Common business use cases for generative AI

Summary

In a panel discussion, industry experts from Google, Vodafone, and Blue Core shared insights on the application of generative AI in business. They highlighted the importance of prioritizing customer value and technical feasibility in AI projects. Notable use cases included Alpha Fold's impact on drug discovery, product cataloging, and customer service operations. The panel emphasized the transformative potential of AI, the need for continuous learning, and the value of hands-on experience in understanding its practical applications and limitations.

Takeaways

• 🌟 Prioritization of AI use cases focuses on customer value, friction points, technical feasibility, and aligning with internal research innovations.
• 🔬 AlphaFold on Google Cloud is a prime example of AI research translated into a practical solution, accelerating drug discovery by understanding protein structures.
• 🛍️ Product cataloging is a significant use case where AI helps in categorizing products for search and creating website copy, improving customer experience and retail performance.
• 💬 Customer service operations benefit from AI through conversational agents and internal support for SRE teams, enhancing post-mortem search and summarization.
• 📊 Generative AI is utilized to improve features on platforms, increase internal efficiencies, and better serve customers by standardizing product data and mapping it to taxonomies like Google's.
• 🌐 Telecommunication companies leverage AI for network deployment, predictive maintenance, and customer call analysis, translating and summarizing calls to understand customer issues and improve service.
• 🔄 Technical solutions in AI involve embedding models, vector databases, and vector search engines to process and retrieve unstructured data efficiently.
• 🚀 The application of Transformer models extends beyond text and images to non-traditional use cases, such as analyzing and predicting user events in software applications.
• 🌐 Global deployment of AI solutions requires consideration of scalability, replicability, data security, and compliance with local regulations like GDPR.
• 📈 Business value from AI is measured in terms of increased employee productivity, improved user experiences, and the generation of new insights or ideas that were previously impossible.
• 💡 The importance of continuous learning, hands-on experience, and staying up-to-date with the latest AI research and technologies is emphasized for maximizing the potential of generative AI.

Q & A

What is the main focus of the session on common business use cases for general AI?

The session focuses on discussing and understanding the top business use cases for general AI, how to prioritize them, and how to get started with their implementation in various industries.

Who are the panelists in the session and what are their roles?

The panelists include Nema Dakiniko, a product manager at Google for AI portfolio, Ignacio Garcia, the global director of data analytics and AI and CIO of Vodafone Italy, Arvind Christian, who runs engineering, data science, and solution architecture teams at Blue Core, and Donna, who leads the Technical Solutions management team for generative AI at Google Cloud.

How does Google prioritize its AI solutions for customers?

Google prioritizes AI solutions by focusing on what will add value to the customers, identifying friction points they face, and considering the technical implementation of the solutions. They also look internally to their research teams to bring innovations to customers.

Can you explain the significance of AlphaFold on Google Cloud?

AlphaFold is a significant research project by DeepMind that focuses on solving the protein folding problem. It's important because understanding a protein's structure can lead to the development of drugs that modulate its function. Google Cloud has operationalized this research into a solution that adds value to healthcare organizations by making it reproducible, scalable, and cost-effective.

What are some of the technical patterns observed in different AI use cases?

Technical patterns include using embedding models to process unstructured data and index it with a vector database, using coding models like Codex to generate SQL or Cypher for database access, and applying Transformer models to non-image, non-text use cases.

How does Vodafone Italy utilize AI in its operations?

Vodafone Italy uses AI for strong model analysis to understand customer behavior and offer models, network deployment for capex efficiency, predictive maintenance, and summarization of customer call data to understand and improve customer service and reduce detractors.

What are the key technical considerations when deploying AI solutions globally?

Key considerations include scalability and replicability, ensuring data security, adhering to local regulations like GDPR, avoiding bias in models, and creating policies to maintain these standards.

Why did Blue Core choose Google Cloud for its AI projects?

Blue Core chose Google Cloud due to its innovation roadmap, co-innovation approach, data governance and security features, and the performance they observed with Google's AI technologies compared to other models.

What are the business benefits of using generative AI?

Business benefits include increased employee productivity, more intuitive and better user experiences, new insights or ideas that were impossible before, and cost-effectiveness in problem-solving and scaling solutions.

