Introduction

Hi everyone once again before we start does anyone of you work in customer service or customer relationship management?

No one okay then this is totally fine I don't as well so so we are on the same page um so yep hi

everyone my name is Maciek I moved to Switzerland nine years ago from Poland I'm both a dog and a cat person, so no discrimination here whatsoever. And I've been working in the data for the past 15 or so years, so for quite some time.

Since the last five years, I've been running Diffuse Data, like a boutique agency that works on all things data.

Context: High-Volume E-commerce Reviews

data and for the last half a year I've been also working as head of data at this plate running the data thing there so does anyone has anyone heard of this

plate something cool so we basically make those very cool metal posters that you can put on the wall just with just just a magnet they are made like in quite nice quality, so for all the textures, etc.

And we sell, if I'm not wrong, about two millions of them every year. So quite a volume.

Why Reviews Spike Seasonally

And basically, because of that, we sell a lot during Black Friday and all the pre -Christmas shopping spree.

We also get a lot of customer reviews like on on some platforms like like Trustpilot but also with Zendesk and all the and all the basically stuff that are there

The Manual Analysis Bottleneck

so within just just December if I'm not wrong we've got over 2 ,000 different customer reviews so so quite a number and basically we needed to analyze them

1And it would take our analysts over three or four working days to go through them and just categorize them, just deciding whether this review is regarding the shipping, this review is regarding product quality, mounting quality, or anything like that.

And what we've done during the last couple of months was basically to automate this process end -to -end. And so that we, because, you know, data people are usually quite lazy, we don't want to do it, so we automated it entirely, right?

But then we built the entire data pipeline for that, all the tech, mumbo -jumbo. But as I was talking with Alberto and Tere already yesterday, I kind of was made aware that we don't really need all these things just to get our output.

Goal: Automate Review Categorization End-to-End

So, what I wanted to show you today is how you can very easily build a use case like that with all the 3D available tools by yourself. And basically, we'll use three different elements, I mean, two different elements, but three different tools.

Tooling Overview (Gemini, Claude Desktop, Claude Code)

The first one will be Gemini in Google Workspace. The second one will be cloud in cloud desktop, and the last one will be with cloud code. So let me quickly show you what I meant here.

So it's quite small, I will make it quite a bit bigger, sorry about that. Let's say that we've got hundreds of different customer reviews with only the date and the customer email.

And some of them are saying that, yeah, the metal print feels premium. Some of them love the League of Legends designs that we've got. some of them are complaining about long shipping etc etc.

Those reviews by the way were also generated by an LLM so they are not our actual customer reviews. But then it's not really that helpful to analyze that or to see how our customer perception changes from from month to month.

Step 1: Enrich Reviews in Google Sheets with Gemini

So in order to To do that, we can basically do it directly in Google Sheets without having to leave it. So we're in Google Sheets.

Generating a Synthetic NPS Signal

I have just added two columns, basically. One is about the Net Promoter Score, and the other one is about what kind of category should this review be assigned to. And as we don't get NPS data here, I want to be also able to kind of generate it synthetically.

So within Google Sheets, you can now use AI directly in functions. So in case of the NPS score column, I just said, like, analyze the following customer review and assign the most likely NPS score that the customer would add to it. So, we are not like 100 % sure whether that makes sense or not, but it is some kind of indication for us.

And if we take a look at it, some of them are quite positive, some of them are quite negative, it kind of seems to work. And when I was looking at it, it turned out to be quite fine.

Classifying Reviews into Clear Categories

the other prompt that I gave it here is basically regarding the review category so what the review is about and in that case my prompt was analyze the following customer review blah blah and assign it to one of the following categories product quality mounting system artwork and design shaping in customer service or price to value ratio and as you can see here the categories are also are also assigned quite nicely so let's take a look at

that one this was assigned as artwork and design and the review was ordered this as a gift for my husband who's a huge space nerd he was absolutely thrilled great gift option. So NPS of 10, customer very satisfied, and it does talk about the space -related designs that we have.

