AWS re:Invent 2022 - JPMorganChase real-time agent assist for contact center productivity (AIM307)
AWS re:Invent 2022 - JPMorganChase real-time agent assist for contact center productivity (AIM307)
Resolving complex customer issues is often time-consuming and requires agents to quickly gather relevant information from knowledge bases to resolve queries accurately. Join this session to learn how JPMorganChase built an AWS Contact Center Intelligence (CCI) real-time agent assist solution to help 75 million customers and help 8,500 servicing agents generate next best actions in the shortest time—reducing agent frustration and churn. Hear how JPMorganChase’s real-time agent assist solution uses Amazon Transcribe to provide real-time transcriptions, Amazon Kendra to provide best answers from knowledge bases during a live call, and Amazon SageMaker for machine learning and training their intent models.
Learn more about AWS re:Invent at https://go.aws/3ikK4dD.
Subscribe:
More AWS videos http://bit.ly/2O3zS75
More AWS events videos http://bit.ly/316g9t4
ABOUT AWS
Amazon Web Services (AWS) hosts events, both online and in-person, bringing the cloud computing community together to connect, collaborate, and learn from AWS experts.
AWS is the world’s most comprehensive and broadly adopted cloud platform, offering over 200 fully featured services from data centers globally. Millions of customers—including the fastest-growing startups, largest enterprises, and leading government agencies—are using AWS to lower costs, become more agile, and innovate faster.
#reInvent2022 #AWSreInvent2022 #AWSEvents
Content
0.21 -> - Hello everyone. Welcome to our session.
2.52 -> My name is Vafa Ahmadiyeh.
4.8 -> I'm one of the AWS principle
solution architect.
8.88 -> I'm working in global financial sector.
13.23 -> I'm part of a large team
that we are looking after
15.96 -> JP Morgan Chase with a migration to AWS.
19.89 -> I'm honored to have Ami with
me on the stage from JP Morgan
23.7 -> and we are, we have created a
25.89 -> fantastic talk to go
through JP Morgan Chase
29.7 -> at how they are using AIML
services to improve the
33.57 -> productivity in the contact center.
36.87 -> So I'll let Ami introduce herself
38.61 -> and then we get through the talk.
40.95 -> - Hi everyone,
41.783 -> I'm Ami Ehlenberger and I'm the
CTO for machine learning for
44.82 -> operations within Chase.
46.74 -> We focus on predictive modeling,
48.75 -> we focus on NLU and NLP
solutions and really
52.83 -> contact center optimization,
which is why we're here today.
57.39 -> - So, let us walk you
through this journey.
60.66 -> Customer service is important.
63.39 -> We all know that the customer
service is important,
65.67 -> but how important is it really?
67.59 -> Let's look at the number.
69.81 -> On a survey on a PWC survey
that, that we have done,
73.83 -> 75% of the people said that
having a good customer service
79.83 -> is a really key factor for
them to choose a provider.
83.52 -> I personally happy to pay a
bit more every month to make
87 -> sure that I risk my bank or is my
91.53 -> phone provider or internet
provider happy to pay more,
94.29 -> and when I've got an issue
and I'm grabbing the phone,
96.72 -> calling the customer service,
I receive what I'm expected.
101.25 -> But that more interesting number
is that half of the people,
104.97 -> or almost half the people,
106.68 -> they will turn around and find
another provider after just
109.89 -> one bad experience.
112.08 -> That's, that's quite
astonishing that number.
115.17 -> But on the positive side,
85% of the companies,
118.92 -> they are telling us that they
actually can improve their
121.74 -> customer service through
the call center where they,
124.623 -> when they introduce technologies
and when they're bringing
127.92 -> AIML to improve the
productivity of their agents.
131.97 -> So we are going to dive deeper
into these If you haven't
136.44 -> done this at this reinvent,
we are trying something new.
139.5 -> So we got the pole in,
you can grab your phone,
142.68 -> scan this or you can go on
AWS reinvent app on the right
146.94 -> hand side, click on more and
the third is live polling.
151.5 -> Just put that number 307177
155.61 -> and let us know that what,
158.01 -> which of the following is
actually your contact center
160.89 -> challenges or your customers are facing.
164.4 -> I'll give you a few seconds
to see how is the room feeling
168.84 -> about these challenges.
171.3 -> Okay,
172.133 -> so all of them basically
that wasn't surprising.
175.98 -> When we were doing other sessions,
177.39 -> It seems that these
are the challenges that
181.71 -> most of our customer are facing.
184.41 -> I think in the equal number, 32%, 18, 25,
187.71 -> that with your existing call
centers or with your customers.
192.48 -> Let's move on.
193.86 -> So, what you can introduce with
199.14 -> AIML to your services is to basically
203.73 -> not only tackle the time
that the customer is on
207.36 -> the call, by two ways.
208.92 -> One is that you want to transfer
211.77 -> the call to the right agent.
214.17 -> I think at Chase we call them
specialist is that right?
216.81 -> - That's right.
218.347 -> - So you want to find
the right specialist.
220.44 -> Especially for example for
for a large organization like
223.26 -> Chase, if I've,
224.46 -> if I've got a credit card
problem or a debit card problem,
227.25 -> there are two maybe different specialists.
228.81 -> So while I'm calling the call center,
231.107 -> I want to talk to that specialist
first before going through
234.72 -> department to the department.
235.89 -> Not only that, we want to make sure that
238.41 -> we empower the agent to
resolve my problem much,
242.064 -> much quicker.
243.81 -> So the other issue is that we
saw the result is high agent
248.73 -> turnout but introducing AIML,
251.61 -> you can actually make
their life a bit easier.
254.19 -> You can, what you can introduce is
256.65 -> things like chat bot or self agent
260.52 -> assist to answer some of those
questions earlier or having a
264.42 -> good knowledge base system to
bring all the information that
267.75 -> that agent needs to be able
to resolve the problem much,
270.63 -> much quicker.
272.43 -> And obviously got tons of data
that you can find out that
275.49 -> what are the areas that you can
276.81 -> improve in your customer service.
279.81 -> More importantly, cost saving.
282.27 -> You can do a lot of cost
saving when you introduce AIML.
285.9 -> Let's dive deeper.
288.09 -> At Amazon, we got two options.
