Copilot in Word: How We Turned 10 SharePoint Docs Into a Review Report
Ten SharePoint documents. One business review at 4:00 PM. And not one person on the team who has read all ten. That is not a reading problem, it is just an ordinary Tuesday.
So we handed the pile to Microsoft Copilot in Word and treated it like a real project. We set it up, pointed it at the documents, squeezed out the facts that mattered, and built three different versions of the report so the manager could pick a winner. Along the way we learned that AI never asks "are you sure?", which is both its best feature and its most dangerous one. Here is the whole run, step by step, with the prompts, the slip-ups, and the checklist we now use every time.
The Setup: Ten Documents, One Manager, One Deadline
Our sample project was a customer onboarding portal rollout. The manager wanted a clear report before the review: where we are, what went wrong, and what decisions we need from the room. The source material lived in one SharePoint library, and a few documents quietly disagreed with each other.
| # | Document | What it holds | Hidden trap |
|---|---|---|---|
| 1 | Project charter | Goals, scope, original go-live date | Date later changed |
| 2 | Scope and change request log | Approved and rejected changes | One change affects the budget |
| 3 | Weekly status, week 6 | Progress and blockers | Shows the old go-live date |
| 4 | Weekly status, week 7 | Latest progress | Says one risk is closed |
| 5 | Budget tracker summary | Planned vs spent | Total differs from the minutes |
| 6 | Risk register | Risks, owners, status | Same risk still marked open |
| 7 | Vendor update emails | Delivery news, saved as a document | Casual, shorthand wording |
| 8 | Test results summary | Pass and fail results | A failure missing from status reports |
| 9 | User feedback summary | Pilot user comments | Mixed opinions, easy to cherry-pick |
| 10 | Steering committee minutes | Decisions and open questions | Budget figure differs from the tracker |
The traps matter. Real project folders always have a few, and they are exactly where a careless AI draft goes wrong.
Our Game Plan in One Picture
Before touching any prompt, we agreed on the order of work. Look at how the AI only appears in the middle, with people on both ends.
One folder, clear names, permissions checked
License, working file, reference the folder
Inventory every file and flag conflicts
Boil it down to one sourced fact sheet
Same facts, three different reports
People check the facts and choose the winner
Here is the same Copilot workflow in video form, from ten messy files to a fact-checked report.
Step 1: Getting the Documents Ready in SharePoint
Copilot can only work with what it is allowed to see. As Microsoft describes it, it uses files you already have permission to open, and those files need to live in OneDrive or SharePoint. A file on your desktop is invisible to it. So the first job is tidying, not prompting.
| Prep step | Why it matters | What we did |
|---|---|---|
| One folder for the project | Easy to reference as a group | Created "Business Review - Source Docs" in the project library |
| Keep the folder to what you need | When a folder holds more than ten files, Copilot has been reported to use only the ten most recent | Kept exactly our ten documents, moved the rest out |
| Clear file names | Lets you tell Copilot which file to trust | Renamed anything called "final2" to something meaningful |
| Check access | No access means Copilot cannot read it either | Opened all ten once as ourselves |
| Working document in the same library | Copilot works best on cloud-saved files | Created a blank "Business Review Report - Working Draft" |
One more thing before you paste anything anywhere: project folders can hold sensitive material. Confidentiality labels and company policy still apply, so give our list of things to keep in mind for safer AI use a quick read first.
Step 2: Setting Up Copilot in Word
This is the part everyone expects to be hard, and it is mostly checking boxes. Menus and names change often, and not every person on the same plan sees the same buttons, so treat our steps as a map and trust your own screen.
We attached our ten-document folder and left the page blank. That was our "everything on the bed" moment.
Step 3: Consuming Ten Documents Without Reading Ten Documents
Our first instinct was to ask for the report immediately. We resisted, and that single decision saved us. Garbage in the middle means garbage in the report, so we asked Copilot to show its homework first. If you want the thinking behind this, our piece on giving AI the right information at the right time explains why the order of work matters so much.
