Say Goodbye to Hold Times with Replicant's AI Customer Representatives

Behind the Bots podcast host Ryan Lazuka and Hunter Kallay interviewed Benjamin Gleitzman, CTO and co-founder of Replicant, an AI-powered customer service automation platform. Key highlights include:

- Replicant uses AI to automate repetitive customer service calls like password resets, freeing up human agents to handle more complex issues. 

- The AI discloses itself as artificial upfront and aims to resolve issues quickly and efficiently without chit chat. Calls are 30-40% shorter than with a human agent.

- Replicant leverages natural language processing and transformer models like BERT to understand callers based on intent rather than keywords. It can interpret complex requests like travel booking details.

- The AI assistant improves customer satisfaction while reducing costs. One client saved millions of dollars in reclaiming equipment by using Replicant outbound calls.  

- Replicant strives to use AI ethically by constraining language models and grounding them in accurate data. Hallucinations and inaccuracies are common pitfalls.

- Gleitzman sees contact center automation shifting from software charged per human agent to an AI-first model with humans for empathy and creativity.

REPLICANT

https://www.replicant.com/
https://try.replicant.ai/
https://twitter.com/replicant_ai
https://www.youtube.com/channel/UCg-GDAYCyLsTTam64PlyBPQ
[e-mail]

FRY-AI.COM

https://www.fry-ai.com/subscribe
https://twitter.com/lazukars
https://twitter.com/thefryai

Creators & Guests

Host

Ryan Lazuka

The lighthearted Artificial intelligence Journalist. Building the easiest to read AI Email Newsletter Daily Twitter Threads about AI

What is Behind The Bots?

Join us as we delve into the fascinating world of Artificial Intelligence (AI) by interviewing the brightest minds and exploring cutting-edge projects. From innovative ideas to groundbreaking individuals, we're here to uncover the latest developments and thought-provoking discussions in the AI space.

Yeah, thanks for having me. My name is Benjamin
Glitesman. I'm a CTO and co-founder here at

Replicant. So we're a contact center automation
platform, and we help with customer service. So no

more waiting two hours on hold to speak to a person.
My background is artificial intelligence and

machine learning, but I've also been an executive
sponsor on some of our largest deals. So I've seen

everything from the ideation of something you
want to create a call flow, you want to automate

with AI all the way through development and
deployment. But I got my start back in West

Virginia. I grew up in Morgantown and I was very
interested in robotics. So I thought that was

going to be the future of what I went into. And of
course, Carnegie Mellon is still a powerhouse

when it comes to robotics technology. And so that
got me a little bit interested in what will that AI

be like in the future? What will robotics look
like? This was me thinking back in 2002, 2003, what

will that look like? And I got a really fantastic
opportunity to go on an exchange program to Japan

when I was a kid. It's through the rotary program.
And so they trade kids all over the world. I went to

Japan, maybe a Japanese student goes to France,
someone from France goes to Guatemala and lived

with a couple host families. I didn't really speak
any Japanese, but I could tell already that they

were living in the future. It wasn't quite the
future that we've inhabited today, but even back

then 20 years ago, people were watching TV on their
cell phones in 2003. And that was really

fantastic. So I got a job there after I'd spent
about a year living with host families at a company

called Advanced Telecommunications Research.
And it's a little bit like the IBM of Japan. And it's

in this bamboo forest. It's really beautiful.
There's deer running all around and it's great.

But the president of that company had created an
Android humanoid version of himself. And when he

would leave and go on trips, he would leave this
robot copy of himself in his office. And it would

move and it would look around. And it was pretty
indistinguishable from a person. At least if you

were walking by the door, you couldn't quite tell
that it was him. But as soon as it opened its mouth,

you could tell that it was fake, because the voice
synthesis was just not there 20 years ago. And so at

that time, I thought, hmm, I wonder how long it'll
take until we're really good, the quality of the

voice, the accuracy, the latency on this kind of
voice technology. And so that was really my first

foray into thinking about what this AI might look
like in the future. Awesome. And that was a long

time ago. So in terms of technology, at least.
Yeah, we've seen a bunch of revisions. I mean, it's

kind of come a long way. I think when I had started
this company with my co-founders, we'd never had a

very good conversation with a machine. And this
was back in 2016, 2017. We said, why is this?

Everybody's got Siri in their pocket. They've got
Google Voice in their home. It's just not working.

I think there's some missing piece as to why we're
not having these great conversations with

machines. And that's really what we set out to do is
to craft great conversations and get this

partnership between the technology that's
necessary in order to make this work, but also the

human aspect. How do you get linguists and
conversational experts and really a whole host of

people around the table because they're
necessary to craft these great conversations?