What advice does the panel have for businesses looking to implement generative AI?

The advice includes experimenting and iterating with AI, investing in foundational architecture, creating a safe environment for experimentation, not restricting innovation, and educating the entire company about the potential of AI.

How does the panel suggest businesses should approach the rapid evolution of AI technologies?

Businesses should stay humble, focus on short-term use cases, invest in tooling and platforms to support AI, and ensure that their AI initiatives are not centralized to avoid stifling innovation.

Outlines

00:00

🤝 Introductions and Panel Discussion on AI Use Cases

The panel discussion begins with introductions of the speakers: Nema Dakiniko, a product manager at Google AI, Ignacio Garcia, the global director of data analytics and AI at Vodafone Italy, Arvind Christian, head of engineering at Blue Core, and Donna, who leads the Technical Solutions management team for generative AI at Google Cloud. The panel aims to discuss common business use cases for AI, with a focus on how to prioritize and implement AI solutions effectively. The conversation starts with Donna explaining the prioritization process based on customer value and technical feasibility, and then moves on to specific examples like Alpha fold on Google Cloud and its impact on drug discovery.

05:00

📈 Product Cataloging and Customer Service Operations

The discussion continues with the application of generative AI in product cataloging and customer service operations. Arvind and his team have worked on categorizing new product labels for search, improving website content, and supporting SRE teams with post-mortem search and summarization. Ignacio shares Vodafone Italy's experience with AI in customer service, emphasizing the importance of understanding customer needs and improving interactions through AI. The conversation highlights the benefits of using AI to standardize product data and the technical considerations for deploying AI solutions globally.

10:01

🛠️ Technical Design and Implementation of AI Solutions

Kevin and Donna discuss the technical aspects of designing and implementing AI solutions. They cover patterns like using embedding models and vector databases for unstructured data, code generation models for database access, and applying Transformer models to non-text use cases. The conversation also touches on the optimization of the inference pipeline for Alpha fold, the importance of reproducibility and experiment tracking, and the use of Vertex AI for automating processes and ensuring security.

15:02

🌐 Global Deployment and Business Value

The panelists delve into the challenges and strategies of deploying AI solutions worldwide. They discuss the need for scalability, replicability, and compliance with local regulations like GDPR. The conversation highlights the importance of data logistics, policies, and the use of Vertex AI for secure and efficient model deployment. The panelists also share their experiences with the business value of AI, including improved accuracy, cost-effectiveness, and the ability to understand customer needs better.

20:05

🚀 Insights and Learnings from the Journey with Generative AI

The panel concludes with insights and advice on working with generative AI. Kevin shares his learnings about the importance of continuous learning, accessibility of AI tools, and the value of research papers. He emphasizes the need for a safe environment for experimentation and the transformative potential of AI. Donna advises on experimenting, iterating, and investing in foundational technologies to combine AI models with proprietary data. The panelists stress the importance of staying humble and focused on short-term use cases while preparing for the rapid evolution of AI technology.

Keywords

💡AI

Artificial Intelligence (AI) refers to the simulation of human intelligence in machines that are programmed to think like humans and mimic their actions. In the context of the video, AI is the central theme, with discussions on its business applications, technical implementations, and the impact on various industries.

💡Product Manager

A Product Manager is a professional who is responsible for guiding the development of a product, from its conception to its launch and ongoing improvement. They often work closely with cross-functional teams and are key decision-makers in product development. In the video, Nema Dakiniko is introduced as a Product Manager at Google, highlighting the role of product management in AI technology development.

💡Data Analytics

Data Analytics is the process of examining data sets to draw conclusions about the information they contain. It involves the use of statistical tools, data mining, and predictive modeling to analyze data and make informed decisions. In the video, Ignacio Garcia discusses his role as the global director of data analytics and AI, emphasizing the importance of data analytics in driving AI solutions.

💡Generative AI

Generative AI refers to AI systems that are designed to create new content, such as text, images, or music, based on patterns learned from existing data. It is a subset of AI that focuses on the creation of new outputs rather than just analyzing or recognizing patterns. The video discusses the use of generative AI in various business contexts, such as customer service and product cataloging.

