Building Smarter Software Robots with Robotic Process Automation & Google Cloud AI (Cloud Next ’19)

Building Smarter Software Robots with Robotic Process Automation & Google Cloud AI (Cloud Next ’19)

November 4, 2019 0 By Kailee Schamberger


[MUSIC PLAYING] FRANCISCO URIBE: Good
afternoon, everyone. My name is Francisco
Uribe, and I’m the product lead for the
Computer Vision and AutoML platforms. In today’s talk, I’ll be
delighted to introduce the Computer Vision platform. And then I’ll turn it over
to our friends at UiPath, Infogain, and
Automation Anywhere so that they can describe how
they’re using our technology to make their robotic process
automation technologies better. So the mission of the
Computer Vision team is to enable our partners
to build the next generation enterprise AI solutions. And this focus on the enterprise
has led, in the last few years, to significant breakthroughs
into the application of AI across different verticals. For instance, we have industries
applying our technology to detect damage in
their facilities, for [INAUDIBLE] maintenance
use cases or [INAUDIBLE] entertainment companies
using our technology to implement intelligent
content management systems. And the focus of
this talk, companies are using our OCR and natural
language understanding technology to derive
insights from their data, provide structure, and automate
complex business workflows. Now to recap, the
Computer Vision platform is comprised of two
sets of products. The first set of products
is our pre-trained products, and the second one is our
customizable AutoML products. The vision API allows you to use
models without having to code. It’s completely plug-and-play
into your application, and all you have to do
is just query these APIs with a rest API,
[? scrape, ?] and perfect for generic,
well-understood use cases. And our AutoML products can be
customized with your own data, leverage behind the
scenes, Google state of the art neural
architecture search, hyperparameter
tuning technologies so you can build
high-quality models with it. And this is great for more
complex, specialized use cases. Now these two
categories of products provide you a pretty
comprehensive set of tools to implement
Computer Vision models. But we know that
AI tools are only part of the equation in what it
takes to build enterprise AI. Enterprises today are still
facing significant challenges wrangling their data and seeking
buy-in to actually produce models at the last mile. So for that reason,
in Cloud AI, we determined that it
was critical for us to democratize this technology,
to partner with key companies, so that jointly, we can solve
the enterprise’s deepest challenges. And one key challenge that
most enterprises face today is that 90% of their
content is completely dark and unstructured, and
requires significant amount of human effort to be able
to understand and integrate into a business workflow. So for that reason,
yesterday we introduced Document Understanding AI. Document Understanding AI is
our [? partnerless ?] solution to help you understand
structured data– things like invoices, legal
documents, tax forms– and then automate business
workflows and improve your decision making. For instance, processing
an invoice today is a very, very tough task. Today, when you get
a physical invoice, a human needs to type the fields
directly into an ERP system. Now, imagine having
to do that across tens or even hundreds of
thousands of invoices. It’s a lot, a lot of work, and
it’s very error prone as well. With our technology, we
can turn a scanned invoice into a digital structure
JSON with all the fields so that then, together
with our partners, we can integrate this data into
a downstream business process, like a P2P business process. So this is, we believe,
very transformational. Now, at a high level,
Document Understanding AI helps you structure from
your unstructured documents with the goal of implementing
better organization of your data, enable search. And on top of that, you can
automate, repair your business workflows, and improve
decision making. And now, to explain
how this technology is being used in a real business
use case, in the RPA space, I’m delighted to
invite Mark, director of AI at UiPath and
Ravi, senior director of digital services at PepsiCo. MARK BENYOVSZKY: Thank you. Thank you, Francesco. So I’d like to start off by
talking a little bit about who UiPath is, in case you’re
not familiar with us as an organization. Today, we are the fastest
growing organization in the enterprise
company software history that’s growing
at the pace that we are. We have over 2,000
enterprise global customers that are using our platform
on a regular basis. And we have over
1,500 UiPath employees that are growing daily inside
our organization at a pretty steady clip. We have over 300
partners with whom we work with in a very
large, rich partner ecosystem that
enables our platform to be as robust as it is. We’re invested by multiple
investment partners. We currently have $400
million worth of investment, $3 billion valuation. And we are recognized as
a category leader for RPA, Robotic Process Automation. From a sponsorship
perspective, we have great sponsors, including
Excel, CapitalG, Kleiner Perkins, and Sequoia as
examples of the companies that are backing
us in this space. And just real quickly, from
an overall perspective, what is robotic
process automation and what’s UiPath’s
specific view of this? It’s really about enabling the
easy deployment of scripts that can take the work that
humans do on a daily basis and automate those tasks on