Using only the files I referenced, list each file by name with: its purpose in one line, its owner and date if shown, and its two most important facts. Then list anything that conflicts between files. Do not guess, and do not merge conflicting numbers into one.
It came back with a tidy inventory and, to our relief, caught the main conflicts. It also had one slip worth knowing about.
| What we found | Where it came from | How we handled it |
|---|---|---|
| Two different go-live dates | Charter and week 6 status vs week 7 status | Used the newest date, noted the move as a slip |
| Two different budget totals | Budget tracker vs steering minutes | Asked the project owner, then marked the answer |
| A risk both "open" and "closed" | Risk register vs week 7 status | Checked with the risk owner |
| One failed test not in the status reports | Test results summary | Added it to the report as a known issue |
| Slip: the inventory covered nine files, not ten | The vendor emails never appeared | Asked "which files did you actually use?" and re-ran |
The lesson is a cheap one: after any multi-file prompt, ask Copilot to list the files it used. A confident answer does not prove it read everything.
Step 4: Distilling Ten Documents Into One Fact Sheet
Now came the important squeeze. We wanted one sourced fact sheet that every report version would be built from, so all three versions would agree on the facts and only differ in style. The funnel below shows how the material shrank at each stage.
pages of everything
two facts per file
sourced, conflicts marked
one per style
From those ten files, build a fact sheet for a business review: current status, milestones hit and missed, budget planned vs used, the top five risks, and decisions needed from the manager. One line per point, and name the source file after each line. Mark anything uncertain with [CHECK].
The "name the source file" rule is the star here. It turns every line into something a person can verify in seconds. If you do this kind of condensing often, our walkthrough on using AI to summarize documents has more patterns, and if the input is mostly meetings, automating meeting notes with AI helps upstream.
We then did the human part: resolved every [CHECK] with the project owner and fixed the fact sheet by hand. That cleaned-up sheet became the only source for everything that followed.
Step 5: Building Three Versions (The Funny Part)
Here is something nobody warns you about. A human colleague, asked for a fourth rewrite, would sigh and ask what you actually want. Copilot never does. It has no ego, no lunch plans, and no opinion about whether the last version was fine. So we used that as a superpower: same fact sheet, three completely different reports.
Each version used the fact sheet plus one twist. These were the prompts, which you can copy and adapt.
Version A: Using only the fact sheet, write a one-page business review for a busy manager. Sections: Status, Wins, Problems, Decisions Needed. Bullets only. Keep it under 250 words.
Version B: Using only the fact sheet, write a numbers-first report. Open with a table of planned vs actual for schedule, budget, and scope, then three short paragraphs explaining each gap. Do not add any figure that is not in the fact sheet.
Version C: Using only the fact sheet, tell the story of this project in plain language: where we started, what went sideways, how we recovered, and what we need next. No jargon and no new facts.
How We Refined: A Conversation With No Pushback
Version A did not come out right the first time, and Copilot never complained once while we fiddled. That is the fun and the risk. Here is roughly how the refinement went.
Four rounds in, a human would have asked us to make up our minds. Copilot just kept going. The final prompt worked because it named a tone, a rule about the delay, and a lock on the facts. If prompt wording is your weak spot, our guide to writing better AI prompts covers the same habits in more depth.
Version C had its own comedy. We asked for a story and the two-week delay came back as "a bold recalibration of the timeline". Charming, but the manager would have spotted the spin instantly. We changed the prompt to "state delays plainly" and it behaved.
What the Manager Saw
We put all three side by side and let the manager score them. The ratings are the manager's impressions in this walkthrough, not a benchmark.
| Question | A: Executive Haiku | B: Numbers Nerd | C: Storyteller |
|---|---|---|---|
| Can I scan it in one minute? | Strong | OK | Weak |
| Are the numbers clear? | OK | Strong | Weak |
| Is it honest about delays? | Strong | Strong | OK, after the fix |
| Would a newcomer understand it? | OK | OK | Strong |
The verdict was the most human outcome possible: the manager picked B as the base, borrowed A's "Decisions Needed" box, and asked to steal the opening paragraph of C for the intro. The AI built three drafts, and a person assembled the best one. That is how this is supposed to work.