290.76 -> Amazon connect, I'm sure almost maybe
294.33 -> all of or majority of you, you
know what Amazon Connect is,
297.75 -> it's easy to use cloud contact center
300.81 -> with a built in AIML in it.
303.06 -> So that's one solution.
305.16 -> If you have worked,
306.06 -> I have worked in the past
with designing call centers.
308.82 -> Sometimes it takes six months
to 12 months even to just
312.21 -> design networking and everything
else with the call center.
315.74 -> In Amazon Connect, five minute few clicks.
317.69 -> You got the call center,
319.35 -> literally got call center
to use every single of you.
322.08 -> You can just go on your AWS
account If you don't have
324.016 -> AWS account, create a free trial,
325.762 -> just have Amazon call
center within five minutes.
329.88 -> The other option that we have
got is what you call CCI.
332.76 -> How many of you by raising
hand heard about CCI?
337.56 -> Okay, not many but quite few.
340.59 -> So today we are going through
the CCI solutions or AWS
344.31 -> Contact Center Intelligence.
347.52 -> What we heard from the customer,
349.11 -> obviously Amazon Connect is,
350.25 -> is a good service for really
easy build contact center for
354.24 -> many, many of our customer but
many of our large enterprise
356.88 -> customer like Chase or many
other financial sector,
360.33 -> they already have got a call center.
362.37 -> Either it's Genesis or
Cisco Y or anything else.
366.03 -> CCI is actually a
combination of the services
370.32 -> that we created the solutions to integrate
372.66 -> to all of these no matter what call center
375.81 -> that you've got to bring
AIML services into it.
379.29 -> So CCI is not a product,
382.23 -> it's not off the shelf product
that you buy and it's not a
384.99 -> new service. Is it too complex?
387.36 -> Not really.
388.193 -> It's just a combination of
the services that we already
391.084 -> created some solutions because there is
394.14 -> no one T-shirt fit all.
396.48 -> With every enterprise
customer that we have got that
398.727 -> got different needs.
399.78 -> So we had to make it flexible
enough that you meet your
403.41 -> goals with the CCI
solution that we've got.
407.73 -> So there are three use cases with CCI,
411.03 -> self-service virtual agents
that for example chatbot,
415.65 -> we talked about it or IVR solutions,
418.56 -> when you would like to resolve
the customer issues before
422.7 -> even you transfer the
call to a live agent.
425.61 -> On the right hand side you
got post call analytics.
427.98 -> So that's where you are going
through all of your data,
431.85 -> either is for a month or for
the whole year or for the week.
435.39 -> And you wanna go through all
of your data and see what are
438.687 -> the common problems, where
are the area of improvement,
442.26 -> where I can focus what,
443.4 -> what are common problems with my customer,
445.56 -> what are the common problems
with my agent that I can
447.96 -> improve that.
449.37 -> So we've got post call
analysis, CCI solutions,
451.95 -> we got self-service virtual
agent or the middle one,
455.22 -> which in this talk we are
going to dive deeper into that
457.83 -> which is realtime call
analysis and agent assist.
462.72 -> Behind the CCI is basically
these AIML services.
466.38 -> If you need to create a IVR,
469.05 -> if you need to do a conversational AI,
471.66 -> you will use Amazon Lex.
473.94 -> Obviously as soon as you try to do IVR,
476.55 -> then based on your business need,
478.02 -> you need to do some text to speech,
479.7 -> that's where Amazon Poly is coming.
481.59 -> If you're in an environment
that you need to do
484.051 -> multi-language, obviously
you need Amazon translate.
487.29 -> So that's come to the picture.
488.82 -> You can think of them as a
Lego boxes and we created the
491.52 -> solution to put them
together. You might say that,
493.62 -> well I don't need the translation
494.79 -> but I do need this and that.
497.49 -> For example, a knowledge base,
499.38 -> you might need a knowledge
base to help your agent that
502.62 -> I call the bank say that
I've lost my credit card.
505.17 -> Then you can immediately tell
the customer that's okay if
507.66 -> it's to do with losing credit card,
509.015 -> these are the things that
you can tell the customer.
511.8 -> For that we have got intelligence search,
514.05 -> which is Amazon Kendra that you can use
516.63 -> for your knowledge base system and
518.76 -> Amazon transcribe and I'm so
happy that we have got one of
522.48 -> our senior product manager from
Amazon transcribe toward the
525.84 -> end of this talk.
526.673 -> We are going to dive deeper
into the architecture and we
528.66 -> have a QA to answer your questions
530.649 -> about Amazon transcribe as well.
533.34 -> So, and we go through the use
case of the JP Morgan Chase.
536.58 -> So if you need to do speech to text,
538.62 -> Amazon transcribe is one of the services
541.11 -> that you're going to use.
543.57 -> Who has heard about TCA?
546.69 -> Transcribe Call Analytics?
549.57 -> Okay, perfect.
551.46 -> That's actually announcement
we had yesterday.
553.56 -> TCA was enabled Transcribe Call Analytics.
556.17 -> We had it for a year or two,
558.33 -> and it was doing fantastic things for you.
560.76 -> So with TCA,
562.92 -> what it was doing before,
it was doing the batch.
565.02 -> So you were putting all your
call recording into a three
567.99 -> bucket and then you were running a job
570.51 -> against that three bucket.
571.607 -> You would say that, okay,
572.44 -> these are a month of call
recording that I've got and out of
576.66 -> the box they will give you three things,
577.98 -> which is really, really important.
579.45 -> So, if you wanna know that what is
582.163 -> the main issue of each call,
583.98 -> you cannot go read all the text
or listen to all your calls.
587.82 -> So one of the things that
TCA will give you main issue,
591.15 -> what is the main issue with that call?
593.58 -> That gives you out of the box.
595.47 -> The other thing that TCA will give you is,
598.02 -> which is actually very,
599.1 -> very good for improvement
in your call center is that
602.85 -> has my agent promised anything?
605.49 -> Is there any action that I
told the customer that will get
608.67 -> back to you and then you can
put them in a category and then
611.31 -> go back to see that have you
actually done what you promised
613.71 -> to the customer or not?
615.96 -> That's the other thing
that TCA is giving to you.
619.14 -> The third one is a call summary.