a highly repeatable basis. So these tend to be rules-based
scripts that we can basically, then, mirror what the
human does on the screen, working with applications
and moving data from one application to another. The way that we
make it very simple is that we work at that
UI layer, hence UiPath. So the integration
doesn’t require us to work if we don’t want to
necessarily at the API layer, or worry about going down into
the services layer or the data layer to be able to perform
this type of integration. So if you want, you
can integrate UI layer or API layer as well. The other compelling aspect of
why our customers are finding so much benefit from
the platform is it’s a low code environment. So business users can actually
start to work with our platform today to simulate maybe
a particular process area that they want to automate. And they discover
that they can actually do that process without having
any type of programming skills and take that process
and advocate it inside the organization,
ultimately to be implemented as part of a production effort. So one of the
interesting things that’s happening in the RPA space
is this beautiful marriage between RPA and AI. And so AI is really going
to enable a next generation capability of smart robots
so that, fundamentally, we can shift from robots that
are basically performing those rules-based tasks that
are highly repetitive in nature and, ultimately, transition
that to more of the increasingly complex tasks that exist inside
an organization that can’t necessarily be
codified by rules, and that require more
cognitive type of capabilities. So when we think about
AI or machine learning, how can we use
those capabilities, like forecasting and
probabilistic modeling and scoring and deep learning
type of capabilities, including computer vision,
to enable robots to be able to have
these new skills to be able to do better work, more
work, inside of the enterprise? So we see this as an
inflection point, actually. And it’s fundamentally going
to transition to industry from this rules-based structure
to a very intelligent-based robot that, again, can
take on more of the tasking that we do inside of
enterprises today. Now, if we take
this and we actually apply it to the space
of document processing– so every enterprise
in the organization deals with large volumes of
documents of varying types throughout the organization. And what’s important
about those documents is that key business data is
locked up in those documents. And sometimes those documents
are in print format. We have to digitize
them to be able to get that data in some type
of electronic format. In other cases, we
have a digital document that we can get that data from. But the important thing is,
how do we get that data out reliably and quickly
so that, ultimately, we can move from data to insights? Because that’s really what’s
important to businesses. I want to unlock that
data from my documents. I want to be able to
get insights and do further knowledge mining
up against that data. So what we can see
here is an example. What role does a
robot have in that? The robot can help automate the
process of collecting documents within the environment. So it can ingest documents
from an email box as an email attachment,
maybe invoices coming in from an email. If there’s a file
storage system, it can inspect a file system
and fetch documents from there. If there’s a blob storage
type of environment, it can also pull
documents from there. So that’s a great
skill for a robot to be able to go look for
the documents that we need. One of the benefits
that we can do is, once we have
that document, then we can assess the document
features and know, for example, a particular
document type is an invoice and that we need to process that
invoice in a particular way. So we provide a lot
of capabilities, working on the capabilities
within our platform to be able to look at the
pre-processing aspects of that document, extracting
the data from the document. And then any type of
post-processing activities to be able to get that
data fit for downstream and into production-based
systems, like an SAP, for example. So the value there is that,
when we extract that data, we get that data in a
very common data format, in an intelligent JSON
structure, which gives us key value pair
information, the extraction of table-level details. In addition to that, we know
the confidence level information that comes from the algorithm
relative to how well the extraction was performed. And that’s really
important because we can set a lot of rules up
against that type of data to be able to treat
high-confidence data differently than perhaps
low-confidence data. Also, when we have this
data, there’s a lot of things that we can do with respect
to leveraging classifiers. So it can have
document classifiers. It can find a document
that’s an invoice or a health care-based document or a
benefits-based document. So it can treat those documents
in a different manner. We can also use
that data naturally to be able to support additional
machine learning models around scoring algorithms or
specific custom fit for purpose models that are appropriate
for a particular business domain or a functional area. And then, essentially,
once we’ve got that data, the robot can actually
work with the human to be able to
accomplish a mission. So this is all about
humans and robots and AIs basically working together. The way that they do that is
through a human-in-the-loop process. So as we’ll demonstrate
later today, we can actually
extract that data, and if it doesn’t meet that
