Conversational AI in Banking

This broadcast was unique in two ways:

1. It's the first time a bank takes the stage.

2. It's the first time two companies join the show together!

Ramtin Matin works as Lead Technological Strategist at SR-bank, the 4th largest lending bank in Norway and part of Sparebank1 Group which is Norway’s 2nd largest financial grouping.

Henry Vaage Iversen is the CCO & Co-Founder of Boost.ai, a Norwegian conversational AI platform, agency and educational institution in one.

They are going to tell us how the conversational adventure started for SpareBank and how Boost.ai got involved in taking things to the next level.

We will be talking about:

- Conversational design

- Natural Language Processing

- Managing & optimizing large amounts of intents

- Labelling data

- Strategy / Digital transformation / Business outcomes

- You name it!

Unchecked transcript

Hi Botrepreneurs, thanks for watching again to a Facebook broadcast.

Yeah. Today it’s a special one because we have guests from two companies. Ramptin Matin,

works as lead technological strategist at SR-Bank.

Hi Ramptin, thanks for joining!

Thank you for inviting me.

Yeah. Sure. So you guys are the fourth largest lending Bank in Norway as part of the Sparebank1 Group

which is in total the second largest Financial Group.

And on the other camera,

we have Henry also from Norway.

So yeah, Henry Vaage,

I've done is to see CCO and co-founder of boost Ai

and Norwegian conversational platform and basically an educational institution in one.

Yeah. Awesome. So we're going to talk about the journey

you guys went went through together basically and talk about conversational AI within the Banking,

see ya around things.

Raja was wondering when the has our bank first realized that conversation was the way to go,

especially for customer support.

And what was the first use case you guys defined and how did that go?

Well, we first, we first figure out that chat or conversational way.

I was the way to go when we devised our chat,

first strategy, and the strategy basically focused on accessibility and speed for the customer.

In order to get there, get there problem solved

and I believe it was the late 2016 that we chose to go into that direction.

So we looked at the problem from two different sides.

We had the internal view in order of being the most effective as we could,

but also from a, from a customer perspective,

trying to to solve their problems as effectively as possible.

Yeah, so instead of letting them shirt

for many minutes on the website and Fa Q's offer them anymore.

Yeah, more Dynamic solution.

Exactly. Yeah read it.

So yeah Henry when and how did boost a I got involved in this conversation or Adventure?

So it's actually a really interesting story

because we started company for like three and a half years ago in 2016,

and we were Given basically a challenge by us a bank.

They want to have the best conversation, AI driven virtual agent on the market.

There's always some Benchmark in the nordics.

We had a couple of banks doing some projects.

Of course. That was The Benchmark for us. So when we started the company,

we didn't have any of the algorithms we have today to be able to understand intense and so on.

So we actually built everything in, I think is two months

and then we actually Hit The Benchmark in the Garrett.

And after three months, we actually suggesting that the lat.

So that means that like four or five months later.

We actually got the first contract and that was with SI bike and we went live with,

I think it was 500 intense in January in 2017.

And from there, we have basically continue building the module we have in banking,

for example, so we start with violent ends.

Now, now that accounts for around two now.

Tell us the intense. So that everything started with s a bank and they're neat.

And of course, as you work with banks, you learn how they structure information.

You understand how you can work with them in order.

Of course, also takes it to the next step, which is authentication,

and also being having the origination being able to handle actions on behalf of the customer.

That's something. Also as a punk is doing today.

Cool, cool. So, you you build templates,

you say that can be placed a starting point

for several of your customers and then they can ya be customized on that.

It's great. Yeah. So typical in a model you have three layers of knowledge.

So one is of course, the generalized knowledge. So being able to understand like hey,

thank you and kind of those things.

The next layer is is the industry specific.

So in this case, we'll be back. In, so we understand the banking firm.

So it's always on specific banking terms which are not usually covered by the the big Tech vendor.

So we specialize in understanding that in the

Nordic but also in the European context and the third layer is the company specific.

So typical, the bank will have their own offering and on their own unique products,

which we need also to understand. So kind of there's three layers,

you need to be able to understand and we have out of the box,

the to talk players and then of course we need to add the company.

Okay. Cuckoo Sorrento and we spoke before briefly

and you said that your virtual assistant is now the only entry point for customers to.

Yeah, to contact the bank. That's quite quite radical.

I have to say, my most companies choose to offer the,

the other channels also out there.

I was just wondering that the clients complain a lot.