Benjamin, I want to get into a little bit of the
technology. But before we do that, what's it look

like from the consumer side when somebody wants to
use your technology? Walk me through that

experience. How do I work it? Yeah, I think we've
all had the bad experience. We've all had that,

I've got something wrong. I want to change my
flight. I call in. I got to wait 30, 45. I think when

some of these flight outages were happening due to
COVID, you're getting three-hour hold time,

four-hour hold time to be able to speak with a
person. So we've all got those stories of the bad

experience. But in our earliest days, we did a lot
with food ordering. This was kind of like an early

breakthrough for us. We worked closely with
DoorDash. We worked closely with Postmates. When

you order from those restaurants, they don't
always have an iPad that's there in the

restaurant. They've got these partner
merchants. They're kind of calling in orders on

your behalf. And so you order your double
cheeseburger, but there's a small army of people

in contact centers that are phoning in those food
orders. And you can imagine, it's long days. It's

very repetitive. It's pretty easy to make
mistakes because you're just calling in food

orders all day. And so our kind of earliest
application of the thinking machine, that's what

we call the technology, the automation that we
create, was can we call the Italian restaurant

down the block and order with someone who's
perhaps got a strong accent. They're in a rush and

so you can't kind of cover up the latency with any of
those click-clack typing noise or doing anything

like that. And it's got to be accurate. You've got
to be calling in these orders correctly. And so

that's, I think, one of the earliest ways that
people might have experienced that was getting

these calls from machines. And as we talked to the
people who are on the other side who are taking

these calls in the restaurants, they said, we love
it. It always runs the call in the same way. There's

no chit chat. I can just get right to the point. I can
get in. I can get out and kind of get exactly what I

need. So that's an example of maybe an outbound
phone calling experience that it's not spamming.

I think that's the other key here is that we didn't
set out to robocall people. We didn't set out to

create more of these unwanted calls in the world.
We're looking to find places where there is an

inefficiency and create some resolution there.
So maybe for you and me, if your car breaks down and

you call many of the AAA auto clubs, if you're in
Canada and you call the CAA auto clubs, you'll be

getting replicant. And so what that means is the
call is answered immediately. There's no hold

time. Our calls are usually shorter. It's about
35, maybe 40% shorter than having to speak with a

person. And you can do a lot of fantastic things
with the technology. You might be, let's say I'm

driving from here. I live in Oregon today. Let's
say I'm driving to Washington. I might break down

on the highway. It's the middle of the night. I
don't even know if I'm in Oregon or in Washington.

And so being able to not only get that answer
immediately, know that I'm going to be safe, but

use some other functionality of the telephone.
Hey, can I send you a link? Open that up. Get the GPS

coordinates from the phone. Hey, can you snap a
picture of the tire? We can see if there's going to

be tire damage. So that's how maybe consumers
would interact with it. But we also take it a step

further. And so once we've realized, hey, it's
Benjamin. He's on the side of the road. He's going

to need a tow truck. We then use the thinking
machine to simultaneously call outbound all the

different tire shops. Hey, I got Benjamin. Can you
go pick him up? Oh, no, we're closed right now. Hey,

can you go Benjamin, pick him up? Yeah, we got some
time. We should be there in about 20 minutes. And so

you get kind of this almost a Rube Goldberg machine
of connecting the inbound phone call to a series of

outbound phone calls to then perhaps another
phone call back to the person who's on the side of

the road to let them know that the tow truck is on
their way or even let them know if there's going to

be a delay in sending someone out to them. Very
cool. Awesome. It sounds like you're in Canada. In

Canada, it's pretty big there right now. Yeah.
We're in North America, Canada. We also just

launched in the EU. And so we're working with some
customers who are in the UK. Just got our GDPR

certification. So that's a really great
testament for the technology and being able to

unlock a bunch of different areas where languages
really matter. I think Replicant works in

something like 35 different languages. And it's a
native usage in that language. So it's not as if

we're listening in French and then translating
that to English and inferring on it in English and

then translating it back into French. The machine
learning models that transformers work in that

native language. So basically, you're selling to
businesses, small business, large businesses,

any kind of business that would need this to help
with their call center because they don't have

humans to answer the phone. And if they do, like you
said, it might be three hours before someone can

answer a customer's call that needs help. Is that
sort of true right there? Yeah. It's for the

enterprise. And so we're looking for people
who've got these larger contact centers. But I

worked a contact center when I was in college. That
was how I helped pay for it. And I loved when I got

those interesting calls. When I got a weird
question, when I got something that requires

research that I really had to dig in and use my
empathy or use my creativity as a human to answer

that call, the 17th password reset of the day was
just not what I was looking for when it came for

having a job that I really cared about doing. And so
what we've heard from agents who are in the contact

center is, hey, Replicant takes all the minutia.
They do all the, I've identified the person who's

calling. I've authenticated them so I know that
they are who they say they are. I've ideally tried

to resolve their problem end to end. But maybe
they've got one extra question. I changed my

flight. I picked my seat. I checked my bags. But
then I got a question about COVID policies. And so

maybe you don't want the machine answering that
because they might change from time to time. So

we're going to escalate that to the agent. But
they're going to get that warm transfer. It's

like, hey, Benjamin's on the line. He's already
done these things. He's just got a quick question

here about COVID. And so you can kind of be in and out
on that call in 30 seconds, where previously it was

a 45-minute hold time. And then the thing that I
hate, re-authenticating people. I get passed

from person to person to person. And they have to
keep asking me who I am. I mean, it's just, it's a

really annoying customer experience. And so that
kind of, it's really like an elevation of the human

in the contact center. I mean, we do it with a lot of
technology. And we do it with a lot of machines. But

it's really taking what humans are good at,
empathy, creativity, building relationships

with a caller, and allowing that agent to not just
have to churn through as many tickets as they can in

an hour, and trying to get you off the phone. So it's
like, part of it is probably the call center.