minimum threshold level of the confidence details that
come in that JSON payload, we can go into a
human-in-the-loop process, so that a human can actually
see what was extracted and then correct or validate certain
features from that data extraction before it goes
to downstream systems. After that process, a
robot can pick it up again and populate Google
Sheets or go into SAP, or go into a variety of
different landing sources that would include, perhaps,
some data repositories for the data mining site. So from and AI and RPA
case study perspective, this individual use
cases for invoices, but the capability can support
documents of many types. Invoices traditionally tend
to be manually processed, so there’s a lot of effort
associated with getting the invoice up and running. And invoices are
interesting because they have many different
multiple feature sets. So we see machine-generated
text, we see handwriting. We have objects on there,
like signatures and the logos. So it’s a consistently
hard problem to be able to actually process
these documents at scale. And so we’re changing the
game by working together and bringing our
investments in AI and our RPA platform in
our partnership with Google through Google’s
document-understanding AI capabilities to be able
to help reduce the costs and errors associated with
that manual heavy lifting of extracting data
or swivel chair when I re-key the data into
a terminal or into a screen. And then I can, as
a result, benefit from increased efficiency
and greater speed, and then start to look at things
like fraud and fraud detection in some of my invoices that
might be in a particular market where we’ve detected
higher fraud. So the approach is to extract
content from the forms and tables, and then basically
allow that human-in-the-loop process system to begin, and
we’ll actually demonstrate that capability later on in
a video presentation. But at a high level,
essentially what we’re looking at is the
supplier invoices can come in, the robot can pick it up. It passes it over to the
document processor from UiPath, where we then interact with the
Google document understanding AI to be able to get
the extracted value out. If we need to perform
that validation step, that begins that
human-in-the-loop process. And then the robot can pick
up the document afterwards, after the human’s
validated or refined the data, to be able to carry it
over into the back end systems. So with that, I’m
going to pass it over to Ravi to talk about how
we’re actually applying it in his environment. RAVI BOGGARAM: Thank you, Mark. So with co-innovation
with Google, as well as UiPath,
here, I’m going to demonstrate a sample
representative use case that we have
experimented with and have proven the use case. And here, we are
seeing a sample that is as close as possible to the
representation of every feature that we just heard from Mark,
ranging from unstructured data to the computer-generated,
machine-generated data is also there. Handwritten
signatures are there. Approval seals are there. And within the seal,
again, unstructured, handwritten
information is there. So different types of
unstructured information is available. So gleaning information insights
from that unstructured data is this use case. Essentially, we have here
all of that information and, most importantly, how the
bot, from the UiPath platform, is able to dissect that
information, different types of annotations, and is
able to progress forward. And when it works
on that information, it makes a API call to the
Google Document Understanding AI feature. And then the response
is coming back, as we heard from Mark, the
inner JSON structure format. And based on that information,
through a certain level of confidence factor,
we were able to decipher the unstructured
content information, whether it is the key
value pairs, including invoice number, the date,
the supplier details– particularly the most important
one, the transportation number. A number of similar key
value pairs are extracted, and it is intelligent
extraction. And more importantly,
it’s the line item detail with the table parsing,
which is another very important and powerful feature
from this Google Document Understanding AI API call. And with that, knowing
that extracted information and the quality and
its confidence level is extremely important. We could predefine
certain thresholds. And then once the
threshold is met, the bot can proceed
automatically to update the back
office systems, whether it is ERP, SAPR,
any other back office system accordingly. Like the swivel chair
type of information is not required
anymore by the humans. The bot is able to do that. You can see the
bottom picture, how that is dissected back
with the responses to the JSON structure. Every element is dissected there
and we have that information. And the threshold,
if it is not met, then we have the opportunity for
human-in-the-loop to override that information, do a
side-by-side comparison like this. You have the original as well
as the parsed information side by side. And then the human,
the analyst, is able to make an
override, if needed, based on the threshold not
being met by the API call returning the JSON
structure, which is seen here in the picture. And after the review,
the bot can continue. Once the human-in-the-loop
is updating, then on the next step, the bot
can proceed and then update the back office
systems as required. Here, like I mentioned earlier,
parsing the table structure is very important. One powerful feature
here worth mentioning is, it’s not a [? parse-specific ?]
