The new Copilot() function can be great for taking disorganized data like feedback from people and grouping it quickly. We’ll explain the best ways to do that and the pitfalls of relying on AI.
One of Microsoft’s examples for Excel 365’s Copilot(() caught our eye and we were pleasantly surprised when we extended their example. Even with more realistic data and better filtering, it worked well. Not perfectly but it genuinely adds new possibilities to Excel
Microsoft’s example is classifying feedback from people about a coffee machine.
Microsoft’s examples are always carefully made to exclude anything that might make the feature look bad.
Their feedback was typically straight-forward and in complete sentences. So we added some more realistic feedback with incomplete and sometimes contradictory remarks.
And we changed the prompt to include specific and more useful classifications; “Classify this feedback as Good, Bad, Very Bad, Don’t Care or Suggestion” instead of letting Copilot decide.
As you can see, Copilot did a pretty good job understanding the feedback and the classifications. A few mistakes:
- The blank answer (row 23) is called “Don’t Care”. Since blanks are common in real world feedback, may be add a “No Response” or “Blank” classification.
- Row 24 “Stop asking questions …” isn’t really “Bad” feedback. Should really be “Don’t Care”.
A better way
Processing external data, especially human responses, is a good use of AI. The key is to give Copilot the classifications you want, instead of just asking it to decide.
As always, carefully check the Copilot answers to make sure the AI has properly understood the groupings and the incoming data.
Here’s the sheet we ended up with:
=COPILOT("Classify this feedback as" & TEXTJOIN(", ",TRUE,D21:F23),
CoffeeFeedback_XTRA[ Coffee Machine Feedback - with extras])
Look at the results and see if there’s more or better groupings. We added “Don’t Care”, “Blank” and the very useful “Suggestion” to better understand the data and separate out actions that could be taken.
The classifications are in separate cells D21:F23, added to the Copilot() context using TextJoin(). The groupings can be more easily changed just by editing the cells.
Contradictory feedback “Very user friendly but confusing to operate”, Copilot took as a “Suggestion”. Oh well.
“The coffee machine is reliable” is called “Very Good” feedback which is a matter of opinion. Some people might think that “Good” was enough since a reliable machine is a basic requirement.
You might first let Copilot decide the groupings by just asking it to “Classify this …” but then adapt those suggested groups into more suitable ones for your needs and the feedback received.
In short: Copilot() in Excel is much better at analysis and interpretation than it is at retrieving hard facts.
Discover the COPILOT Function in Excel
Facts are a slippery thing with the Copilot() function in Excel