People working at call centers might be scared
because their jobs might be taken. But at the same

token, it might elevate their positions. So it's
more meaningful for them to answer the phones

because they don't have to deal with all these
minutia calls about password resets, things like

that. They can try to solve something that's a
little bit more complex and fulfilling on their

end. Yes, yes. And I think it's a great question to
ask, like where and how is this going to take jobs?

Or as I see it, it's really taking tasks away from
people. There's a lot of attrition that goes on at

contact centers. I've worked with some of our
partners who have seven months of retention is

about all you get from a person in a contact center.
It's a difficult job. You're churning through

these tickets. People are angry. They're yelling
at you. You have to get up to speed on changing FAQs

and different regulations and things. It's a
difficult job. I don't think that should be

underestimated. And so anything we can do to
diffuse the situation, I ideally completely

resolve the call without having to get to a person.
But when you do speak to a person, it's a bit of an

elevated experience for the agent and hopefully a
good experience for the caller as well. And

something surprising that I heard is we just had
our customer conference in Nashville last week.

And this was actually very heartening. So when
Replicant was deployed in a particular

customer's contact center, the disability days
that their agents were taking dropped by 70%. So

fewer people are sick on the job when they had
Replicant to sort of be that front line of the tier

one customer service calls. And I don't think we
can take credit for all of that improvement, but it

really just goes to show you that when you shift
from how many tickets can I answer in an hour to,

hey, how do I build a relationship with this person
on the phone and kind of understand and solve their

problem, it's a real elevation of the role. That's
crazy. It's like that brings up the question, is

the job causing their sickness because it's so
much stress or they just hate the job and they're

not showing up and calling in sick. So that's
really, that's eye-opening there. What about

like so when Replicant, how does it work? Like if
someone's using Replicant for their call center,

does the, when the phone call starts, is there some
kind of like cue to the person on the other end that

their AI artificially created voice or how does
that work and then does the call just start like,

did the people, whoever answered the phone know
that it's AI that's talking to them, I guess is my,

the heart of the question. Yeah, ethics has been a
big concern of mine since the beginning. I did not

want to create a technology that was going to be
doing this the way I saw a lot of companies were

doing it when we began, which is I think the wrong
way being disingenuous to callers. I think one of

the early demos I saw of dialogue flow and Google
calling was not identifying itself as a machine.

It was using dialogue covers. It would put the ums
and the a's and the disfluencies and that

click-clack typing noise in the background. They
were doing everything that they could to keep the

caller from understanding that, hey, this is a
machine. And I think it's a lie. I think it's a, it's

a little disingenuous. I think there should be
ample disclosure that you're speaking with a

machine. And so every time we get on the phone,
it's, uh, hi, thanks for calling AAA. I'm a

thinking machine on a recorded line. And so, you
know, we're always disclosing ourselves as the,

as the thinking machine and that's, uh, it can be
confusing, I think, for people. They're like,

wait, what, what's, what's a thinking machine?
And so we might have a little bit of that

conversation or they might immediately push back
and say, hey, there's no way a robot's going to

answer my question. Like get me to a person right
away. And so we've developed a number of

conversation design principles. One of those is
what's in it for me. So when someone says, Hey, I got

to speak with an agent. It's like, you know, Ryan,
I'm happy to get you over to an agent, but it's going

to be about a 25 minute hold time. In the meantime,
can I get your policy number? And so you're sort of

pulling them back into that conversation.
You're, you're training them that, Hey, I know

you've had some really negative experiences with
machines in the past, but this is really a, it's a

new generation of this technology. And so give us a
chance. Yeah. We interviewed sales GPT and

they're doing something very similar, but
they're doing it for sales calls. Like they, they

have automated calls that call out to people and
they do the same thing. Right. When they start

their call, they say, okay, we're an automated
call. I don't know what their exact wording is, but

thinking machine or artificial intelligence
call. They let people know right away. But yeah, I

think it seems like the more and one thing they
brought up when we interviewed them is that we're

at the beginning of this. So it's going to take some
time for people to get used to these calls and

they're going to start to happen all the time. But
once people do get used to them, it's like anything

with technology, I think for my end, at least it's
going to make life easier. Like I'd rather almost,

I think it's going to come to a point where you're
going to get a call from an AI phone call. You're

going to be like, thank God, I don't want to talk to a
human because chances are they're going to handle

it better than a human could. And if it does, if they
don't handle it better than a human could, we can

always escalate it to a human, you know, maybe if
they can't handle it or towards the end of the call.

So it's kind of fascinating. To see how this is
going to all play out. Yes. I think this brings up a

great point of when and how should we use the
technology and how do we know that it's ready?

There's a professor out of MIT, Diren Asamoglu,
and he's writing a lot about so-so technologies.