So basically, what you can do now, after the sandwichment, and again, it's in Google Sheets, right? So if you know how to use spreadsheets, then you know how to use AI, which is, I think, quite a cool use case for that.

Let me just move a bit here.

So what I've done next is that, you know, I'm just a lazy data person, I don't want to spend time on analyzing this, right? And to see, you know, what's the operating PS, how does it differ between categories, etc., etc.

Step 2: One-Off Analysis in Claude Desktop

So using cloud desktop. I just exported the spreadsheet to a CSV file, told cloud to analyze the data and show timeline chart of average NPS per day.

It did some basic analysis saying that the overall average is 7 .7, it highlighted some best and worst taste etc and it also created a chart of this with this NPS score built into that so quite a cool case for like one -time analysis when you need to

dig deeper into that you could of course also ask it to do further categorization of the results or in my case like what are the users concerned the most with And it turns out that the customer service is the biggest problem, then the shipping times, then the price -to -value ratio, et cetera, et cetera.

Turning Insights into Customer Outreach Emails

And then I was like, okay, so as a customer service person, I also want to be able to automatically improve the experience of the people who are affected the most. and then basically what I've done is to say create a list of customer service detractors along with an email to apologize for their experience listing

improvements to the customer service process that we will be implementing and offer them a 20 % discount code for their next order, right? nothing nothing more than that and then basically what's done is that it went

through the list through the reviews with the lowest customer service score and basically wrote like quite a nice emails hey we let you down Brian and we want to make it right and then like the whole and like a cool strategy on how to

fix that here's your discount code and if we were to automate the send outs like it could be automated end to end right so so that's quite nice thing but then this is like using cloud desktop in that or any kind of those those like chatbots it is like a one -time thing right you need to open it that you need to talk to the chat, but you need to ask it to do something for you.

Step 3: A Repeatable Dashboard with Claude Code + Streamlit

But then I was like, okay, so let's say that now I want to do this analysis once a week or once a month, right? And I want to have like a dashboard that will just grab the data from any kind of database that it lives in, or a Google Sheet, and then automatically visualize it so that I can analyze what's happening the most, I mean, what's happening with the current setup.

up so now moving on to cloud code basically this is like a cloud for coding as the name suggests and you can ask it to do anything

for you so I asked it to build and run a streamlit app streamlit is basically like a like tableau or power bi but but built with python allowing for full analysis of the data in my csv file so i didn't tell it what what kind of metrics i want to take a look at i didn't tell it how do i want to break them down or what what i want to analyze i only asked

basically for the best practices, right? And this was my only prompt.

The other stuff that I've done here is that I asked it particularly to use a very specific Python method to do that, and then as it builds the dashboard, I asked it to replace the scale that the score was actually presented at with the average NTS.

So that's my prompt here. And all the other stuff was done automatically based on my original prompt, like build and run a stream it up, allowing for full analysis.

And, you know, my career was built on defining APIs and building reports and making people to use them right, etc. And make sure that they reflect their needs, the organizational structures and everything. 1And this process now took me like two minutes to go through this, right?

What the Auto-Generated Dashboard Includes

and what it built basically looks like that so it's like a full -fledged dashboard with all the average average net promoter score and the whole distribution of promoters passive detractors etc broken down by category

broken down by the score itself showing the trend over time both by day and by week, showing what the score distribution by different categories, showing what kind of keywords were occurring the most often in the reviews, especially across the happiest customers and least happy customers, and then

showing, let's say, what are the actual keywords used, and lastly allowing the end user to actually be able to just preview the reviews if they don't have access to the original spreadsheet.

And that's the work that would have taken me or any analyst in my team at least a couple of days. And in that case, having it all built in two minutes, I was

quite astonished with that and I think it's a really powerful tool and yeah if you if you get an opportunity to do something like that I can highly

Conclusion and Next Steps

recommend it to all of you so this pretty much is for the demo if you have

any questions please let me know and also if you would like to connect on LinkedIn you can scan the QR code