💡ML Infrastructure

Machine Learning (ML) Infrastructure refers to the underlying technology and systems that support the development, deployment, and maintenance of machine learning models. This includes hardware, software, and frameworks that enable data scientists and engineers to build and scale ML applications. In the video, the panelists discuss the importance of ML infrastructure in operationalizing AI research and bringing it to customers.

💡Customer Service Operations

Customer Service Operations involve the processes and strategies that organizations use to manage and improve their interactions with customers. This includes handling inquiries, resolving issues, and providing support to ensure customer satisfaction. In the context of the video, AI is being used to enhance customer service operations by automating responses and providing insights into customer needs.

💡Technical Solutions

Technical Solutions refer to the application of technology to solve specific problems or meet particular needs. In the business context, this often involves the development and implementation of systems, tools, or processes that address challenges faced by an organization. The video discusses how technical solutions, particularly those involving AI, are identified and built to add value for customers.

💡Data Structures

Data Structures are specialized formats for organizing, storing, and managing data in a computer's memory. They are crucial for efficient data access and modification and can significantly impact the performance of a computer program. In the context of the video, data structures are important for combining with AI models to create effective solutions, such as mapping product catalog data to standardized taxonomies.

💡Replicability

Replicability refers to the ability to produce consistent results or outcomes when the same process is repeated under the same conditions. In scientific research and technology, replicability is essential for verifying the effectiveness of a method or solution. In the video, the panelists emphasize the importance of creating AI solutions that are not only effective but also replicable across different regions and languages.

💡Google Cloud

Google Cloud is a suite of cloud computing services offered by Google, which includes Google Compute Engine, Google Cloud Storage, and Google App Engine. It provides businesses with a range of tools and services to build, deploy, and scale applications, store data, and process large amounts of information. In the video, Google Cloud is the platform where AI research like Alpha fold is operationalized and made available to customers.

Highlights

Session on common business use cases for general AI

Nema dakiniko, a product manager at Google, introduces the session

Ignacio Garcia, global director of data analytics and AI, discusses Vodafone's AI initiatives

Arvind Christian shares Blue Core's work in engineering, data science, and solution architecture

Donna leads the Technical Solutions management team for generative AI at Google Cloud

Prioritization of AI use cases based on customer value and technical feasibility

Alpha fold on Google Cloud as a notable AI research breakthrough

Product cataloging as a key business use case for AI

Customer service operations improved with AI conversational agents and internal support

Vodafone Italy's use of AI for call center call summarization

Technical design patterns for different AI use cases

Importance of data structures and taxonomies in AI solutions

Global deployment of AI solutions with considerations for scalability and replicability

Use of generative AI in retail for customer movement and shopping behavior analysis

Google Cloud's role in providing AI tools and infrastructure

Business value of AI use cases and measurable outcomes

Advice for businesses on adopting AI: Experiment, iterate, and invest in foundational technology

The transformative potential of generative AI across industries

Transcripts

foreign

[Music]