And so think back on the ATM. The ATM was heralded as
the end of the bank teller. This was going to take

all the bank teller's jobs. There was never going
to be any bank tellers anymore. It was seen as like a

disaster for the banking industry. But what
happened? Bank tellers can do more interesting

things. They can deal with loans. They can get to
know people more. They're not just their

dispensing cash, you know, during business hours
and consumers one as well. Because they can get, I

can get money at any time of the day. And so there's
sort of like a win-win technology that was able to,

you know, see a benefit for society and a benefit
for the people who are working that job. Contrast

that with something that I would say is below the
line on a so-so technology, the robots in checkout

aisles. It's frustrating. I can never quite get
it, you know, to happen properly. It requires

babysitting by, you know, a person. So it didn't
really remove that job. It just sort of shifted to

like an oversight. It was not focused on really
resolving that problem of the checkout aisle. It

was a kind of short-term cost-cutting measure.
And so, yeah, maybe you should save a job or two, but

you're not seeing that benefit for the consumer.
You're not seeing that benefit from the customer

side. And so whenever we're launching a new
thinking machine, a new capability, a new product

feature, we're always measuring it against, is
this technology ready for prime time? Is it

something that's going to make the callers life
easier, or is this going to be a stumbling block or a

road block in their way? Yeah, so I want to comment
on that, the checkout lines at the stores and so

frustrating. I don't understand it. If I'm going
to have, this is my mini rant, if I'm going to check

out by myself, I want to check out by myself, I just
wait for the moment. Every single time it happens,

I try to scan something, it doesn't work. Now I got
to call somebody over. Happens every single time,

very frustrating. But yeah, I wanted to say also
the video on your website was helpful for an

example. It's about a minute and 30 for whoever
wants to see it. Just go right to the Replicant

website, and you guys have a little demo of what it
sounds like when you're talking to one of these AI

entities. And it's very cool, very cool, very
seamless. Since the beginning, we've been really

focused on listening, thinking, and speaking.
Like when we're doing transcription, not only the

words that people are saying, but understanding
the vernacular that people are using. In West

Virginia, they're not going to speak like people
do in California. In Detroit, they're not going to

use the terms of phrase that people do in New
Orleans or in LA. And so transcription has always

been a key focus for ours that we want to understand
everybody equitably, not just people who sound

like me or use the words that I do. On the thinking
side, I think that we should really be doing the

heavy lifting from the caller. And so I've seen
some technologies where it's imagine you're

doing a hotel room booking. And how many adults
will that be? Two adults. I got two adults. Is that

right? Yes. And how many children will that be? One
child. It's such a long and kind of clunky

conversation. And so being able to level that up,
powered in large part by large language models to

say, hey, it's going to be me and my hubby and my nine
month old and understanding that that's two

adults and one child. It's this kind of contextual
reasoning that really elevates the

conversation. So that's on the thinking piece.
And then on speaking, the quality of the voice

really does matter. We will always, always be
upfront that you're speaking with a machine, but

the better that that machine sounds, the more
willing people are to give it a chance. If it sounds

like a robot, doesn't matter if it's the smartest
robot you've ever spoken to before, people are

going to escalate, people are going to press zero.
It's just too easy for them to say, you know what?

I've been burned by technologies like this
before. I'm not going to give it a chance. Yeah.

Like you said, it's like when you release the
software, it's sort of got to be ready for

primetime. Otherwise, it can bite you and
nobody's going to use it again. Or they're just

going to store it. Oh, yeah. Or I'm shopping for a
shovel, and then it tells me that I might be in a

loveless marriage and I should leave my wife. I
mean, some of these early experiences with large

language models there, I think there was a kind of a
rush to get it out into the market and also a

misunderstanding of the guard rails or perhaps
lack of guard rails that some of these companies

put in place. So you mentioned LLMs. Could you just
tell us a little bit about what you're using? Where

are they pulling the data from? Is that how do you
help prevent hallucinations? Can you just talk to

us a little bit about that back-end tech and what
that looks like? Yeah. So we've been using

Transformer since before they were cool. And it's
been a lot of learning from the AI and the ML team

over the years. But we used to use a lot of, let's say
in 2017, it was like BERT. And so we were able to do

entity extraction, entity linking, and tent
recognition. So as opposed to previous keyword

matching, where you've got to call in and say
billing. And if you say my bill is too high, it's as I

started to understand that. And so BERT and that
Transformer technology really unlocked our

ability to do intent-based matching entity
extraction without having to rely on people

saying exactly the right thing at exactly the
right time. Now, large language models have just

pushed this forward in a way that you can get this
kind of single turn capture of a lot of different

pieces of information. It's going to be me and my
hubby and my nine-month-old. And we want to check

in on Thursday and we want to leave two Tuesdays
from now. And being able to do that kind of slot

filling of all of those different pieces of
information, it's a massive time save for callers

on average when we've put these LLM components
into our conversations. It's a 30% reduction in

time on the line, which is great for the caller.
It's also good for our customers because we charge

per the minute. So the shorter the calls are, the
cheaper it's going to be for them. And we saw a 10

percentage point improvement maybe in things
like make model collections, so trying to collect

vehicles for AAA on the side of the road, being able
to do collection of complex pieces of data that you

need to get. And the CSAT, the customer
satisfaction score went up because people could

really talk in their own voice. You mentioned
hallucination. I think that's a really key aspect

here. I think the first thing I would do is I would
caution people against thinking about large

language models as human or reasoning or thinking
in the way that humans do. I would add that they're

more than human. They are kind of beyond or
different than a human intelligence. And I think

that when you think about it in that way, it kind of,
it doesn't prohibit you in the way of thinking

where animals have more than human intelligence.
Fungus has more than human intelligence. And

until we created the internet, we really didn't
understand how do mycelial networks like talk to