621.63 -> It gives you a call summary
of the of each call,
625.53 -> but more importantly TCA has
got an engine that you can,
628.835 -> a rule engine that you
can create the rule,
631.71 -> you can create as many rule as you want.
633.39 -> I'm sure there is a total limit,
634.44 -> I can't remember top of my head.
636.15 -> But you can create many rules
to say that for example,
639.63 -> I want to know all the calls
that there is more than a
642.81 -> minute silence in it.
644.43 -> I want to know all the calls
that for example in the last
648.96 -> 60 seconds, the sensitive
of the call was negative.
652.59 -> I wanna know the calls that ended badly.
655.95 -> One of the things that
our customer using a lot,
658.62 -> I wanna know all the calls that the agent
660.87 -> at the end didn't say that,
662.46 -> can I help help you with anything else?
665.34 -> I want to know all the calls, which is,
667.53 -> that's one of the most use
cases that the customer on the
670.74 -> phone says that I wanna
talk to your manager.
673.35 -> You wanna, you wanna know those calls,
675.03 -> put them in a category and
then take some actions.
679.08 -> So this was exist in TCA
or transcribe chronologist,
684.03 -> what we announced yesterday,
you can do that all live,
687.36 -> you can do that all in real time now.
690.18 -> So imagine a call center,
I'm an angry customer.
693.66 -> I ring and I say that I
wanna cancel my subscription.
697.05 -> Live, you can send it,
698.94 -> you can flag that call and you
can send your supervisor to
702.45 -> go and help that agent. Before
that call get escalated.
705.84 -> I'm gonna show you them
at the end about this.
710.22 -> Gonna ask you another question.
711.75 -> Let's have the phones ready.
714.57 -> Before I hand over to Ami,
716.25 -> I quite like to know that what
are these call centers you
719.82 -> are actually dealing at the moment?
721.62 -> Let's see where we are with the results.
725.52 -> Oh wow. Again, we see that
that's not unexpected.
730.35 -> That's what we are thinking
that we get today and that's
734.49 -> what we are actually
experiencing with most of our
737.46 -> enterprise customers as well.
739.8 -> In exact same. I've got one.
745.74 -> Perfect. Let's go back and I've got,
749.73 -> this is the last question,
750.84 -> I promise I'm not gonna do that again.
752.73 -> So I'd like to know that what you have got
756.18 -> in the pipeline for next year.
759.69 -> Let's check the result.
763.74 -> Oh wow. That's interesting.
766.14 -> 'Cause in another talk was
lot more on the self-service
768.69 -> visual, it seems that that that's perfect.
771.45 -> So we'll go through at
the end of the slide,
774 -> we'll go through how you can
get help from what we have
775.98 -> built in CCI solutions
and how we can help you.
780.39 -> But let's go through the
fantastic journey that JP Morgan
784.44 -> went to build live agent assist.
788.76 -> - All right, so I like
to start at the top.
794.057 -> Technology at JP Morgan.
796.8 -> There's a lot that JP Morgan
Chase has to be able to do.
800.88 -> We focus on scale, on
reliability and security.
805.35 -> And then we overlay the ability to
807.48 -> innovate using AIML as a part of our
810.15 -> key investment strategy.
812.1 -> This is a across all pillars
and it is something that we're
814.8 -> intentional on.
818.19 -> When I break it down to Chase,
820.89 -> we actually stood up a focused
team a couple of years ago to
825.33 -> start solving operations
problems through automation and
828.54 -> applying AIML.
830.49 -> And the reason we're focused here,
833.76 -> is some of these key pain
points that you see today.
837.54 -> Right now Chase has 66 million
households and 5 million
842.19 -> small businesses that we run and service
844.95 -> on a day to day basis.
847.14 -> With those customers,
850.02 -> 62 are digitally active,
854.67 -> 48.9 million have our mobile app.
858.78 -> These are double digit
growth year over year.
861.69 -> Every, I was gonna say every
day, it's not every single day,
865.62 -> but we are continually pushing
new self-service capabilities
868.8 -> out into the market.
870.36 -> We wanna make sure that we're
meeting the customers where
872.46 -> they are and where they wanna be.
874.65 -> Personal finances and
banking become emotional.
878.13 -> This isn't something that I
can drop the ball on when they
881.46 -> try to make a payment, when
they wanna do a transfer,
883.47 -> this is affecting their
day-to-day lives and we have to be
886.53 -> there for them.
887.67 -> One of the other driving factors
for the kinds of work that
890.13 -> we're after is the fact that
not only is money personal,
894.9 -> but people are using multiple channels.
896.88 -> 50% of our customer base
are starting online,
901.972 -> going to the app.
903.9 -> They, they interact with us in all facets.
907.71 -> Every channel they come into
our kiosk, they're at the atm,
910.62 -> they're in our branches and
then they call our call center.
914.34 -> Bringing everything we know
about them together allows us to
917.25 -> personalize and experience and
understand what they wanna be
919.95 -> able to do.
921.63 -> So with all of these digital
safe service capabilities
924.45 -> coming to life and new
capabilities coming live,
928.26 -> not every day but really frequently,
931.5 -> how do we keep up and why do
I still see 32 million calls
936.33 -> coming into a call center?
939 -> So, let's take a look at the
Chase call center experience.
944.73 -> I don't know how many of you
here are Chase customers.
946.86 -> If you are, thank you.
948.27 -> I'm hoping to continue
to earn your business.
950.49 -> But while you're here,
951.84 -> the first thing we do
is put you through IVR.
956.1 -> Sometimes it's great,
sometimes it's terrible,
958.95 -> but we are focusing on improving,
961.23 -> using NLU and other automated
prompts and right now,
965.91 -> I can cut out probably two
thirds of the calls by allowing
968.73 -> self-service and automation
to take over in that space.
973.23 -> So once you're authenticated
and if I haven't gotten you
976.47 -> through or self haven't delivered
an ability to self-serve
981.27 -> in that channel, we'll
connect you to a specialist,
984.84 -> but the specialist will
then validate your identity.
988.62 -> But problem identification is
the piece that continues to
992.07 -> add swirl. How do I know at
which level of granularity,
995.97 -> how do we be able to knit
through everything they say,
999.54 -> there's a backstory for
how they lost their card,
1002.63 -> there's a reason they need to
be able to transfer things.