let's get started hey everyone uh

welcome to this session on common

business use cases for general of AI I

am Nema dakiniko I'm a product manager

at Google for our gender of AI portfolio

but I will turn it over to our esteemed

guests here who can introduce themselves

because we do have a very packed panel

and we're going to start with you hi

everybody I'm Ignacio Garcia I'm the

global director of data analytics and AI

for both of them and I'm also the CIO of

Vodafone Italy

everyone arvind Christian here I had

engineering at Blue Core so I run

engineering data science and our

solution architecture teams

hi I'm Donna I lead the Technical

Solutions management team for generative

AI at Google cloud and together with the

solution architecture teams with Kevin's

team we identify design and build AI

Solutions

foreign

Solutions and we also do ml

infrastructure as well

very cool uh okay so let's start first

by talking about top business use cases

and Don I'm going to start with you on

this one but I really want to understand

many people here are like look Jenai

fantastic this is great I'm sold but

what are those top use cases well how do

you prioritize them how to get started

sure yeah so um I can start with

prioritization and then go into a few

use cases so we really in our

prioritization we focus on our customers

what will really add value to them where

are they seeing friction points and then

the Techno technical implementation of

that we also look internally to some of

our research teams and think about how

can we take those Innovations and then

bring those to our customers so one

example that we worked on around one and

a half years ago

was Alpha fold on Google Cloud so

deepmind had this amazing research and

maybe to give a little bit of context on

the protein folding problem scientists

have long been interested in solving the

protein folding problem because once a

protein's structure within a cell is

understood then scientists are able to

develop drugs that can modulate its

function

but for our Healthcare organizations in

order to be able to actually leverage

that they need additional requirements

so for example reproducibility

scalability it has to be cost effective

um and so we took their their amazing

research and we operationalized it on

Google Cloud through a solution

some other areas more recent where we're

seeing traction or for example product

cataloging so being able to categorize

when there's a new product label it for

search and then also create the website

copy and arvind and his team have done

some amazing work in this space and

customer service operations and it's not

just a conversational agent which you

may have experienced on a website that's

answering questions but also internally

for example supporting SRE teams with

post-mortem search and summarization

which we've done some work on but also

supporting support agents with

summarization and next steps and Ignacio

and his team have done some some great

work here

what are some of the use cases and

prioritizations that your organization

does yeah

um before I jump into a use case maybe

just a little bit step back and what we

blew core do so we can connect the dots

so we are an identification and a

customer Movement platform so we work

with large Enterprise retailers to

identify and then convert Shoppers to

uh repeat customers

so we've used traditional Ai and created

over 20 retail models using first party

data so Shopper information

behavioral data and then product

information

so these models are baked into our

platform so marketer can use these

models to create campaign campaigns and

audiences

so the content that needs to be

generated

the channels in which to deliver these

uh the the content and finally the

timing when to deliver are all

personalized on a per Shopper basis

with the Advent of gen AI we looked at a

couple of areas you know one is how can

we improve uh our features on our

platform we looked at internal

efficiencies as well and option

opportunities to better serve our

customers

so um the the problem that she was

referring to uh is core to our value

proposition so taking unstructured

uh product catalog data and mapping it

to Google's product taxonomy

so for example

um a retailer could call this a t-shirt

another retailer could call it a

categorize it as a true cut T-shirt and

so on and so forth but if you

standardize it in the Google's taxonomy

this is probably labeled as a t-shirt

which is under categorized under a shirt

which is apparel and so on so there are

a number of uh advantages to

standardizing product catalog one is we

will improve our models and our wrecks

our customers will be able to now

analyze performance within their product

catalog and finally we will be able to

deliver Trends within verticals across

the retail space

so using gen AI

we were not very successful using

traditional AI so with Gen AI we were

able to solve this specific problem very

cool absolutely I love the phone

hopefully people are a little bit more

familiar with it in terms of what you do

but love to understand your use cases

and prioritizations oh thank you I think

that I still will take a couple of

minutes just to explain the the

complexity that we have and that context

maybe help to understand how are we

using what we're using so we are a

telecommunication company we do mobiles

we do

televisions we do fixed lines and we do

iot so the whole package across the

world we have more than 300 million

customers we have billions of iot

devices so I'm just talking about the

scale and then it's in Europe and in

Asia and in other areas so languages are

completely diverse and this is another

point that is very important on the on

how we're using and the type of problems

that we need to resolve probably a bit

of background as well is we have been

very focused on partnership with Google

on cloud and data so Google is our

partner on the data domain and we have

been very focused on making sure that