one another. And so here we have another
technology that even though it presents itself as

a friendly agent that you can talk to, like the way
that it learns and the way that it thinks, or

perhaps interprets the data that it's seen, is
kind of not like people. And so hallucination is

kind of the name of the game with these
technologies. I don't think you can fine tune or

reinforcement learn the hallucination out of the
system. You know, they choose the next token and it

could be the correct thing or it could be the
incorrect thing. But once you've chosen that

token or once the machine has chosen that token,
it's going to vehemently defend what it's chosen,

even if you tell it that it's kind of wrong. So in
addition to putting guardrails on those language

models, I think there's a few ways you can do this.
You can constrain the technology so it's only

allowed to do certain things. This would be like a
make model collection. We'll use the large

language model for, oh, it's my 1993 Chrysler
Sebring. And then they say, oh, actually, it's my

2024 Tesla Model Y. That's really great for the LLM
to be able to collect. And you can speak

colloquially and you can do all of these things.
But we don't necessarily allow the language model

to do any natural language generation. So you lock
down exactly what the LLM can do and you defer to a

dialogue policy, which might be constructed with
the customer. You defer to scripts, which can be

written with the customer so that they've got it in
their own language, in their own tone. And it

prevents some of these embarrassing issues that
you can imagine where, imagine you're calling

into an airline and you're complaining about the
cost of your bags. And I can imagine the LLM would be

like, well, why don't you fly Southwest next time?
You get two bags for three. And so it's not always a

hallucination. It's not always a dangerous piece
of information that's being responded. But it

could just be a little bit damaging to the brand,
even if it is factual. Sure. But it sounds like

that's going to happen no matter what. It should be
a very, it's going to be over time, a smaller and

smaller percentage of the times it's going to
happen. But it's going to happen. And you might

even see social media posts on something funny
that happens on a phone call. But it's just going to

happen and it's eventually going to get better. So
you just sort of have to deal with it because it's

not human thinking, it's machine thinking, and we
don't really fully understand how it's

capabilities right now and what it's going to do.
Yeah. And my guidance would be in the face of that

uncertainty, it's a little bit difficult if you
want to go build these things in-house. You can

certainly do it and you can hook up a raw LLM to your
callers, but you're going to have these

hallucinations. So then the question becomes,
well, what are you going to do about that? And so

we've come up with, I mean, we've spent so much time
thinking about constraining what the LLM can do,

coming up with dialogue policies, grounding the
LLM in, could be FAQs for a customer, it could be

logged in data for a customer. Hey, why did my bill
go up this month? And so you could compare this

month's bill with the last month's bill. And then
we're also building ancillary models that are

looking at the output of the large language model
and then in parallel saying, that looks like a

malicious query or, hey, that looks like
something that's not factual or, hey, that looks

like something that is not befitting for the brand
guidelines that the customer wants. So it's

possible that someone could go and build all of
these pieces, but it's a heavy lift and it's

something that I don't know if all of the LLM
companies are being that kind of honest about. It

might work in the ideal case, but the failure modes
are quite numerous. Yeah. Like, so for example,

say I'm talking to an AI customer agent and I start
swearing at them, telling them they suck or

something like that. How do you guys respond to
those situations? Well, it really depends on the

situation. We work with a number of trucking
companies in addition to AAAs. And so with

trucking companies, I would be surprised if we get
through a phone call without a little bit of

expletive. Like it just happens. It's the way that
some people choose to speak. And so I don't think

you can always be so prescriptive as, oh, as soon as
we hear someone say a certain word, let's go ahead

and escalate that call. It's another reason why we
really care about the intent. Behind what's

happening on the call or being able to, you know, if
we got all of the entities and, you know, they put an

F bomb in front of it, let's just go ahead and
continue with the call. Because really that

person, even though they're frustrated and we
know they're, you know, if you've gotten to the

point where you're calling customer service,
you're probably nine out of 10, you know, you're

90% frustrated already. And so it's okay if you're
using the words like this, as long as we're still

pushing the conversation forward. And that's a
call that each one of our customers can make. And

there's self-service tools that they can use for
like, well, what do you want to do if this is

detected? What do you want to do if the
conversation starts to go in these different

directions? It's almost like de-escalating
things is maybe the best practice in those

situations. situations. Yeah, yeah. What are the
most frustrating things to do? Maybe someone

needs to just blow up a little bit. You know, hey, I
feel like I'm not being, you know, I got I ordered

this product, it's not working the way that I want,
and I really want to kind of let loose on the company

a little bit. But if you can do that to the machine,
and kind of, you know, guard or shield the agent

from having to get the full brunt of their, you
know, annoyance, by the time they either

completely solve the problem with the machine, or
maybe they do get to the agent, they're not quite as

like riled up as they initially were on the call.
Yeah, that's a great point. I mean, it saves the

company to dealing with the mental health of all
their agent, like human agents, not having to deal

with all that stuff. So that's awesome. I didn't
even think about that. One of the main one thing

that's really annoying when you're talking to
customer services, like you said, if you're

talking if you're making a call to customer
service at any company, usually you're at like a

nine out of 10 in terms of frustration. And then you
start talking to someone a human and they're very

nice. You're trying to be polite to them and try to
solve your problem. But since you're nice to them,