1007.34 -> What is the disambiguation we
have to go through in terms of
1010.31 -> clarifying the right intent so
that we're creating the right
1013.7 -> task and workflow for our agents?
1016.49 -> And then finally we wanna make
sure that this is handled and
1019.55 -> dealt with and closed out.
1021.35 -> First contact resolution is our goal.
1024.35 -> So, after the call we do a wrap up.
1029.09 -> Now the the quality of the
wrap up really drives the
1031.76 -> insights and the analytics
capabilities of our teams as well
1034.55 -> as the models and the accuracy
of the predictive power that
1036.77 -> we have using the entire
data chain and not only just
1042.069 -> thinking of it as an exhaust
but thinking of it as a
1044.78 -> product. How can I consume
it and make more, more value?
1050.03 -> So what can I do with AIML?
1052.58 -> Again, self-service
automation, just take it out.
1055.79 -> How can I make it easy for a customer
1057.44 -> to do what they wanna do?
1059.69 -> Second, understand why someone's calling.
1062.81 -> This is,
1063.643 -> this does more than just making sure
1065.21 -> that the agent is ready.
1066.35 -> How do I feed that back in to
the capabilities and feature
1069.68 -> function on our,
1070.513 -> our backlog for the feature
roadmap for digital?
1073.88 -> How do I make sure that our chat bots
1075.62 -> can handle these things?
1076.58 -> How do I make sure that live
chat can handle these things
1079.43 -> and how do I make sure that
these are easy and seamless
1081.77 -> workflow processes? Omnichannely
1087.35 -> One of the things we haven't
talked about yet is making sure
1090.14 -> they have the right
answer at the right time.
1093.23 -> When you are accompanied
the size of Chase,
1096.56 -> we do have a ton of different
product offerings out there.
1100.34 -> How do I make sure that the
right one with the right answers
1103.52 -> are in front of the agent at
that time so that you don't
1105.833 -> have to sit on hold for 90
seconds while we're going through
1109.55 -> a 500 page document to find
that one answer about that one
1113.63 -> APR, about the third card
that you have in your wallet,
1117.2 -> making sure that this is
at their fingertips is key.
1120.26 -> And then having accurate data.
1121.79 -> I talked about that call
summarization capability and we're
1124.31 -> very excited about the
announcement yesterday.
1127.61 -> So, a couple of years ago,
1132.38 -> we went through a pandemic.
1134.36 -> We had just launched this
idea of we're gonna stand up
1138.23 -> an AI focused organization,
1140.03 -> we're gonna optimize the call center.
1142.85 -> Then the world changed and our
agents started working from
1147.53 -> home, hybrid environments
two days in the office,
1150.32 -> two days back somewhere else
and the ability for their
1153.14 -> natural ability to team work together
1155.306 -> and ask for help went away.
1158.78 -> So it created this opportunity
for us to really understand
1163.514 -> and prioritize the development of
1165.65 -> a virtual agent assistant tool.
1168.08 -> You can no longer just tap
your body on the shoulder and
1170.63 -> say, hey, have you seen this before?
1172.1 -> Hey, which article do I need?
1173.75 -> What is the form that I have to fill out?
1175.19 -> How do I press that button?
1176.45 -> You can't hear the heated
voice and tone of the customer
1181.7 -> and raise your hand so that your team lead
1183.62 -> can come and help you out.
1185.48 -> So, we decided to build a virtual agent.
1192.08 -> So what we wanted to be able
to do is make sure the agent
1194.674 -> had the intent at their
fingertips as fast as possible.
1198.47 -> We wanted to make sure that
while we were doing that we
1201.56 -> would automatically surface
the article so they didn't have
1204.35 -> to search more than once.
1206.18 -> We wanted to make sure
that we had next action,
1208.85 -> so the right next service
action to avoid the second
1211.43 -> callback and wanted to make
sure that at the end of the day
1215.06 -> customer NPS went up.
1217.28 -> Now it's also great if the
employee experience goes up.
1222.14 -> So we created this tool and
1225.44 -> it side seats our virtual agents.
1228.26 -> So now when you sit there you can see
1231.02 -> the transcript go on live.
1233.3 -> We are passing in information
from the IVR so you know what
1235.73 -> the customer thinks their intent is.
1238.04 -> We have the ability to pop up
other things and it cues off
1241.25 -> of both our agents and
our customer's voices.
1243.53 -> So you can hear Chase speak
and different entities
1247.34 -> and ideas will pop.
1248.66 -> In the associated knowledge
article will pop and populate
1252.29 -> on the right hand side.
1254.03 -> We are measuring success and we have
1255.71 -> a ton of experimentation
going on right now.
1257.63 -> We're learning into what's
valuable to the customer,
1260.15 -> what's valuable to the
employee experience as a whole.
1262.88 -> So you'll see some of the
metrics for tracking at the
1264.74 -> bottom, whether it's
first call resolution,
1267.89 -> average handle time,
reduction in transfers.
1270.17 -> What are the things that we can
do to improve the end to end
1274.015 -> experience for both sides of this coin,
1276.8 -> the agent and the employee.
1280.82 -> So this is what we came up with.
1283.28 -> If you take a look,
1285.83 -> we've got, our audio is streaming in.
1290.3 -> It was, to be fair, the
first time we did this
1293.93 -> we did not use managed services.
1295.82 -> We built a ton of homegrown
manual solutions that took a lot
1300.2 -> of care and feeding and
optimization and it took us a while
1303.5 -> to get it to a point where we
were comfortable exposing this
1306.634 -> to some of our agents.
1309.2 -> We learned a ton along the
way and that's really what I'm
1312.68 -> hoping you guys get outta this talk.
1314.81 -> So, yes,
1316.73 -> we were able to have customer
audio streamed in real time.
1320.48 -> It was transcribed using our
existing incumbent engine that
1323.45 -> we used to store a lot of our recordings.
1326.3 -> We had the ability to, after
the call was transcribed,
1331.07 -> take it through some homegrown NLP models,
1333.5 -> extract intense, entities,
1335.33 -> pop them up on the screen and
then use that to queue off an
1338.36 -> elastic search query in the
back end through our knowledge
1340.88 -> management articles.
1342.83 -> And in doing this it created
a lot of churn and in the
1347.42 -> experience itself as new
entities were popping as we were
1350.018 -> refining some of these things,
1352.04 -> the knowledge pain would change
in the experience although
1356 -> not optimal was progress.