our data is in the right place is safe

for our customers we have the

anonymization we have all the

regulations that we have in Europe

around privacy and we're super focused

on that and we have been using Ai and we

have been using Google tooling and for

many of our normal operations so if I go

it will be strong models analysis on

trying to understand why the customers

are going up next best offered models

that is on the customer side but then if

you go to the network we we do analysis

in where to deploy the network so capex

efficiency which is super important or

and we call it a predictive maintenance

so trying to understand what is going to

break and and be sure that we can

replace the components and and that was

successful but then with gnai we have

been now experimenting and and it's a

completely different dimension for

example the use case that you you are

saying with it implemented in Italy so

we are getting all the calls that our

customers are making to the call centers

We're translating them into text and

then we're getting a summarization of

the problems what was the original

intention to reduce the risk that the

customers are calling us and to drive

automatically the Deep detractors so the

amount us is the attractors this is only

possible now that we have large language

models available and we were able to do

it very fast because we have been very

consistent on creating the data

structures to combine combine the the

models with our data in a good way but

this use case in particular so we're

taking these 50 000 calls translating

them into text summarizing and getting

the the reason of the problems and it's

a complete Game Changer because then we

understand what the customers are really

saying we don't need to do surveys and

get high level data we're really getting

to the actual details on why are they

calling us and then we can intervene on

that and then that data that was a

regional reason and then that data has

become key to do other things to

understand behaviors and understand

potential upselling and other areas and

the other important thing on that is we

need to replicate that across all the

countries so replicability and scale is

fundamental it's not only doing and what

is a use case is can we do it fast and

with it secure and with it across a wall

in a in a way that we can repeat and we

will talk later about vertex Ai and

different components on Google have

allow us or is allowing us to us

absolutely it's interesting I think both

we actually spoke about data and like

taxonomies and data structures so this

goes into our second question around how

do you actually technically design these

Solutions and Kevin I'll start with you

um okay so Donna's team and my team have

worked on a number of different uh

business use cases applying to an AI and

uh what's interesting is to see a few

technical patterns kind of surface up or

just kind of permeates across the

different use cases right so the first

one is how do I get data that's outside

of the llm right into my application

right so that use case is very pertinent

to you know customer support right for

example so uh one of the very uh one of

the very common technical patterns is to

use

um like a embedding model to process

your unstructured data and then index it

with a vector database you know so

matching engine are now called a vector

search right yeah is a very popular

option and now we also have lodb with PG

Vector as well right so there are a lot

of options for that yeah and after you

index that then you can very quickly

retrieve your unstructured data images

and text and so forth the other type of

data received we're now seeing right is

to use a coding model like code to

generate the SQL right to action to

access your relational database or to

generate Cipher to access neo4j right so

that's that's kind of another kind of up

and coming type you know pattern that

we're seeing and another area that we're

seeing is around application of these uh

language or you know these Transformer

models to non-image non-text use cases

right so one of the partners that we're

working with full story we're helping

them build a sequence model to analyze

and predict user events right and

finally with the Donna mentioned Alpha

fault right well allothold actually is a

Transformer model it's very interesting

yes so instead of generating language or

image it generates protein structures so

we took deepminds research

we broke the inference pipeline which is

actually a multi-step pipeline right so

we applied uh different types of compute

to the different processes the earlier

parts has a lot of data retrieval so we

use high uh IOP CPU nodes for the later

compute stages which is extremely

compute intensive we use Nvidia a100

gpus right so that's how we can optimize

right going from research to production

to production we also need to make sure

that we're we we have reproducibility

right and experiment tracking and for

that we use vertex metadata right and to

kind of stitch it all together we use

vertex pipelines to automate the process

so adaptation is another very

interesting area that we're seeing a

more customer uh demands

very good you can also I'll go to you

first what are those technical

considerations that you have to have

when you're trying to deploy this around

the world essentially uh the first thing

and again is escape scalability and and

replicability of what we are doing

because we have to secure the data we

have a lot of local regulations uh about

gdprs in different countries have their

own flavors and we have to to create or

we have created policies around and

making sure that there are no bias in

the model oh making sure that we can

detect those buyers is very regulated

the world in Europe probably very

different to America but there we we

have to complain with a lot of things

and you can do these things what takes a

lot of times and by the time that you

deploy then your data signs are going to