they feel like they can sort of drag this call out
for a freaking half hour, whereas you just want to

get off the phone and get this problem solved in
five minutes. So that it sounds like that's

something you guys can help solve as well. Yeah, I
mean, the reason the calls are, you know, 30, 40%

shorter than speaking with a person is, yeah, I
don't think you expect small talk from a robot. I

don't think that you, you know, you might ask it how
the weather is and we respond to that. But I, you

know, or just like that one with it. Yeah. And
practice, it doesn't really happen. And, you

know, you can go to our website, you mentioned that
the video is there, but it's also if you go to

try.replicant .com, you can speak with the
thinking machine. And it's really interesting to

see the way that people talk to the thinking
machine when they've actually got a problem

versus the way that they talk to it from trying it
out or being at kind of like the top of the funnel, if

it comes to buying, I think there's a little bit of
like, well, let me try to throw the entire kitchen

sink at this, you know, machine and, you know,
what, what's it going to do when I bring up, you

know, whether you like the president or not?
What's it going to do when you bring up, you know,

like cultural issues? In practice, we never
really see that on phone calls. I mean, people, you

know, they just want to get in and they just want to
get out and kind of go on with their way. And it's a

little bit of like when Amazon was in the early days
and people were like, well, don't, don't you want

to smell the books? And it's like, yeah, you, you
might want to have that book reading experience

some of the time, but most of the time with customer
service, it's very transactional. In terms of it,

so you guys have been around for quite a while, when
did replicants start? And when did you start

offering the product, your customer service AI
product? Yeah, we began in 2017. And it really was a

lot of research, it was, it was kind of
foundational blue sky research into listening,

thinking and speaking. What should the
conversation design look like? What are the tools

that you want to use to build this? I think an early
observation I had was that in order to craft these

interactive voice response systems, IVR
systems, that's where you're pressing one for

this and pressing two for that. And sometimes they
talk and sometimes they don't, they've been

around for a long time. But to make that it was all
XML. And so you had to be a programmer in order to

make this stuff. And so is it any wonder that you
ended up with press one for this and press two for

that and press three for this? It sounds like a
mathematician like made this kind of technology.

So if you can make something that's graphical, I
used to work on graphical programming languages

for teaching children or teaching maybe your
parents how to code. And so pulled a lot from those,

graphical programming tools, app inventors is
one that has been released from MIT. That was an

early project that I worked on. But we brought in a
lot of that graphical programming tooling into

how you would make the machine. And it was a really
fun time to not only see technologists play with

it, but like what will happen when you give this to
people who work in contact centers? What will this

look like when you give it to their supervisors?
What about when you give it to, you know, the CFO and

a company? It becomes understandable. It becomes
something that you feel like you can actually

impact and change rather than the kind of, you
know, inscrutable XML that is not only hard to

understand, but it's hard to program. I mean, like
imagine making like if statements and while loops

and all of this in a, you know, in an XML document,
it's just hard to do. And so a lot of those early days

of research was like, can we make this fast enough?
Can we make it smart enough? Can we make it sound

good enough? Can we create the tooling that will be
really intuitive for people to use? And then, you

know, it's really only been like three years that
we've been, you know, selling the product. Some of

our earliest customers were, as I mentioned, like
DoorDash, Postmates doing the food ordering

expanded into Because Market. That's what they
sell products to older folks. And so maybe you're

calling in about incontinence products. Maybe
you're calling in about, you know, CBD, topical

appointments, things like this. But that was
another challenge for us because we're working

with an older population. And so, you know, will
the technology that we built for restaurateurs

also work for, you know, folks who are a little bit
older that need us to speak a little bit slower than

might need us to repeat ourselves. And it really
worked with flying colors. I think the ability of

the machine to be very persistent, you know, it'll
repeat something as many times as you want. It's

happy to like stay on the line there with you, you
know, for asking for a policy number. And you've

got to, you know, dig it up out of a, you know,
physical piece of paper that's like upstairs in

your filing cabinet, like we'll just stay there on
the line with you where I don't know if an agent is

necessarily going to, you know, want to expend
that amount of time because again, in that older

model, they're gold. They're, you know, sort of
paid on how many calls they can get through and

that's sort of an important metric for them. Yeah.
And plus the person that's actually needs the

help, he's talking to a human, so he feels bad about
taking forever to do something, you know, to find

an ID or whatever it might be. So there's that
pressure there too, even though you don't even

know who this person is. So it's cool. Yeah. And if
you're talking about incontinence products, I

mean, I would personally feel more comfortable
talking to a machine about that rather than a

person. You know, I think it's a little bit easier
for me to be honest. It's a little bit easier for me

to like actually say what's going on. And so that's
a, you know, an element that might play into it as

well. People say, I'm kind of glad I'm speaking to a
machine about this. Yeah, I can see why, you know,

it's, I don't think people are going to realize
that until they actually start using it because

the first thing you think about is I'm talking to a
machine. I already talked to recorded voices and

people in India. I don't want that. I want a real
human, but there's going to come a time where

they're going to realize it's actually better to
talk to the AI, you know, and I think that's coming

soon, which is very cool. Yeah. You know, I think
three years ago, four years ago, when we would show

this to people, they would say, no way, this is
real. You know, we think you're, we think your