1358.67 -> So I am a big fan of progress
over perfection and you'll see
1361.85 -> that in some of the iterative
development approaches that we
1364.04 -> take here at Chase.
1366.74 -> The last piece is making sure
that we had the ability to not
1370.22 -> only pop the intent but if
that knowledge article had some
1372.98 -> guided content,
1374.12 -> then we could show them the
scripting and do some coaching
1376.16 -> in real time.
1379.31 -> What we found is that transcription
and the accuracy of it
1383.78 -> was really important.
1385.34 -> This was the foundation
that we were building
1387.11 -> a lot of our intent
models, our entity models,
1389.63 -> our knowledge base,
1392.24 -> we were looking at call
reason and volumes to even
1395.21 -> prioritize the things that
we would want to be able to
1397.58 -> execute, contain and self-service
1399.92 -> and we need to be able to get it right.
1402.29 -> So, we cut over and decided to
start using Amazon transcribe
1408.135 -> and we went live this summer.
1411.02 -> What we saw was an approximately,
1415.91 -> depending on line of business,
1417.29 -> 12% reduction in word error rate.
1420.74 -> That is important for two reasons.
1422.96 -> One, no frontline agent or specialist
1426.77 -> is going to trust a tool that confuses
1429.17 -> the words chase and cheese.
1432.68 -> So making sure that these things
are relevant and they could
1436.94 -> actually see and feel the
improvement of the engine as it
1439.37 -> goes creates a level of
confidence in the tool.
1442.25 -> If it understands and can
see these things and it stop,
1444.8 -> it stops getting simple things wrong,
1447.47 -> I'm gonna have a better,
1448.493 -> better faith that it's gonna
give me the right advice.
1452.15 -> Trusting a tool to tell you
what to do takes some time.
1456.26 -> The other thing we learned is that I,
1459.367 -> I said in the beginning we are very
1461.27 -> security focused here at Chase.
1462.767 -> Not only do we have regulatory
compliance things that we
1465.32 -> wanna make sure we're we're handling,
1467.18 -> but personally identifying
identifiable information is key
1471.41 -> and I don't wanna be the
person on the front page of the
1473.21 -> Wall Street Journal having
messed something like that up.
1475.91 -> So we had a very manual three
step process to extract any
1480.47 -> PII from a transcript before
we could store it or use it or
1483.59 -> push it to downstream systems.
1486.17 -> It came out of the box and we
were able to test it and hit
1490.76 -> our accuracy requirements that
we have internal to the bank.
1496.25 -> The other thing that I, I as a,
1498.56 -> an engineering lead loved
was that because this was a
1503.09 -> managed service,
1504.59 -> although we did work through
optimization but I didn't have
1507.86 -> to have the same level
of manual data wrangling,
1511.25 -> we didn't have the same kind
of level of effort to try and
1514.22 -> pull all of these things
together, with our pipelines,
1516.71 -> our constant maintenance and support,
1519.77 -> I was able to save three FTE.
1521.851 -> Now that doesn't sound like
a ton but that does help me
1525.5 -> advance my UI.
1526.64 -> The other features I've gotta do and
1528.5 -> work through and not just,
1530.96 -> I was gonna use the word
babysit but really focus on just
1534.65 -> one piece of the puzzle.
1536.39 -> So having a solid foundation
that our agents believe in,
1539.36 -> that we see improving, that we
have the ability to optimize,
1543.53 -> really started to unlock
our capability here.
1546.41 -> So, I talked before that we
wanted to be able to use these
1550.58 -> models for inclusion or
1554.57 -> the the output of the transcription
1556.88 -> capability to go into modeling.
1559.76 -> But let's talk a little
bit more of the why.
1562.55 -> Call summary ,the way we use it today,
1565.04 -> is we have an incumbent tool,
1567.83 -> takes all of our transcriptions and
1570.71 -> categorizes everything
with a main call reason.
1574.07 -> We also have some agent tools
depending on the line of
1576.74 -> business where they select
the main reason and if it was
1579.29 -> handled or not.
1580.94 -> 40% of those calls are
labeled general inquiry.
1586.52 -> That's not insightful.
1588.11 -> That doesn't help me figure
out what I have to fix in a
1590.87 -> digital workflow.
1592.04 -> It doesn't help me figure out
how I improve our knowledge
1595.1 -> management structure.
1596.54 -> If it's about a payment, what
was wrong with the payment?
1598.79 -> If it's about a transaction,
what do you need to know?
1601.73 -> How do I improve the customer
customer experience when all I
1604.28 -> know is that 40% of my 33
million interactions are labeled
1608.93 -> general inquiry and
that's 33 million a month.
1613.79 -> So speed to insights,
that's what I'm after.
1618.71 -> How can I change and improve
the customer experience?
1621.59 -> How do I improve the support
that I'm giving my employees?
1624.74 -> That's one of our keys.
1626.93 -> Now, call summary is also an
interesting beast because if I
1630.11 -> can figure it out from a
call reason I can get more
1632.222 -> predictive, I can catch them and chat,
1636.29 -> I can email them before it
happens and I can learn more
1641.06 -> about next call avoidance.
1643.79 -> How do I get to a suite of
capabilities that says while I
1646.28 -> have you, why don't we walk
through and set up autopay?
1650.69 -> Why don't I help you set up
some alerts so you'll get a text
1653.66 -> message the next time your bills do.
1655.58 -> Why don't I, you know, show
you how to use the app.
1658.64 -> Let's do this together.
1660.38 -> Every interaction with a customer
at Chase is a opportunity
1665.54 -> to deepen the relationship that we have.
1668.69 -> We strive to be the bank for all.
1670.94 -> I can't be the bank for all if
I don't understand the needs
1673.1 -> and goals and help you
deliver it as a partner.
1675.83 -> That's what we're after here and
1677.06 -> that's what drives us every day.
1679.22 -> So, I talk all the time about data
1683.09 -> and needing data for modeling.
1685.22 -> These are the kinds of things
that I can get through,
1687.95 -> in justice features into our model
1689.75 -> and change the way we work.
1692.12 -> And lastly, that more seamless
1695.6 -> omnichannel customer experience.