kill themselves because they have in

spending 99 of their time in activities

that are not related to the model so

architecturally what we have done is

take out the problem of data transport

so making sure that their data arrived

to the right place is something that we

do and we do in with your and

Engineering that we have created for all

the data Logistics that allow us to

monitor to make sure that the data is

encrypted the data is anonymized and if

we change anything on the policies that

applies for all the the data pipelines

that we have across so that that has

been designed and then we have the

process inside which is a bigquery

standard and then in the top and here is

where vertex and and your product is

fundamental we we have in a very early

adopter so vertex AI we created

something that we call the AI booster

based on vertex AI these are our

adaptation but that is where we receive

the models and where we then exchange

and make sure that we can do what you

are saying which is running models with

your own data and data that is subside

so architecturally we have data

Logistics we have all the where we have

policies security encryption and

monitoring all the good stuff then we

have the engine that runs all the

queries which is our real Nerf systems

if you want to call it like this and

then in the top we have created with

vertex AI the interface to run all the

models and to co-create what is that

allowing us that now our data science

are now working on encryption data

engineering uh all the all the

bureaucracy and equally we can share

models across the market so the data

Engineering in Italy that have created

this model are now just passing the

information and the the guys in Germany

that are going to run it I can do it in

weeks rather than in months which was

um the previous setup so it's very very

important for us to take it

and spend that time on making sure that

we have the foundation right then in

parallel we allow a lot of

experimentation because very important

that the people can experiment and see

the power in a safe environment where

they can play and they see and they see

that the model is right but the

deployment is very automatic to

automatized and and it's very secure

that's excellent so it's a lot of

engineering behind to to allow us to run

properly Auburn can you double click

into the technical aspects of yours sure

um

so the team came up with a really

ingenious uh two-step process so on the

one hand we have thousands of product

catalog in our database and then the

Google product taxonomy has around

roughly around 5500 classes in

subclasses

so what what the team did was first use

uh gecko to create the embeddings and to

sort of reduce and narrow down the

options and then pass that along to

create a prompt and pass that through

text bison to create the final results

very cool very cool um okay so this is a

little bit of a self-serving question

but why did you choose Google cloud and

it can't be Donna and Kevin so so that's

not on the table but um for us was a

proper processor like five six years ago

and we were defining our Cloud strategy

first and and we did a very sort of

analysis and the three reasons I would

say is your header teaching data so you

you have that the The Innovation and

roadmap that you were proposing and the

approach and the ways of working and it

was very refreshing to see that it was a

relation on co-innovation and trying to

tackle problems together rather than

this is a price list and just consume

these products and services and we have

done that and in we have been very good

partners so far I always always have to

make sure so far

um so blue core is natively built on gcp

and so the team is very very comfortable

using the tools and technology and

Google has done a fairly good job

building the Gen AI Technologies

alongside the existing Technologies the

other things which you mentioned around

data governance security are important

to us and a lot of those are also baked

into the jnai technology

so I think that's the first the second

one is we actually started this project

on

GPT 3.5 and we got really good results

and then we elevated and moved on to

gpd4 and we got really good results as

well as well and then when Google

released Palm 2 we decided to try it out

and so far the results have definitely

been better than what we've seen with

open Eis models absolutely Okay so we've

talked about use cases we've talked

about like technical you know Solutions

but at the end of the day it's all about

that business value right like how do

you actually see those business results

Donna like I want to talk to you about

this like start with you

how do we see those business results and

can you give me some examples because

without them it's kind of pointless

right like you want to see those results

so I'm good

yeah um so let me start with Alpha fold

um so in the case of alpha folds the the

customers were able to conduct

experiments much faster get much quicker

insights and also minimize the high

ratio of failures from more traditional

methods and I mean the impact was really

incredible to see and it's also one of

the reasons why I'm in this field

because they were able to accelerate the

drug Discovery process both biotech and

pharmaceutical companies alike

um more generally I would say within the

generative AI space we see customers

getting benefits in terms of employee

productivity allowing them to focus on

more value-added tasks we see more

intuitive experiences being built with

generative Ai and also better user

experiences and then also new insights

or new ideas that were impossible before

but I think arvin's and Ignacio will

have great input here Armin we had two

overarching goals for this specific

project so the first one was obviously

the accuracy of the data and the results

and then

the cost of

building and maintaining this feature

and what we've found

was using llm and gen AI we achieved

both goals so sort of to step back we