videos are fake. We think your demo is fake. Like
there must be, you know, a person hiding inside the

machine somewhere kind of coming up with these
answers. But since open AI put chat, GpT into the

hands of, you know, 100 million people very
quickly, I think it became obvious that this is a

technology that is here, is only going to get
better. And you probably need to put some serious

guardrails on this because I think everyone's
probably dealt with a hallucination or something

that feels inaccurate. And that's great if you're
writing a novel or you want this to be kind of a

writing partner for you or you're playing a game
and you want kind of a lot of creativity. But in the

customer service space, I think you want, you want
factionality, you want factuality and you want to

be grounded in truth. And you brought up like one
open AI came out, which is it, like the AI hit the

scene in about two, late 2022 is when it got really
popular, at least with the general public. How did

that change the world for you guys? Because you've
been in the game since 2017. You said you released

your product three years ago. How did that
technology change the technology that you're

using to do all this at Replicant? I'm glad that it
showed people that this is a technology that's

here and is ready to be used. I think it changed
people's mind, which is to your point earlier,

it's convincing people that maybe they should
hang in there and talk with the machine because

this is this kind of new batch of technology that's
not like what we had in the mid 2010s. It's not like

what we had in the early 2000s. It's definitely not
what we had in the voice synthesis of the 90s. And so

it's changing people's minds, which is great. And
there's kind of a rising tide that's lifting all

boats. But being able to understand how and when
and really if you should put these LLM tools into

particular turns, it could be data collection,
like I mentioned before, it could be

summarization for the handoff of the call.
Imagine I have a whole conversation with the

machine and then, well, I've got to escalate that
over to Ryan the agent. Let's generate a quick

summary. Let's even come up with the first line
that the agent can say. So it gives you a little bit

of cover while you say, okay, it's Benjamin and
he's authenticated and he's on the side of the road

and here's some pictures of his car. So that's
another place where we can kind of use LLMs to level

us up. It's also changing the quality of the voice
and the kind of realness that we can put into the

thinking machine, even if we are telling people
that they are talking to All right, awesome. And

what can you get into? I mean, feel free to tell us as
much as you want or as little as you want. But what's

the actual technology you're using right now?
Like how do you generate the voice? What LLMs are

using? Things like that. And if you can't let us
know, we understand. But that's my question.

Yeah, I think our thesis from the beginning is that
technology is going to progress quite quickly.

And so it may be technology that we use in house. It
may be technology that's coming from a third

party. It may be technology that becomes kind of,
you know, ubiquitous like transcription. I think

Google will give you transcription and Azure will
give you transcription. And we did understand and

do understand the right turns to use the right
providers. Google, because of Google Maps, very,

very good at collecting addresses. And so
they're, you know, quite good in that realm. When

it comes to alphanumeric collection, that's a,
hey, my policy number is X is in X-ray, Z is in zebra,

75, 52. Google's not good at this. Azure's not good
at this. Whisper is not good at this. It's just,

it's a difficult problem to solve. And so for that
piece, we built our transcription in house so that

we could, you know, understand the words that are
being spoken, but then think about it to know, you

know, X as an X-ray is not double X, but that's just
one, you know, X there. And so it could be

alphanumeric data collection. It could be email
address collection. That's something else that

we're quite good at. And I mean, that's hard to do.
If I asked you for your email address, I might need

you to repeat it a couple of times. It's just kind of
a hard thing to get across. There were some other

early models we made. We worked with Hawaiian
Telecom, so a telecom company out of Hawaii. And

we're getting people who are calling in with names
that don't sound like mine. And so the standard

transcription providers, again, they're kind
of, they're a little bit biased toward the data

that was used to train that model. And so being able
to do phoneme based matching of names, even for my

name, I'll call anytime I'm talking to a, you know,
an AI system, and they say, what's your name? I say,

it's Glikesman. And usually what I'm getting is,
it doesn't appear that you're talking to me. And

like, what a rude thing to say to a person. It's
like, I've just told you my name, and you're going

to tell me that, you know, I'm not, you know, you
haven't understood what I've said. And so we built

a lot of models in house to make sure that we can nail
that and like get that done properly. So the keys I

think are in each turn of the conversation,
knowing which providers to use could be third

party providers could be in house providers for
transcription for intent and entity recognition

entity linking, whether or not we're going to call
out to like an LLM model, and then choosing what

voice we're going to use to synthesize it. Another
key point here about LLMs is that they're not HIPAA

compliant. They're not PCI compliant. And I
expect that we'll be able to get there. But there's

a lot of turns that might have, you know,
personally identifying information, credit

card numbers, things like this. No way we're going
to pass that to an LLM, even if we have data

redaction, you know, policies in place so that
they don't use it for training. It's just, it's not

something that we can actually pass, you know,
legally to the system. So we still use a lot of our

in-house models for those things. And you can also
imagine that we might use multiple providers.

What if we use multiple transcription providers
at once? We'll use our in-house and we'll use, you

know, Azure. And then over time we'll learn sort of
which turns are more kind of high value for us.

Which ones, you know, do we mostly get the right
thing on the right turn? So it's quite a complex

system, but I think the key is being like malleable
and able to swap in new technologies as they come.