Maybe, especially in the beginning and What percentage is percentage of conversations,

do you now see being handed over by the book human agents?

Just a quick elaboration on that.

We have chosen the chat Channel as the preferred way to communicate with with our customers.

And as per that decision that we have removed,

all the phone numbers and stuff like that from from our website.

Yes, but we have not shut down the the phone line so to speak.

So people are still able to contact us via VIA phone,

but when it comes to to the chat,

we try to push that channel 4 because we see that it is an effective way.

It is an effective means of solving the customer client problem,

and we actually also see that Twenty-five percent of our clients

or customers actually escalate the the conversation to,

to a human agent. And 75% actually choose to keep the conversation within the verb conversational AI

or virtual assistant,

which is a quite a high number

and we actually use that as a as a guiding kpi in order to measure our own.

Quality and to see how relevant we are enable to,

to actually solve the the customer problem building on that theory

that if we were not able to solve the problem,

customer problem a larger amount of people would actually choose to

to escalate the conversation to the human agent.

Yeah, exactly. And I was wondering already a huge amount of

intense was mentioned that this is usually needed to

advance or other big corporates.

What what are the problems intelligence you run into their teaching structure them

and used entities in combination,

of course. So we are actually currently in here in a project where we are integrating our CRM system

and that is in order to be able to give our customers who actually use this service,

a more personalized dialogue.

We have the capabilities to have a very personalized dialogue based on

what type of products they have,

what type of Interest they have shown towards products,

but also from their To situation understanding

what their net worth is and actually be able to guide them

with Good Financial advice in that regard.

So we are always looking into integrating more and more systems into the conversation Alejo.

Aye, in order to be as relevant

as possible because in the end people would like to have specific advice to their situation.

And not just some type of General response,

which Which might not be that relevant and I can also add something on the complexity of the model.

Of course, as you add task of intense,

you had the complexity and we have structured intent in a hierarchy.

So the meaning that we have a really nice visual over Google on the intense,

but also make make the use

or the need of utterances at the low point

because we can actually since it's divided in Herrick and In the hierarchy,

we can actually being able to add more and more specific intent as we go forward.

So typically we start with the root intent which could be,

for example, Insurance. Next could be car insurance.

And then we talk about price car insurance,

and so on to go deeper and deeper in the intent hierarchy.

So typical what we need for one.

For each of the intense is around 10 after answers or training data samples.

Let's face it. What? the the model learns from Who

go and so depending on where the both ends in the hierarchy.

It's the follow-up questions are also,

of course, triggered by the steps below.

Like, what would you like to know about your car insurance?

You, I'm interested in pricing if it wasn't clear.

Yeah, but that's the best part of the intent classification.

So when you classified intent, you can end up in different part of the path.

Basically, you go deeper and deeper but of course,

when you want to do and text Semester rump and said connect with the back and of course,

you need to be able to extract entities.

And of course, if you're missing some of the entities you

need to be able to handle the conversation.

So you can be able to extract them from the customer is not always there

giving all the information you need the first time

when sometimes you need the extra extract that in the later,

part of the conversation,

so the pain till it be done on basically if it's connected to the backend or not level,

of course at the terminal left, on the, on the structure of them,

of the end of the conversation. Yeah,

and you also keep in mind,

previous conversations customer already helps with with God,

that is definitely something that we are aspiring to do.

We obviously do store the conversations and we try,

and we follow the GDP, our rules, of course,

but we would like to use this information actively within the conversation.

So so we are also looking So saying,

alright, the last time you spoke with us we were speaking about did you actually get this resolved

and to make sure that the customer actually understands that there's a collective memory across

the different channels that we that we were working with in the bank.

So regardless of you contain contact your house beard,

human agent, or the chat,

or the email that you actually have a collective memory across these channels.

That is something that That we're trying to picture that we trying to obtain,

but of course, there's some there's some issues in terms of some

Legacy systems that we need to to take into account,

but the conversational AI are fully capable but actually handling this.

So we see that there's a big strength in doing that.

Okay, cool. Cool. Yes.

He quite some viewers at the moment, please.

Just let us know if you have any questions then we're going to try to tackle them right away

and goes.

That's the fun part about doing Facebook live interview,

of course. And yeah,

how does within boost AI labeling and annotation work?

What tools? Do you provide there? So meaning regarding is to to label it rinses.

Yeah. Yeah. So sure, so typical in the solution.