attempted to solve this problem last

year using traditional AI techniques and

like I mentioned you know you're talking

of thousands of product catalogs we have

around 5000 Plus classes in the Google

taxonomy and trying to build a model

using that is extremely difficult and

now to scale that across multiple

verticals within the retail space makes

it even more challenging so for us

it was extremely expensive both from a

human time perspective

computational resources and you know the

amount of data we needed to train and

build those models and a lot of those

were solved with Jin AI absolutely

all right I think that I will I will

talk about three dimensions of benefits

first I will go to the use case and the

reason why I'm here which is that

summarization and the the understanding

on on what the customers are calling us

and then I will go to the three areas

where we are working on generative AI

because that's one use case but really I

mean we are going ahead in in many areas

and finally from technology point of

view the benefits about velocity and

cost of building because that is what

I'm responsible for in order to learn

more on the on the recent side so the

these particular use case which is the

summarization on all the calls and why

people is calling to our call centers is

very simple we were not able to do it

before so it's a paradigm change in the

past that was impossible to do

simple as that

um and now we can do it and that

information has changed completely our

understanding and our relation and

proximity with our customers because we

were doing surveys and getting scores

and getting some comments and then it

was a team trying to understand what

does that means and then we were cross

uh Crossing that data with technical

problems that we have and then making

interpretations and you you should see

the debates that we had in our customer

boards that we create to try to

understand and be better and now we have

three really really granular information

that is precise is what they are saying

it's not what we think that they're

saying so that is a paradigm change I

cannot put money on that but we are

trying to reduce 30 percent our lead

detractors and we are in in that track

so he's doing very well then if I go to

the second one which are the different

areas I would say we are working on all

the checkbox and interaction with the

customers it's early days we see an

incredible opportunity in there so we

have checkbox but we have to invest a

lot of money and time and training the

different languages and hence the

explanation about the fact that we we

operate in a multi-language is is

expensive it's not precise the

experience for the customer is not

correct and our early tests are showing

that this can be again a paradigm change

in comparison with what we were doing

before so one big area by the different

domains of customers operations all the

different call centers that we have is

is a chat box the second one is what we

call copilot so it's making sure that

the people can really do their job

better and that imply a lot of areas so

we're talking about

if you are in my team in the in the

technical team coding yeah and we see

productivity that's 10 to one in

comparison with no use in it I still

experimental I'm no I'm not talking

about

details but

legacies so we grow through mergers and

Acquisitions so we have a lot of

legacies with documentation that is not

existence and knowledge that is

disappearing of there

we are testing how to get that those

system documented and that change again

our ability to to run operations in a

different way or to maintain systems

that were not possible to maintain so we

are experimenting in what we call

copilot and you can add every different

business area and we are doing

experiments on that and then it's

Knowledge Management we're a massive

company we have a lot of information

managing knowledge is is a big problem

for us and doing it well can change

completely how our customers are

receiving our services and that is a

certain area so the the first benefit

was in the in the proper bit the second

is in in these three areas and finally

from me something like vertex Ai and all

the engineering is the velocity so the

fact that we were able to do in in weeks

the real application that is working and

for taking all the calls in a full

country is just because the tools are

there and the cost and the velocity of

the plane is is helping us to to really

get the value yeah

I really like this framing of again it's

it's what are you what are you changing

how much does it cost how effective is

it and then can you measure that right

because otherwise it's just a boondoggle

you're spending money and who knows

what's happening but then you can prove

it you can double down on it and build

on top of it absolutely Uh Kevin uh can

you talk a little bit about what you

learned along the way of this magical

genie Journey

um in short I learned that we have to

continue learning I mean there's just so

much to learn I mean think about where

we were six months ago right look at

where we are today you can imagine

what's like six months from now right so

it's a new normal The New Normal is

you're gonna really have to keep the

learning up

um but then the good news is

accessibility right imagine think about

a year ago how could you get your hands

on to llms right you have to spin up

your own VMS spin up there you know just

set it all up and download some

framework install it yourself before you

can even prompt it all right and you

probably have to troubleshoot a bunch of

libraries and stuff like that along the

way I mean we've gone through that so

yeah

um today it's available everywhere all

the cloud vendors have it we have cohere

we have anthropic it's there's really no

excuse to not have that hands-on

experience right so go for it you know

and there are all these you know if you

sign for free and try it out right so

definitely try it out and we have