1697.85 -> Has anyone ever been talking
to a bot and then really needed
1700.43 -> a person and then ended up
having them call and then you get
1704.09 -> put into the IVR and then you
have to authenticate 700 times
1707.51 -> and then the person who
finally answers the phone says,
1710.51 -> what are you calling about
after you already told them on
1712.73 -> the chat and then told them
in the IVR and then told the
1715.01 -> person a third time and
then they say, hang on,
1717.11 -> I can't help you, let
me transfer the call.
1719.93 -> I never want that to
be my chase experience.
1722.81 -> So, that's what we're after.
1725.06 -> How do I seamlessly pick up a
workflow and help you finish
1728.75 -> that transaction?
1731.69 -> So let's talk about knowledge
management a little bit.
1735.08 -> This is exciting for us because our
1739.76 -> knowledge management answers repository.
1742.22 -> I don't know how many
other people feel this way,
1744.77 -> but my opinion is it reads kind
of dry like an encyclopedia.
1749.93 -> So when you're trying to get
the right answers and you're
1752.12 -> sifting through all of these things,
1754.16 -> it's a lot easier to try to
use a tool that allows you to
1757.82 -> get to sentence level accuracy
of what is the answer.
1762.35 -> If I just wanna know what
is the APR on that card,
1766.04 -> I can get there using a tool like this.
1769.19 -> So, the curation of doing the
1773.72 -> elastic search took us a while,
1776.69 -> making sure that we had the
ability to kind of get through
1779.09 -> and have relevant answers.
1781.76 -> Was a very manual task.
1784.4 -> We also needed to have the
right set of training data and a
1789.458 -> representative sample of all
those intents that like aren't
1794.21 -> frequently asked questions,
how do I answer those?
1796.61 -> Because that's what the
agents need help on,
1798.38 -> not the calls that they
see every single day.
1800.09 -> They know how to do that,
1801.17 -> they can process a payment
with their eyes closed,
1803.21 -> but what happens in those niche
cases and how do I get them
1806.18 -> to the right answer as fast as possible?
1808.76 -> So we did a proof of concept with Kendra
1811.88 -> and the experience was great.
1815.751 -> I gave the team and and this is important,
1818.15 -> I gave my team a week,
1820.16 -> you have one week to hack this
together and let's just see
1822.38 -> what it can do without us
over-engineering and trying to
1826.67 -> make sure is absolutely perfect.
1828.71 -> Let's just see how good it can be
1830.69 -> and what comes out of the box.
1833.18 -> So in that week, they
were able to plug it in,
1838.04 -> get the piping all done.
1840.619 -> But this is what I saw
that made me excited.
1843.56 -> I had the ability to out of the box,
1846.14 -> incorporate agent feedback.
1848.09 -> We have that capability
in our existing tool,
1851.24 -> but it's at the overall experience level.
1853.07 -> Like every call, did this
tool help you, yes or no.
1856.37 -> This allows me to get just
about the knowledge article,
1859.46 -> is it relevant or not?
1860.93 -> And how do I make sure that
I listen and learn over the
1864.05 -> course of time?
1866.81 -> It also had an out of the
box filtering capability.
1869.6 -> So a customer might have more
than one account with us.
1873.17 -> How do I make sure I
can toggle between them?
1875.39 -> I don't just make an assumption,
they have one question.
1878.03 -> They very often say while
I have you on the phone,
1881.6 -> you need to be able to toggle.
1883.43 -> That capability came outta the box.
1885.68 -> We also had the ability to
add FAQs from other sources of
1890.15 -> you know, any kind of alert system.
1892.79 -> So instead of them going through an email,
1894.83 -> we could create something that said, hey,
1897.56 -> there was an event that happened
yesterday and there's gonna
1899.87 -> be some people who were affected by it.
1901.64 -> Here's what happened, here's
the people who are affected.
1905.57 -> And allow that to just be a
part of the the the initial
1908.764 -> experience.
1911.27 -> And like I said before,
1912.59 -> I gave them one week to get
this going and the accuracy was
1916.13 -> really comparable and the team
was really excited about the
1918.95 -> out the box capability but
also the ability to optimize,
1922.46 -> you know, what if we
enhance our intent modeling?
1925.1 -> What if we combine this and
Kendra, what if we, you know,
1928.61 -> how could we take Lex
comprehend and Kendra and make
1932.39 -> something really impressive?
1934.52 -> So more to come on this.
1937.82 -> And then lastly,
1939.26 -> any engineering team knows
that your end user does not use
1944.75 -> your product the way
you initially thought.
1948.02 -> So one of the things we're
able to do is create champions.
1952.04 -> We have a model office
now we sit side by side,
1955.28 -> we listen to their calls,
1956.45 -> we watch the tools that they click on,
1958.04 -> we see their struggles through
knowledge management searches
1961.61 -> and what's gonna be relevant or not.
1963.65 -> But more importantly they've
taught us about the business.
1966.56 -> So I sit down and have
them grade my papers.
1969.56 -> How are the intent models doing?
1972.41 -> What level of granularity are we getting?
1974.42 -> We can go through the
ontology and the parent child
1976.52 -> relationships and the entities
that we have built out and
1979.94 -> they will tell me,
1981.65 -> you are great up to here but
there's a slight nuance between
1986.81 -> a stolen card and a loss card
or the way we execute this
1989.93 -> versus this or why they can't
get a new card right away
1993.5 -> because they changed their
address or they did whatever.
1996.02 -> And when you go through the
conversations with them as an
2000.61 -> engineer and as a data scientist,
2002.32 -> you can start to see the nuances
and the things you need to
2004.69 -> pick up on to be able to optimize it.
2008.14 -> The other benefit you get is
that these model office users
2011.02 -> become advocates and champions.
2012.76 -> They're part of your development process.
2014.77 -> So now, now not only are they
thinking with you and saying,
2019.84 -> Hey, I want new features that do this.
2022.54 -> They're saying to their friends,
2024.58 -> I'm a part of this project and
it's actually helping me in
2028.51 -> my day to day job. So let me
teach you how to use this tool,
2031.63 -> which is half the battle in
creating any new employee tools.
2035.02 -> Behavior change is the biggest
hurdle that we have to do to
2038.98 -> adopt some of the capabilities in AI mode.
2043.18 -> So, this is where we are today.
