request?" Expected: state that the sources don't name an owner. Invent none.
04
Out of scope: "What's the travel reimbursement limit?" Expected: decline, because no source covers reimbursement.
Four written expected answers where the agent cites its source, preserves the training conflict, marks the equipment owner missing, and refuses the reimbursement question.
Key takeaways
Choose source scope, operation, artifact, and verification before you prompt. The last one is what most people skip.
When comparing files, keep conflicts visible and send them to a content owner. Never let Copilot reconcile them into one rule.
A SharePoint page publishes one reviewed message. A SharePoint agent answers many bounded questions, and you test an agent at its boundaries.
Planner tasks, Forms questions, and video answers are all drafts. Use one action per task, neutral labeled questions, and "Not found" when a video is silent.
A source Copilot can use in one surface may be unavailable in another. When a value changes, propagate it everywhere and reconsider any conclusion built on it.
Check your understanding
1. You need one curated, reviewed explanation that readers will consume the same way every time. Which fits best?
2. Your onboarding agent, grounded only in onboarding files, is asked for the travel reimbursement limit. What's the correct behavior?
3. Which survey question is ready to use as written?
4. A generated plan contains one task: "Approve the handbook and schedule orientation." What's the problem?
5. A feedback value you already used in a document and a deck changes from 34 to 29. What must happen?
Frequently asked questions
Terms used in this lesson
source scope
The specific set of files, pages, or knowledge you allow Copilot to use for a given task.
evidence log
A running record of each claim, its source, and its status (supported, conflict, or missing) across every artifact.
SharePoint agent
A conversational Copilot experience grounded in selected SharePoint knowledge that answers bounded questions.
picker eligibility
Whether a Copilot surface offers a particular source for selection. Eligibility in one surface doesn't imply another.
prerequisite
Work that must finish before a dependent task can begin. The relationship between the two is a dependency.