Step 6: The Human Audit Before It Leaves Your Desk
An AI report that looks finished is the most dangerous kind, because the polish hides mistakes. Copilot can state things that are not in your documents with complete confidence, which we unpacked in our look at AI hallucinations. So we ran this checklist, and we now run it every time.
| Check | What to look for | Done |
|---|---|---|
| Every number and date | Matches the fact sheet and its source file | ☐ |
| Conflicts resolved | No leftover [CHECK] tags, no silently merged figures | ☐ |
| Bad news still visible | Delays, failures, and risks not softened into jargon | ☐ |
| Nothing invented | No new names, commitments, or "benefits" you never wrote | ☐ |
| Source coverage | All ten files represented, not nine | ☐ |
| Formatting and template | Headings, fonts, and spacing match your company style | ☐ |
Beginner to Advanced: Where to Start With Your Own Documents
You do not need ten files and a looming deadline to try this. Start small and climb.
Level 1, Beginner: reference one document and ask for a two-line summary plus open questions.
Level 2, Intermediate: reference three to five files, build a sourced fact sheet, and resolve the conflicts yourself.
Level 3, Advanced: reference a whole folder, produce multiple versions for different readers, and save your prompts as a reusable template.
Level 4, Beyond Word: when ten documents become hundreds, build a dedicated assistant over your SharePoint content.
Level 4 is a different tool entirely. If you are curious, our guide to the Copilot Studio knowledge base shows how an agent can answer questions from your own documents. And if the manager asks for slides next, our look at AI and PowerPoint problems covers that handoff.
Limitations and When Not to Use This
- It can only use what it can reach. Local files, files you cannot open, and some protected files may be skipped without much warning.
- It may not read everything. Our inventory dropped one file. Always ask which files it used.
- Folder and file limits exist and change. Reported limits on items and recent files have shifted over time, so check your own version.
- Conflicts get smoothed over. Unless told otherwise, it may quietly pick one number. Tell it to flag, not merge.
- Tone drifts in both directions. Too bold, then too timid, and sometimes spin.
- Do not use it for legal, financial, or confidential-by-policy material unless your organization has cleared the tool and the content.
- This was one walkthrough. Your files, formats, license, and Word version can behave differently, so rehearse on low-stakes documents first.
Final Thoughts
Copilot did not write our business review. It read faster than we could, sorted what it found, and drafted three starting points, while the people did the judging, the fact-checking, and the choosing. The best habit we picked up was boring and powerful: build one trusted fact sheet first, then let the AI dress it up as many ways as you like.
Frequently Asked Questions
Can Copilot in Word really read my SharePoint documents?
Yes, if you have a Copilot-enabled plan and permission to open the files. They need to be stored in SharePoint or OneDrive, and your organization's settings can limit what it can see.
How many documents can I reference at once?
It has changed over time. Microsoft has raised the number of items you can reference and added folder support, with a note that large folders use the most recent files. Check your own version before relying on a number.
Why did Copilot skip one of my files?
Possible reasons include permissions, file type, size, location, or limits. Ask it to list the files it used, then fix the cause and run it again.
Why build several versions of the same report?
Different readers need different things. One version for a quick scan, one for numbers, one for the story gives your manager a real choice and often leads to a better blended final report.
Does Copilot ever say my prompt is bad?
Rarely. It tends to keep going rather than push back, so you have to be the one asking whether the output still matches the facts.
Can I do this with another AI assistant instead?
Yes. The method (inventory, fact sheet, versions, audit) works with any capable assistant. Copilot's advantage is that it works directly with your Word and SharePoint files.
Is it safe to use confidential project documents?
That depends on your organization's policy and settings. Follow your company rules and never assume a tool is cleared just because it is available.
Comments
Post a Comment