table structure that you have to templatize, no. This one is a powerful,
computer vision-based, intelligent extraction feature
in that API call we make. And that’s why it’s
worth mentioning. That parsed information is
shown in the UiPath Studio on the left side image. On the right-hand side is
the actual invoice on which the extraction was done. And the validation station
is providing that view. And then if your
human-in-the-loop is there, all those fields can be
updated and overridden. If it is meeting the
threshold, then the system proceeds to the next
step automatically. And also, our additional feature
that Mark mentioned is robots can use intelligent
classification. That is inbuilt feature for
geography-based detection or candidate for audit
kind of detection. All these are
classifications that are available in the platform,
with UiPath platform. And then you could
dissect that information and classify them into
these kind of categories in the picture we
are seeing here as the representation
of categorization. For example, on the
right-hand side, you can see, after the
threshold, certain amount level you want to automatically
have an audit mentioned. So the bot catches it, and the
UiPath platform automatically mentions that it is a
candidate for auditing. Those type of features,
are all built in. This is ML-based, AI, computer
vision-based, combinatorial use case between the
Google Cloud Platform and the UiPath platform. And we did the co-innovation
with both the partners. And with that, the benefits
are of numerous cases. It is not just the unstructured
information dissection that we saw just for invoices. Honestly, the benefits
are much bigger than that. It registers the manual
effort phenomenally. It registers the
invoice processing time because the bot is doing,
in an automated manner, like hundreds of
thousands of documents processing in some cases. And here, the data is reviewed
and validated by humans. And also accuracy is
not going to be an issue because here, the bot is the one
which is actually acting on it. And also, another
powerful feature is you are actually able
to glean that information and catch it if it is an error
before updating the back office systems, like SAPs
and other things, because here it
is, upfront, known. And it allows humans and
bots to work together faster and with higher
performance over time. And as an extension,
this could be extended to various other
domains, like supply chain and in the warehouse and
various other places. And with that, I will
pass it back to Mark for the video to
demonstrate these features. MARK BENYOVSZKY: Excellent. Thank you, Ravi. So the best part,
if we could start– actually, let me tee this up. So the best part is to
actually see it an action. So you saw the slideware,
which was great. But we’ll actually demonstrate
how these capabilities work in the platform. In the first instance, we’ll
go ahead and get this started. If we could roll the
video, it would be great. So here, we’re showing
how we can actually start the UiPath
robot that’s already been built to be able to
look into a particular file directory and see that there’s
documents to be processed. We have a number
there that we can take and we can actually upload
those to Google Cloud if we wanted to
have some storage up in the cloud for that
particular document. It can then be further processed
by the Document Understanding AI. And then we can
determine whether or not we want to actually
look at that document based upon the
extraction results. So if we never had seen
the document before– the first time that
we saw an invoice, for example, we may not
have all the features from the extraction associated
with the taxonomy that’s associated with invoices. So the human can
go in very quickly and actually highlight
the actual aspects of the invoice number,
the purchase order number, the date,
who the supplier is, the details with respect to
address information, billed to and shipped to– the
typical common header type of information
that we would see, as well as the subtotal
and the total information. So it’s a very quick
process, again, how the robot can make that
available for the human to be able to do this in a
very easy UI type of way, and then being able to
complete the process. So another example is,
what if we actually meet the minimum threshold value? So if we uploaded
the second document, it goes into Google
Cloud, into Buckets. But we can see that the
Document Understanding AI is capable of extracting
the features there. If it meets our criteria for
the minimum threshold level, then we wouldn’t necessarily
have to have that human loop process to be able to exist. The document could
continue along its way. If we have another document
that we take through here to demonstrate
the table capability, as well, as Ravi
pointed out, that’s a really powerful
feature here to be able to actually parse tables. The confidence level might
be a little bit lower. But what you can see
here from the invoice is the detailed level
data that’s coming out of the invoice itself. So that’s a particularly
challenging task to take. Oftentimes, it’s pre-scripted or
templatized, as Ravi mentioned. Here we’re just using computer
vision and the ML type of capabilities to be able
to detect that information, parse it out correctly. And again, through a
validation station, the human is basically able
to see the extraction side by side with the actual document
itself for that verification. So another thing that’s
interesting is just to see, what does a robot see? So there’s a JSON payload. We can take that data
and naturally, we can feed it into something like
Google Sheets if we wanted to. And we could see