Frameworks like Lang chain to help you

build up these applications as well

right so go Hands-On I think the last

thing also is we used to say read the

manual read the manual right I think you

know in this world we may still want to

want to start saying read the papers

read the papers papers as in research

papers right so you'll get a lot of

insight of what's going on right so this

a lot of this started with a Transformer

paper back in 2017. it's still a really

good paper to read right so there are

good papers

um if you don't have time read the

abstract all right but I think that's

where you could really get up to speed

on what's coming down the pipe I would

used to go to the Google AI blog and

just read all the papers they were

publishing because I was like listen if

Google's publishing it it's probably

state of the art so yeah that's that's

how I used to cheat sheet my way through

yeah exactly the papers or you could run

it through an llm to summarize get a

summarization right yeah

um

I think for a subset of problems like

the one that I mentioned I think llms

provide a phenomenal Foundation because

there is so much information encoded in

these models because there's craned on

so much relevant data it makes the lift

for us a lot easier right like I

mentioned we tried this with traditional

Ai and it was really not sustainable for

us to scale this and what uh gen AI is

now able to provide or the tools for us

to be to not only solve but able to

scale problems such as the one I

mentioned

for me three things one is is really

transformational everybody can see it

don't try to centralize it in Ito and in

any area let the people experiment so

make sure that you create a safe

environment for experimentation and you

can measure that to then decide where

you're going to invest but let the

people play because the people need to

play don't don't restrict in the other

hand investing in Foundation because it

is the only reason or the only way to

get real value is combining the LL

models with your own data so investing

foundation invests on getting the right

architecture because then you can go

super fast when the experiment shows

that these are use case is worth 2x

scale so that that is my my learning

thing and sorry one more stick with your

CEO make sure that they they understand

the potential the value and they are no

um bombarded by vendors telling them

that is a

bullet point a silver bullet to resolve

every single problem so invest in

education in based on bringing the the

whole company and do not stop do not

centralize it because then you kill it I

I've heard so many customers say the

board of directors the CEO have told me

we need to do this ASAP and so yeah it's

in every single boardroom every single

you know CEO office

um okay one last question and then we'll

open it up to q a so you know get ready

there are mics here in case you want to

ask questions but last question advice

any other pieces of nuggets of wisdom

that you have for the audience here in

terms of business use cases

um sure yeah so I would say and I think

we've all touched on this already but is

experiment and iterate so generative AI

truly has the potential to transform

each industry we're living in a very

exciting time but nothing really

substitutes hands-on experience right to

understand what the value is

specifically for your business and also

to understand the limitations so I would

look within an organization and

customers that we've seen be successful

do this is really look at who's excited

to Pioneer this which domain experts and

which machine learning experts get them

together to identify what does success

look like to work on the use case and we

did this as well so we actually worked

with an SRE team to on the postmortem

search and summarization use case where

we worked with experts that really

helped Define what success looks like if

our outputs are good we started with a

very small set of experts we gradually

expanded to a trusted tester group we

started with a small data set of 100

postmortems we expanded to a thousand

postmortems and now we're rolling this

out more broadly and so um that that

would be my advice is just get started

um

so I I consider uh gen AI or llm as

being yet another tool in your toolbox

right so as you go through building

features I'm sure you have some success

metrics and if Jin AI is going to help

you reach those metrics and those goals

then yes that's the right tool right so

that's the first thing and given as you

mentioned Kevin like how rapidly things

are changing uh pick use cases that has

a tolerance for getting it wrong so for

example pick use cases that are internal

facing or pick use cases that have a

human in the middle so you can Rectify

if there are issues uh the last one I

would say is

invest in tooling and technology and

platform what I mean by that is today we

run 20 plus models we have customers

asking us questions about the

performance the output of the models we

uh our teams have to debug fix issues so

we've built a lot of tooling to be able

to observe and scale the systems so I

keep him for uh reminding my team if

you're gonna build a ship be prepared

for a shipwreck right so while you're

yes focused on llm engine AI ensure you

have enough tooling and platforms to

support that

I'm sorry nothing else to add I think is

I I already mentioned everybody is if

that is precisely that I think my piece

of advice is if anyone tells you what

they know what channel is going to be

doing in five years they're lying to you

so you know stay humble stay stay

attached to the actual you know

short-term use cases because it's I'm

sure it's gonna be crazy in five years

but but we're we're to your point we're

all just like wait what did we announce

so absolutely all right uh we are out of

time uh love it if you uh you know give

us feedback here and I think we'll be

around if you have any other questions

as well thank you everyone

foreign