2048.25 -> We still have the audio
streaming in real time,
2050.8 -> but now we're using transcribe.
2052.51 -> We've had the ability
to paralyze some of the,
2055.66 -> the audio stream to not only
hit transcribe but to be able
2059.2 -> to get those NLP models straight
away and then have Kendra
2063.981 -> kickoff based on the entities
and intents that it's
2066.97 -> extracting and retrieve
knowledge articles for us.
2072.25 -> All right, with that I'm
gonna hand it back to Vafa.
2075.357 -> - Thanks Ami.
2077.2 -> So this is the architecture
that at the moment EV program is
2081.31 -> running and so the calls
are coming from the PSDN,
2087.34 -> it goes to the, obviously
I only put two boxes here,
2090.94 -> but you can imagine how complex
is contact center of chase.
2094.6 -> It goes through a number of
fire firewalls and number of
2099.55 -> services,
2100.383 -> homegrown services as well as
other third party solutions.
2104.02 -> And then we have got a
server that we call it secure
2107.56 -> conversational gateway.
2109.33 -> What it does actually got
two options when you want to
2112.6 -> stream the idea more than two options.
2114.79 -> But one of the options
that we, we chose for EV,
2117.61 -> which I'm not going through
that why because we had lots of
2120.46 -> experiment there was to use
an open source protocol called
2123.67 -> GPRC that at the moment we are
2126.572 -> transferring the voice to AWS.
2130.03 -> Another team actually at Chase
that after that we've done
2132.97 -> that.
2133.803 -> They are using Chime CK and using Kineses
2137.86 -> to transfer the voice there.
2139.33 -> And we are trying to compare
these two projects at
2141.43 -> Chase to see that which one
has got a better performance
2146.128 -> because as you know the
performance is the key.
2148.54 -> You want to have this in the real time,
2150.1 -> you want to stream the
audio in the real time,
2152.02 -> you want to get that answer back.
2153.67 -> You cannot if five second gone,
2155.65 -> then the moment of
helping the agent is gone.
2159.13 -> So the key every millisecond
for us was matter when we were
2162.73 -> designing this.
2164.26 -> And then it goes through
the AWS direct connect
2166.24 -> obviously goes through
the transit gateway.
2167.98 -> So the latency of the network
was for us was the issue
2173.29 -> that we were doing lots
of low testing there
2175.33 -> and the stress testing there.
2176.83 -> And then you got two (indistinct).
2178.66 -> So the first one that the
audio is getting stream to is
2182.08 -> responsible to call the transcribe,
2185.08 -> get the transcription and be
using MSK because it needs to
2189.1 -> send it for lots of auditing
purposes and other places
2195.07 -> that needs to get stored goes there,
2197.08 -> you call the transcribe, you get back,
2198.58 -> you got the intent and
everything that you already done
2200.74 -> machine learning before that
was happening on prem non
2204.244 -> SageMaker that you know,
okay, what are the intents?
2206.89 -> Oh customer says I've lost my credit card.
2209.26 -> And then you'll find the answers
through that Dynamic DB or
2213.374 -> Kendra,
2214.207 -> which is not at the moment
because we are going live with
2216.49 -> the Kendra at the moment,
2217.93 -> the POC is done.
2221.02 -> And then there is a UI on
2224.47 -> EKS as well that the agent
can real time while they're
2227.56 -> talking to the customer.
2229.03 -> The things pops up here
to help the customer.
2232.99 -> I said I lost my credit card.
2234.22 -> These are the three links
that will help you with with
2237.185 -> customer.
2239.17 -> So that's the chase.
2240.07 -> But obviously was lots of
challenges and for us the main
2243.43 -> challenge was to do the low
testing for the number of calls,
2247.15 -> about 33 million calls,
10 thousands of agents.
2251.08 -> So we needed to make sure that
all of this infrastructure
2254.32 -> can scale up and down
wherever they needed.
2257.38 -> When you got a thousand
agent or 10,000 or 15,000,
2261.16 -> this, this should,
2262.03 -> this should do exactly the same thing.
2264.79 -> So most of our time was
spent on low testing.
2268.27 -> So maybe this was complicated
at the beginning when,
2271.63 -> when we started and and doing it.
2273.43 -> But going back to CCI solutions
that I was talking through
2277 -> my talk, these are the things
that we already built for you.
2280.51 -> Everybody in this room you can
just go do a couple of click
2284.35 -> and you'll have this up and
running in your account and
2287.17 -> I'm going to show you how,
2291.43 -> and there are two things
we always say at Amazon,
2295.57 -> we always say that
everything in IT goes wrong
2300.4 -> and very vice person told me
that don't do a live demo,
2303.85 -> but guess I'm gonna go do a live demo.
2306.91 -> So let's see.
2310.51 -> Right.
2311.343 -> Let me jump on my AWS account
or one of the AWS accounts.
2319.69 -> Right, you can just go on Google
2321.367 -> and just search for live colon analytics.
2323.92 -> You'll see this demo or
we call it LCA or you can,
2327.611 -> I think the first item
that comes on top is this,
2330.183 -> this blog that you have
got for live colonized.
2333.07 -> That's one of the solutions
that we built for CCI.
2335.98 -> Goes through all the configuration.
2338.14 -> You can actually watch this,
2339.253 -> this really interesting
presentation that we've done in the
2341.65 -> past. It goes through all the setups,
2344.2 -> everything that we've got and
it gives you cloud formation
2345.687 -> there so it can actually
run this cloud formation.
2349.72 -> And as I said, few clicks,
call it LCA reinvent
2355.9 -> if I can spell 22.
2357.85 -> And these are all the options
that gives you based on your
2361.45 -> need. For example, you can have a demo or,
2364.481 -> or you're using choice,
2366.31 -> you would like to use Chime
voice connector or you got
2368.542 -> Genesis Cloud for example or
Amazon Connect, whatever it is.
2373.36 -> So obviously this is a starter kit.
2375.94 -> This solution is going to give
you a starter with just few
2378.4 -> clicks.
2379.233 -> So you'll have your live agent
run live agent as this up and
2382.54 -> running and then you start
customizing it for your business
2386.29 -> need. I'm not gonna do any of that.