Of course, there's so much pre-trained that you don't need the label that match data,

but we also have some built-in tools where you can actually generate a lot of the expenses you need.

So for example, and what we see is that

for some of the products it's you have almost the same question attached to it.

So for example for a bank you can have different insurance.

Says then, of course, price for insurance X terms,

for insurance X. Have the same structure is basically that you need to replace

some of the key words there in order to build a princess.

So that's actually part of that the platform that you can actually just add the keywords

and then you have add the training data.

And that's basically taking like two seconds adding that training data.

And then we have something which are working.

That's really convenient when you wanna start looking at,

adding a lot of Hence, so typical,

we have experienced a couple of clients wanting to add maybe 10,

20 new products. Then it's really nice that you can actually utilize

the training data between across these products.

Yeah, yeah. Yeah,

you guys have built a native

and will be engine natural language understanding also and also an intent trainer if I'm correct.

Yeah, so they are intent training.

We actually called AI trainer. So a typical day,

a trainer will be able to the trainer model.

So in our solution, we can build a model without having that technical background.

You don't need to be a data. This, you don't need that technical background.

So that's the reason. For example, for Esteban,

retrain, the resources on customer service center,

to be the role as AI trainer with

what we have experienced this that they have a lot of insight

and understanding and how to communicate with the customer,

and they also work daily with the customer.

So we definitely it makes a lot of sense to build the interface

for them so they can train their own virtual agent without being having that technical background.

We call it a a trainer and we actually had an ax in a convention.

Now, in Stavanger, for a couple of months ago,

where we have over 200, a, a trainer certified coming in

and think we have over 2,000 AI trainer certified in the platform.

So of course, they have their own community and they of course,

get help from the AI to build intense.

So they are going into more like an approver role.

They approve. So the AI will give suggestion,

here something you need to learn. Do you approve it?

Yes. Now, so that is that the key role for the Air Trainer.

In addition to add a good response to the question or the intent,

just a comment to what Henry is saying there.

If I may I think it's very important to stress the fact that

and although it is called an AI trainer.

There's not there's not needed to have technical competency to actually fulfill that role.

And I think it's worth mentioning that we need to to point that

out because I was Henry and mention

as well that the greatest effect that we have had was the resources.

We actually recruited from our Customer Center,

mainly because they know how to how our customers actually ask questions

but also because they have understanding of the quite complex product mix of a bank

or any other company.

You need to have a good understanding of what type of product makes you have.

What type of attributes. Those products has a 12 in order to set up an efficient dialog where

you actually want to answer the question of the of the client

as efficiently and as quickly as possible.

Of course, the agents know, the lingo also the best

as using might also differ from looking at the age of the customer,

but I got a question coming in for Henry about AI trainer.

Is it built as a custom interface or is it something that's integrated in your foot platform?

What do you mean by? So,

that's what they are trainer and the customers get access to is their own platform.

So that means that they get their own interface where they have all the tools.

They need to build a virtual agent.

So, of course, we need to certify them. That's also done online for the same platform.

So that means that they open up what we call an admin panel,

and then give you overview of the activities,

the suggestion from they are, but also the possibility to build the conversation as well.

Ben 10, and also that the conversations conversation flow,

So it's part of your studio,

your platform. Yeah, correct. That's great.

Yeah. And designing a happy part of a conversation is yeah.

Fairly easy. Relatively speaking.

Of course, there are always the edge cases that where labor goes for conversational design.

What does boost a I offer? To make it easier to tackle those unhappy Parts the college.

Yeah, we're conversations. Don't derail and it's a good question.

And of course, we see a lot of limitation

and a lot of deployments there because most of the chat box available on the market.

Like you have asked one question,

get response and then everything is forgotten.

We don't have contacts. You don't have to memorize school.

Go forward. And I think that's kind of,

its not a good customer. Experience, you definitely

want to even if it's a robot you wanted to have some context during the conversation.

So we have two really important features in that regard.

One thing is what we call a goal. So as you mentioned,

the happy path is a way to start the conversation.

And we also look at more like a process.

So if you want to order card, we want to order.

You want to get a loan, there was actually a path you need to take.

And as I said, there could be a happy path. With the goal,

we can set that when they have come to this point of the conversation.

It's a success. It's a goal. That means that during that conversation.

The customer is able to ask questions of off the topic

and then of course the guided back to the conversation you continually flow

with information you already had from the conversation

so you can basically as your human will do you start talking about something

and then you ask about the weather so small