Instructor Guide

Welcome to a hands-on session with Copilot on hotel performance.

The strip shows the seven stages we will work through: orient, question, scan, connect, chat, challenge and verify.

Copilot can help analyse. The hotel team verifies.

Let me introduce today’s hotel. Riverside Grand is a made-up hotel with 220 rooms, and you are its Hotel Operations Coordinator.

There are five workbooks covering the first three months of 2027: daily performance, guests, facilities, service and reservations. Every value in them is synthetic.

Eight pressures show up across the files, from uneven demand to reservation friction. Your job is not to fix them today. It is to ask better questions and trace every answer to a row.

Copilot can help analyse. The hotel team verifies. In the next hour you will use Excel + Copilot Chat, and learn to spot when an answer goes beyond the evidence.

This is Riverside Grand in numbers, all from file 01.

Occupancy averaged 73.7 percent over 90 days, the average guest rating was 4.39 and there were 134 complaints.

Every label has a plain word beside it, so read that word out. These are averages and totals, not a verdict.

You are the Hotel Operations Coordinator. The general manager wants to know where the hotel is under pressure, and what to look at first.

You will make three things: a performance scan, an evidence grid and a checked briefing.

Two modes run all lesson: Copilot in Excel on one file, and Copilot Chat on the web with one or two files and no Excel open.

Five focused files: each has a job and a warning, so read the card before you open the file.

The conflict at the bottom is real: file 02 says groups have unequal sample sizes, but its own summary shows 36 responses in every segment.

We say what the rows show.

Now we get the files.

Click a card to download one workbook, or take all five in the ZIP.

Everything is made-up data, so it is safe to download.

Copilot in Excel needs the workbook saved in OneDrive or SharePoint, with AutoSave on.

Create a folder called Hotel Performance and upload the five files.

Your setup may look different, so check in your own environment.

Before we ask Copilot anything, we look at the file.

File 01 has three sheets and 90 daily rows. Look at the Day column: it does not match the dates, so we quote dates, not weekdays.

A good question says what you will decide, what you will compare, and what the file cannot show.

Click one choice in each group and watch one sentence take shape.

Read the rule at the bottom aloud: describe timing and observed relationships, and do not say caused by unless the workbook supports it.

Now the other side. A leading prompt asks Copilot to agree with you.

Click the leading prompt in each pair, then read why.

A useful prompt asks Copilot to test the idea, with the evidence for it, against it and missing.

Open the workbook in Excel and click Copilot.

Type a forward slash to see how you point at a file.

The workbook must be in OneDrive or SharePoint. Check the exact method in your own environment.

This is Try Out 1, in Copilot in Excel on file 01.

Paste Try this first, to map the file. Then paste Then ask, to scan performance.

Read the Verify line before you send anything.

Now we check Copilot against the rows.

Click Verify with workbook on each day and compare the source row with what Copilot wrote.

A status chip is earned by the row, not by how sure Copilot sounds.

Now reveal the two discussion cards, one at a time.

Day 23 is 23 January, and Day 58 is 27 February. They have similar demand and opposite service.

A quiet day can still have a queue, so hotel-wide occupancy is an incomplete measure.

This is a checkpoint, a pause to see where we are.

We have oriented, questioned and scanned.

Next we connect performance to service.

Try Out 2 is in Copilot in Excel again, now on file 03.

Stay on this one file. It has three related sheets: usage, downtime and maintenance.

Paste the Try prompt, then the suggested follow-up, which is not from the course pack.

Copilot has built an evidence table for three claims.

Look at each claim for evidence for it, evidence against it and what is missing.

A claim with no contradicting evidence is not finished.

Now the evidence grid: click each claim to read all seven fields and give it a status.

Then reveal the pool trail and the arrival congestion.

Both warrant investigation, and neither proves a cause.

Another checkpoint.

We have an evidence grid with a status for every claim.

Next we move to Copilot Chat and brief the manager, without opening Excel.

Try Out 3 is a different mode: Copilot Chat on the web, with Excel closed.

Type a slash, attach files 01 and 03, then paste the Try prompt.

Two files at most. The approved pairs are 01 and 03, 02 and 03, or 02 and 04.

Here is the one-page briefing, point by point.

Tick a point only when its sheet and row are named. For the rest, use Verify later in the workbook.

If Chat cannot open the file, use the prepared output and keep discussing the evidence.

Third checkpoint.

We have a source-linked briefing.

Next is Refine, and it is mandatory.

Refine is mandatory. We challenge a draft that sounds sure of itself.

Copy the draft, paste it into a new Copilot Chat with files 02 and 03, then paste the Refine prompt.

The follow-up is a suggestion, not from the course pack.

Click Highlight unsupported language and read the labels.

Each highlighted word claims a cause, a certainty or a value that the rows do not show.

Then read the supported rewrite, and check its figures yourself.

Six cards: sample size, proxy bias, confirmation bias, missingness, causation and operational context.

Click each one to test the draft against the files.

Then reveal the counterexample and the bias correction.

This stretch is optional. It is counted in the 75 minutes.

Open file 05 in Excel, and paste Prompt 7.

If you are behind, skip it and give the time back to Refine.

Compare rates by channel, not counts.

The extreme payment failure is real, but it sits in the channel with the lowest error rate.

One status conflict stays visible, because the fields agree and only the note disagrees.

Final verification is mandatory, and it is short.

Paste your revised analysis into Copilot Chat with files 02 and 03, then Prompt 8.

Copilot checks every number, date, record and recommendation.

Read the verdict. It is Not Ready.

Two figures and one record ID have no source location, and the contradicting evidence is missing.

Final status cannot be Ready while any material figure lacks a source location.

Run your result through the same six questions every time.

Mark each one for yourself. Nothing is stored or sent.

Completing the checks does not approve the result.

Copy any prompt from the pack, or download everything.

The pack names the file each prompt uses, and how to verify it.

Prompts 4 and 7 are optional. Try one on your own data this week.

Here is what you can do now: scan, connect, brief, challenge and verify.

This workshop is not about asking AI for a dashboard. It is about asking better operational questions, tracing every answer to evidence, challenging bias and causation, and deciding what to investigate next.

Copilot proposes, and the learner verifies.