all the data that’s coming out of the actual
extraction process itself in a human,
legible type of way, as well as the detailed table
data that comes out of that. Or we could use
this and analyze it further if we had some
reporting or analytics dashboards that we wanted
to be able to put up against this data. But more importantly,
probably what we want to do is we want to take that
invoice data and we want the robot again now
to pick up that data and start to log in
to SAP on our behalf. So the robot’s basically
then authenticating with user and password
credentials here. Automatically knows
that it’s in the task to be able to
input invoice data. It can go to the appropriate
screen with an SAP to be able to input
that data and put all the corresponding
information that it extracted from that process automatically. So again, the robot is
actually performing these tasks that a human would typically
do in the environment– if we could stop it
here, it would be great– without having to
necessarily go in and perform that function itself. So what we’ve
demonstrated here really is an incredible end-to-end
type of capability. This particular example
is for invoices, but it can be applied to any
type of enterprise document– unstructured, semi-structured,
and structured. And another powerful
aspect of the platform is that it can be truly
used at an enterprise level. And just finally,
to land here, we have different graphics
that Ravi showed before as a classification technique. Our time is out. We really enjoyed speaking with
you, and thank you very much. FRANCISCO URIBE: Thanks
a lot, Mark and Ravi. That was an awesome
presentation. And thanks a lot for
being a great partner. So next, I would
like to introduce Abhijit, our SVP of
product and engineering at Automation Anywhere and Gans,
BPO of the digital experience and insights at Infogain. ABHIJIT KAKHANDIKI:
Hey, thanks Francisco. All right, guys. So I’m from Automation
Anywhere and we are, again, one of the leading vendors
in this entire space. We’re the ones who
actually coined the term “digital workforce.” And the idea is,
just like a human, you actually execute
on things or do work. Then you think about
things, you analyze things, and you go drink your coffee. Well, there is a digital
equivalent to that. And that’s actually what our
product portfolio consists of. So we have our core
enterprise RPA platform. And the bots can do– they can
work with 400 different types of frameworks. Actually, the way I put it is
bots can see a couple of levels deeper than a human can. And they are less
likely to make mistakes. So the core RPA
platform can work with any of the applications
that are out there. Then we actually have
had a cognitive component to our product line since
the last four years. And again, we have
several applications of that, including
understanding documents, as well as processing documents. We have a bot store that
is public and open where you can find 30 to
40 different domains with all kinds of different
documents being processed. We also have a smart BI
platform that actually gives you real time, not just operational
insights about the bots– what each of the bot is up
to, all kinds of bot security analysis as well. But what we do is we also
work with the content that is processed by the bots. And that gives you
real-time business insights. So we believe that
RPA can actually play a major role for you to
spot business opportunities. And it’s not just about cutting
down process costs and cycle times. So this together is
the digital workforce. The digital workforce
is available to you in our bot store by process
or by digital workers. If you look at any
business process, it consists of these steps
with respect to data. You are either
capturing the data or you’re enriching it,
validating it, processing it, reconciling it, and analyzing,
and reporting on it. So traditional RPA, while it
tackles majority of the steps, AI is still needed in
several different aspects. So we have AI embedded
in the core platform. So core, computer vision
based on deep learning that will actually, even if we
don’t understand the framework, even if we treat the
screen as a giant image, the bots are still able
to work 100% reliably. But we have expanded this and
formed this true partnership between RPA and AI to be
able to embed any AI models. So one of the highlights
of the platform is the fact that you can
actually embed traditional– you can embed AI models from
providers, such as Google, as well as you
can have custom AI models that you build yourself. So for example, running a
Python script and so on. So we are delighted to be here. We are partnered with
Google at a very high level. And it is not just
about using Google AI. But it’s also about running
bots in the Google Cloud and, finally, humans
interacting with the bots through your Google
productivity suite. So we have a very
interesting use case here. And here to talk
about that use case on how we have leveraged
the Google Computer Vision API to prepare for 10x
growth at a customer, where achieving this kind
of transaction growth was just not possible
without using bots– and so here to talk
more about is Gans from our partner, Infogain. GANAPATHY SUBRAMANIAN:
Thank you, Abhijit. A quick note about Infogain. We are a 30-year-old
software engineering firm. We are based down in Los Gatos. And we’ve got offices
all over the world. We are partners with some of the
leading technology companies, and we assist Fortune 500
companies on their journey to digital. We are partners with Google
and with Automation Anywhere. And we’ll talk about a
very interesting case study that we’ve got here
for our customers. What we’re going to talk
