Introduction

Hello everyone, my name is Antonio Osuna.

I'm working here at Celonis as a staff software

engineer and lately more closely to the AI and LLMs like I guess most of you today.

And

today I'm bringing you an example of how the industry of software is introducing is integrating

integrating AI capabilities in already or consolidated products like Celonis.

Celonis has a platform.

We already heard about it before by AI.

And yes, Celonis has a platform that provides

customers the way to ingest, process and analyze the data from the databases, right?

Not something saying, hey, my process is working like that.

It's more like we are looking at timestamp, we are looking at data, and we are figuring

out what is going on in the real process.

Use Case: Logistics and Customer Complaints

Setting the Scene: Conflicting KPIs

So today, in order to illustrate you a bit the use cases that I bring in, it's a logistic

company, right?

Imagine something like Amazon.

Everybody knows Amazon, I guess.

and yeah Amazon has a lot of departments and processes inside and one of them is

customer support for example and deliveries process right so what is

going on here in this scenario we have the customer support VP which is saying

to the deliveries VP hey what the fuck man I have a lot of complaints and I

I have a lot of claims from their customers.

What is going on?

So kind of the KPIs are exploding, skyrocketing.

And the deliveries VP says, hey, have you

seen my on -time delivery?

Everything goes well, right?

Orders are delivered.

It should be fine.

But the customer support VP says, hey, I have a lot of claims.

I have a lot of complaints.

lanes so yeah let's figure it out something is not working here so they all agree that we have

to they have to look at that idea look at their process so here is where process mining a step

Turning to Process Mining

in so a data analyst analyst help with Solonis is gonna investigate the case it's gonna get the

insights about the situation right so let's move to the a real scenario in the data so let's see

Exploring the Data

Deliveries and Complaint Records

that here we have the deliveries data and in the bottom we have the customer claims

so yeah we have a kind of um the name of the customer we have the shipping address we have

all the data that we need and also the way that the order was delivered I mean

I bet that all of you have happened that the delivery guy called you and hey you

are not at home where can I left the the order right the the parcel in the step

door in that with a neighborhood something really imagine imaginative so

So here we have in a box, in the neighborhood of the 1B,

in the mailbox, I don't know, any things.

The front door, this is generated by AI

with my loyal friend Gemini.

And in the bottom, we have the complaints, right?

So yeah, one user says the package is completely crashed.

Here, there's a lot of really funny claims like this one, which is the package left at

the kebab shop and now it smells really bad, something like that.

We have a bunch of claims.

We have a bunch of orders that we have to figure out what is going on.

Let's start.

Free Text as an AI Opportunity

Here, we have a lot of free text, which is quite complex to understand, but it's the

the real and the best case for the LMS and AI.

So let's start using Celonis product

to go drill down into details of what is going on

and how can we analyze this in a better way.

From Unstructured Text to Actionable Insights

Clustering Delivery Notes

And starting from, hey, I want to analyze the way

that the delivery guy has left the order, which

which is many free text like I've let this in in the door but something else

means the the same right so we have a bunch of groups of ways of delivering

something kind of to the neighborhood in the door in a mailbox whatever right but

explain in many different words that's the challenge right so here what we're

We're going to apply a clustering algorithm, which

is kind of embedding and generating the embeddings

from all the text, the free text,

and then grouping them in different groups.

So we can get kind of leveled from the information.

And here we can quickly run a test.

And here we have, for example, all these inputs,

inputs, which is different neighbors and different ways of explaining the same, they have left

into the same group.

This is the first step to understand what kind of delivery is producing

more claims.

LLM-Assisted Categorization

If I execute this, then I can go to another tooling that we have in the

the platform, which is, OK, I'm going

to introduce the data, which is all my groups that has been

generated from the previous step.

So here we have concatenated all the free text, group it.

And then I say to the LLM, hey, analyze these groupings

and tell me a category that I can understand,

and I can filter, and I can organize.

So the output is the category, and if I run the test, I will see the output, which is okay for this group.

It's led with a neighbor, then placed in a recycle bin, which is not ideal.

Delivery to balcony, there's many ways that people left the orders when you are not at home.

and yeah here we have the AI generated output into the platform the idea here

is is not only using AI is using it quite well integrated into the product

so you have all the data ingested available and then you will have back

the data generated into the platform to still using it in the rest of the

Operationalizing: Sentiment and Resolution Suggestions

product then we have this first step which is okay let's take actions right

so for the for the inputs that we have from the claims tell me about what is

the sentiment this is the problem that I that it gave to the to the LM hey for

For this customer complaints, tell me about the sentiment.

It's a bit hungry, considerably hungry, and super hungry.

And also the customer resolution.

For example, refund needed, or just apologies

and communication about what is the next step to give them

a resolution.

So if I trigger this with the configuration,

I also asked for an email in case of apologies so you can yeah automatically

generate emails for the customers and here we have the sentiment as output

resolution and also they made email suggested considering the claim also the

delivery situation and yeah the customer name and so on so far and coming back to

Driving Decisions with Dashboards and a Copilot

KPI Visibility and Root Cause Links

the real situation here i have my dashboard bps are happy with every with all the numbers right

so bp really likes kpis and dashboards to understand the situations and here you can

see hey this way of delivering something is producing a lot of claims from the customers

so you can easily match what is going on and what has to be improved in your in your process

and at the end just for imagine that i'm working on it i'm data analyst i know about

salonis product but people may know may don't know about salonis or data so here we have the

Copilot for Explanations and Outreach

last step which is a co -pilot that helps you to understand the situation the real situation right

Here I have, for example, a question which is, give me the emails prepared to be sent

for the most affected customers.

If I send it to the agent, which is linked to the source data, the generated one, and

also all the context of the data, it is giving you the proposed emails for the customer,

customer, the most customer affected, which are really angry about the situation.

And you can copy -paste and send it, and also you can coordinate this and take actions automatically

into the platform, kind of automatically sending the emails.

Scalability and Orchestration

Maybe you can say here, hey, this is a really simple example.

example, we have a very limited data set.

Why It Matters at Scale

But right now, our customers are working with

millions of rows.

So this is why it is really helpful, really impactful for the time consuming

and the impact for the customers, which are actually processing a lot of data.

And on

Automating the Workflow

On the other hand, we have the orchestration, right?

Because this is a process that I've done manually,

sequentially, but you can also put this

in an orchestration engine,

which is automatically triggering all the actions

that we have seen in the demo.

Conclusion and Q&A

And at the end, everybody's happy,

everybody knows what is going on,

and the customers also receive the compensation

or the apologies from the company.

And yeah, that's it.

This is the demo.

Do you have any questions about...