2387.67 -> So it goes through Kendra,
2389.8 -> do you wanna do call recording
on a three how you want the
2392.606 -> transcription configuration,
2394.63 -> all of your transcript
configuration is here, which,
2397 -> what language do you want
to take the PII out or not?
2400.06 -> So on that blog that I've mentioned,
2402.7 -> it's going to explain all of
this configuration to you and
2406.39 -> all I need to do say that yeah,
2408.07 -> I'm happy with everything and
then create this stack for me.
2412.18 -> Wait for 10 minute,
you've got a live agency,
2414.79 -> I'm not gonna keep you waiting,
I've created it before.
2418.03 -> Let's jump on,
2425.62 -> another account.
2428.71 -> So that cloud formation that
I've built and if you look at
2434.8 -> the output of what it gives you,
2438.58 -> they actually give you an UI.
2440.32 -> So let me just go back
to that architecture
2442.48 -> for a few seconds.
2446.83 -> To explain a bit more what's happening.
2453.4 -> So basically what I've created
here is going to put a demo
2456.94 -> for me on AC two, which
I'm gonna call it now.
2459.91 -> Just basically a sip trunk and
we are using chime connector,
2464.2 -> stream it.
2467.26 -> We are going to stream that
call over here we are using
2470.92 -> Lambda, we called the transcribe.
2473.11 -> We actually created one which
I think went live yesterday or
2477.55 -> day before. We created
one with TCA as well.
2480.34 -> So this is using
transcribe and comprehend.
2482.17 -> We created one with TCA
as well. So you can,
2484.21 -> you can search for that and
and you can use it and we put
2488.83 -> stuff in the three bucket
obviously again another can stream
2491.89 -> that you do your comprehend
to find out what's happening
2495.31 -> with the call and we obviously
it's a very simple UI.
2498.97 -> Then later on can replace that,
2500.77 -> that API that it creates your UI,
2502.84 -> very simple UI that is on the
three bucket and we are using
2505.78 -> app sync to update that UI for you
2508.24 -> and the agent can can see that.
2511.09 -> And as you can see here,
2514.06 -> things that this cloud
formation has created for me is
2516.58 -> already created this UI.
2524.057 -> Obviously you can create this
UI the way that you like and
2529.63 -> Let's see how it works.
2531.16 -> If it works.
2535.03 -> Let me check that it already give you
2536.543 -> by the way that demo it already create
2538.75 -> a phone number for you.
2539.65 -> Make sure I'm not calling
my manager the demo.
2553.27 -> - [Automated Voice] Thank you
for calling mechanic Michelle.
2554.71 -> Seattle's only all female on
(indistinct) this is Maria,
2557.92 -> how can I help you today?
2561.57 -> - [Vafa] You can see the call there.
2563.02 -> We already got a call in real time.
2565.6 -> I'm talking obviously that's
the demo. Nobody's there.
2567.46 -> Don't worry. So, and I can
call to my phone and say that,
2570.67 -> hello, let's go to this
call, see what's happening.
2576.07 -> I'm very happy with your service.
2578.361 -> I'm grateful.
2580 -> And you'll see that live is
happening there and you can see
2583.96 -> it's kind of happy, neutral.
2585.97 -> And then maybe we are
talking in the middle.
2589.18 -> I'll wait for the demo to say something.
2601.33 -> Maybe less straight.
2602.163 -> Say what? I'm not happy with your service.
2604.69 -> I need to talk to your manager.
2609.04 -> You went red, you'll see that.
2611.002 -> - [Automated Voice] Yeah, sure.
2612.88 -> So just to confirm,
2613.96 -> would you like me to try to
troubleshoot over the phone with
2616.27 -> you or would you like to
bring it in and have one of my
2618.74 -> tenants check out and sit in?
2619.816 -> - [Vafa] No, I just don't like
anything about your company.
2622.75 -> I wanna cancel my subscription
and I, I wanna go.
2627.004 -> - [Automated Voice] Okay
great, are you with your card?
2630.46 -> - [Vafa] So you'll see it's already
2632.02 -> actually picked up an issue,
2634.72 -> an issue detected there. We've
got the issue detected there.
2637.687 -> You can already alarm it,
2639.19 -> you can already send a
supervisor to go and,
2643.03 -> and you can see all the unlinked.
2644.74 -> Let me just disconnect the call
2649.48 -> As soon as I disconnect the call.
2650.68 -> You got the full transcription
there as well for you.
2652.597 -> You can see your call,
2653.68 -> you can see what happened during the call.
2656.05 -> Obviously this is a demo,
so it was only 15 second,
2659.23 -> but it can go longer and,
2662.98 -> and you got all the insight
of the call that you can see
2665.74 -> exactly what happened,
what time it went good,
2667.6 -> what time it didn't.
2668.71 -> You got all the transcription
that you wanted on the call.
2671.5 -> If you remember we've talked about
2672.73 -> the call categories in TCA,
2674.379 -> you can have those call issues
detected and do something
2677.89 -> about it in real time.
2679.87 -> A very quick demo, but it
just create two clicks.
2683.53 -> Everybody in this room can
do it. Please go try it.
2686.89 -> And we have just launched
one with TCA as well,
2689.11 -> not with transcribe.
2691.93 -> Let's go back to some of the
resources that I'm going to
2697.66 -> share with you.
2698.59 -> So we just went through this
how we can help obviously there
2703.24 -> are many workshops,
we've got the CCI team,
2705.13 -> we've got the business development team.
2706.84 -> We are more than happy to
reach out to your solution
2708.79 -> architect that they're happy with you.
2711.34 -> You can go to that website,
2713.11 -> you see all of our solutions
on CCI that we can help.
2717.16 -> There are many sessions,
2718.72 -> although some of them are yesterday,
2721.15 -> but you got today more
sessions on tomorrow,
2723.46 -> we got more sessions as well.
2725.92 -> And these resources like to
take a picture. They are.,
2729.213 -> or if you just search CCI
or live call analysis,
2733.72 -> you'll find them on the Google
as well of all the blogs that
2737.59 -> we've written and all the
solution that we've created that
2740.56 -> you can use it as a starter
kit and then you can evolve it
2744.28 -> for your customer need.
2746.23 -> The way that we have done for
Chase and thank you very much,
2751.63 -> we're gonna go to our QA.
Source: https://www.youtube.com/watch?v=r9N8-OHoiSI