about is a customer here, which is basically the world’s
largest travel marketplace. And they’ve got more than 400
million guest arrivals per year on their web properties
and mobile properties. And one of the interesting
problems they had was the customer segment where
you had more than 10 properties that you wanted to
transact on their platform. These properties– if you
ever went to their website, you would see that there’s
almost 10 pages of data that you have to
provide in order to list your property
on their platform and go through all the stuff. So highly manual process. It took a lot of time. It took up to 30 days for the
entire process for a property to be listed. Five to seven days was
just the data part of it. And though there was a lot of
onboarding assistant provided, the whole onboarding assistance
was provided manually. So you had to call up a team. The team would go
through it and stuff. What was causing a lot of
friction in this experience for the customers was basically
the fact was, because this was a manual process, they were
seeing a very low activation rate. Activation is the process
from a new sign-up to have your property on
the site to actually doing the transactions on the site. And that was in the low
30s as an attrition rate. And that’s primarily because
of being a manual process. You had poor data quality. So you had a lot of
errors in the data that has been processed. There was incomplete
attribution. So you couldn’t
run those metadata to run campaigns and
promotions and search. And this process was being
done on a much smaller scale of volumes. And the customer was
anticipating 10x growth. And this was just not a scalable
process for them going forward. And when we worked with
Google and Automation Anywhere to help this customer, there was
one simple goal that they had, which was, how do we optimize
the conversion funnel? Which means if I sign
up a number of customers on the site, I need them
transacting more and more. That was the primary thing. So if you look at the
conversion funnel, there are three steps in
the conversion funnel. One is you have to create
a listing on their site with a property. And that was something that
was taking a lot of time, like we talked about. Second is, how do you enrich
this listing with amenities, data, and other data so that
you can activate this listing and put it in front of
more users out there? And the third thing
was, how do you create more and more
better data so that you can earn promotions and search? So this was the goal
with which we did it. And that’s where we first
brought in the robots that Automation Anywhere
was, and we hosted that on Google Cloud. And we basically had the robots
doing the listing creation, and also gathering
the data so that a lot of that can be
automated and someone does not have to do it manually. And once that
listing was created with all the data on it, it
went to a human person, a team out there, which approved the
data, put the listing data out there. And once that was done, we
used the API integration into Google Cloud to
basically transfer the data to the web property
that our customer held. And once that was
done, the listing was happening for the website. And this was, earlier,
all done manually without APIs and without
a cloud infrastructure is all being done automated. Once we did that and we
started seeing some benefits for our customer, we
moved onto the next stage of the project, which is
basically the fact that you’ve got something automated. You’ve got something– how
do we scale this automation? So scaling that automation
had three basic parts to it. As you can see, the first
thing is, in the first phase, we already improved
the turnaround time and we made it a self-service. So we also integrated Jira into
the Google Cloud infrastructure so that customers can
look at their own tickets and figure out where
their property was in the whole process
of onboarding. But we also wanted to move
this to international scale and move it to more countries. And that’s where Google Cloud’s
whole scale and infrastructure came into play. And for them to roll it
out to multiple countries was much easier. And the last stage
of the product was to basically use
Google’s AI services to see how we can monetize
that better data that we can pull out of the properties. And how do we move it to
new product categories and new language categories,
in the sense of, how can we use NLP, how can you use
language translation services that Google has
so that we can move this? And this was the
roadmap that we used working with Google and
with Automation Anywhere to move the customer
to a larger scale. What we particularly used, and
what I want to talk about here is we actually embarked on
a very fast and effective integration of
AutoML vision models into the existing
robots and the process. This was a whole
project that we did in eight weeks that goes
to show that ease of use and of the AutoML product. And we basically took all
the properties that they had. We took 1,000 images. So each property comes
with some images. And in the previous
world, someone had to sit with the images
and look at each image and see which amenity
was in the image. And amenities are very
important if you ever tried doing a vacation
rental or a hotel, because you want to
see the property. You want to see what it
is in it, whether it’s got a microwave, whether it’s
got a fridge, whether it’s got a bathtub. And that’s more so in
the international market. So we took the 1,000 images. We did the training
on the images. And we did multiple iterations,
five-plus iterations on the training. And each iteration
took around two to four hours of model
training time to do that. So it was fairly swift,
fairly good to do it. And we looked at three
to four major amenities– the TV, bathtub, and a king bed. Based on these amenities, the
pricing recommendation engine of our customer could actually
recommend more effective prices that you should be charging
for your properties. And in eight weeks, we
were able to demonstrate a 90% accuracy and upwards
of 90% across all categories that we see out there. And that’s where the
AutoML vision was not only fast and efficient,
but it was very effective in generating the
data that the customer wanted. If you look at what the customer
is looking at, going back to the problem statement
that existed out there, one of the things
was this customer was doing this entire
process manually, which means they were
probably taking a month to do this entire process of
which just the data capture part of the property
was five to seven days. And if you know anything about
properties, in that one month, you lose your customer
because he’s focused on to something else and stuff. And that’s where the
activation problem really came in because in a month,
you were losing them. You didn’t know where
the property was. And activation was also
missing because, like I mentioned earlier, the
amenities data was not there. A lot of metadata was
not there for them to promote their property
among the search engines and promotions and
stuff like that. So by integrating Google’s
AI services and APIs and automation
[INAUDIBLE] robots, we were able to demonstrate
100% increase in the activation rate, which is direct
revenues for the customer that exists out there. So that’s something
that we were able to do. The total cost went down 40%. I mean, being a digital
unicorn, they really never worried about cost. But it’s very important
as you scale the process, you want to make sure that, per
transaction and per property, your cost goes down. So we were able to demonstrate
a 40% reduction in cost. But I think, to me, the
most important fact was– and which was causing
the biggest friction among the users– was we were able
to turn around– the turnaround time was
reduced by almost 70% to 50%, which is basically the
fact that something which took five to seven days
was being done in two days. This is very important
for them because they could get back to
their property managers fast and tell them that,
hey, we signed up you. You are on live and
stuff like that. And last but not
least, I think, as you move a process to scale– they are looking at upwards
of 150,000 properties over a calendar year to scale up
to a million-plus and stuff– it’s very important
that they go about doing this in a self-service model. So for example, if there
are exceptions happening, if something needs
to be changed, something needs to
be done offline, we really put in a self-service
automation module in place. And this is where the
ability of integration with Google Cloud and
Jira and the RPA robots really came to speed. It was not a custom project. We could do that very quickly. It did not require generating
custom APIs and stuff. So that’s something that we did. And by doing all this,
like you see out there– I talked about it– 96% accuracy and 90%
recall is what we came up with the training model. So this is now being implemented
across all countries. And the future state is we
are going to increase it across multiple
amenities so that, as you go across using
these properties, you will see that this is
going in a much faster way than you can. This is one of the
use cases that we have seen with our customers. And Google’s object detection
AI was very useful to it. I’m just going to
turn it on to Abhijit to talk about other cases
that we can come across. ABHIJIT KAKHANDIKI:
This is amazing, right? I mean, what can bots do next? I think what I like
about this, it’s the application of
computer vision. But it’s also the
realization of quick results. Given that this entire project
was done in about eight weeks, it really prepared the
customer to scale their number of transactions with bots. And that’s one of
the key value props that RPA brings to the table. RPA plus AI will give you
a completely new level of [? auto ?] [? I ?] because
RPA helps AI in three ways. It brings data. It brings training
to AI, as well as it brings the
business context to any of these services. So that’s just one
of the use cases. We have plenty more that
are actually scattered across different verticals. So if you want to
learn more about this, please visit us at our booth. We also have a video of this
process with the bots in action where how they actually go from
that beginning of the property listing all the way
to recognizing all of the different
objects in there, and being able to provide
a response and then human-in-the-loop, right? So the same concept of having
the confidence threshold and all specified, and if it
is below a certain threshold, a human needs to be in the loop
to be able to review and train the bot further. So this is basically AI. As it learns, it becomes better. And the bots can actually manage
more and more by themselves, requiring little or no
supervision eventually. So like I said, visit
us in our booth. It’s Booth 1660
[INAUDIBLE] down here. And we can talk to you about
several other such use cases. Thank you, guys, for attending. Francisco, I don’t know
if you wanted to say– [APPLAUSE] Thanks. FRANCISCO URIBE:
Thank you very much. Just to wrap up, I wanted
to invite you to go to cloud.google.com/document
understanding ai. If you want to learn more
about these capabilities, sign up and be connected to
one of our prime partners. Thank you. [MUSIC PLAYING]