Because if you're keeping up with the papers every
week and every day is just like massive gains in

this industry. And, you know, it reminds me a
little bit, I was young, but like the newness of the

internet, the kind of like, you know, the how quick
features were coming out and how rapid the

evolution was. It's an exciting time, but it's a
time when having a really strong ethics grounding

and having a really strong like security, how are
we going to make sure that we keep these things on

track? How are we going to make sure that we can
responsibly, you know, manage data is really key.

So what is the next step in your development and
where do you see your project going in like maybe

five years, 10 years? You know, when you think
about contact center automation, the contact

center is quite big. And this was conversational
AI. And conversational AI is so broad, it could be

analyzing what agents do and giving them
recommendations, not in real time. It could be,

you know, churning through emails and being able
to respond to those. It could be things like we do,

which is kind of, you know, voice or SMS or, you
know, chat on websites. And so contact center

automation has a lot of additional pieces to it.
It's got that agent handoff and making sure that

that's really solid so that you don't have to
repeat yourself and kind of reauthenticate. It's

got knowledge based kind of agent assists for
agents as well. Maybe you're new on the job. And so

you, you know, want to get some FAQs or you want to
get some recommendations from the machine of what

might the customer be calling about or what are
some ways that you can kind of answer their

question. I think it comes into like workforce
management and how are you routing particular

calls to the right individual. And so, you know,
there are many players that are in this market, you

know, the five nines of the world, the genesis of
the world. They've built these technologies, but

they haven't built them in a way that's kind of
modern facing and not that is, you know, powered by

AI from the start. And so the big shift I see in the
industry is Genesis is charging per seat. So

they're saying how many agents do you have in your
contact center? Your cost will be number of agents

times, you know, some amount of money that they're
charging per month. The future of the contact

center, it's not 90% agents and 10% AI. I think it's
90% AI and 10% agents who are there to like give that

empathy or to give that creativity. And so it
really flips that whole business model on its

head. And so you've got a little bit of an
innovators dilemma when it comes to these legacy

providers who need to move away from the, you know,
like towards AI, which is moving quite rapidly,

but they're stuck a little bit in that per seat
model. And so that's something that I get excited

about thinking about the future of like, how do
you, how do you transform the entire experience

before the call is made during the call, after the
call, the kind of data and analytics that you can

like gather from that. It's a, it really is an
underserved market. I feel bad for folks that have

been in this industry since the 80s and 90s because
they've been sold technology that was supposed to

make their life better for a really long time. Hey,
I've got an IVR touchtone system. You can't wait to

see how, you know, this is going to change your
life. And then, hey, we've got an IVR touchstone

system, but now you can say billing to it. Like you
can't believe how this is going to change your

life. And it just hasn't had the, it doesn't have
the benefits that haven't had the improvements

that were promised. But now I think we're finally
to a point where you can get end to end resolution

using an automated system that's going to be
faster than talking with a person, better

informed perhaps than, than that person too. And
the caller experience is going to be really

quality. So people are going to ask for these
machines. It's fun talking to people in the AI

space because when we talk to them, all these ideas
pop in my head at least of like what, what can

actually happen with these tools that are coming
out. And one of the things that's, maybe that will

happen with these phone calls is one of the stories
that we did was on how AI can listen to someone's

voice and tell if they're under duress or they have
anxiety, depression, something like that. So I'm

thinking maybe when the AI is talking to the
customer and the customer is like, oh, yeah, I'm

good. This is all good. You know, thank you for
taking care of me. The AI, the AI can subtly

understand if they're really saying that just to
say it, or they still haven't had their issue

resolved. And then you guys can follow up with them
with a phone call or an email or something like

that. Like just the, there's so many
opportunities out there that seems like that's

something that could eventually happen. Yeah,
it's, it's a promising future. But I also, I

sometimes caution against using like, is there
really a problem there? And one example I'll give

is roadside service. So we get a lot of people
calling in on the roadside, they're anxious,

they're nervous. We know that they're on the side
of the road and they're, you know, in danger. You

don't necessarily have to kind of like read the
tone of their voice there. They'll probably tell

you, you know, kind of the information that they,
it's a little bit like the truckers and

profanities. Like just because it's there, it
doesn't mean you have to act on it. And I think

that's another kind of key. It's a key thing to
consider when you deploy AI. I think about Twitter

for, I don't know how long it was, months or years
would crop photos to find faces. And so you'd

upload a photo and it would crop it in the preview to
find a face. Well, it cropped white faces more than

black faces. It cropped people who were standing
rather than people who were sitting in

wheelchairs. And so the decision was made by that
team to not use the AI anymore for photocropping.

And you have to kind of ask, well, was that even a
problem in the first place? And you've got these

ancillary folks who are excited about it as well.
Photographers who spend like painstaking effort

to like crop the photo in the way they want were
like, thank you so much for not re-cropping my

photos. I don't even like want this. And so it could
be things like analyzing stress and voice, which

could be useful in certain situations. It could be
voice as a password, which I think is very

dangerous. We're in the industry of synthesizing
voice. I know how easy it is to synthesize voice

based on a snippet. I don't know if it's
necessarily safe to use voice as a password. It

might be a username. You could say, hey, when when
you hear my voice, that's probably Benjamin's

username, but you don't necessarily have
authorization to do things on my behalf just

because if you can synthesize my voice. So
choosing what and what not to do with these

technologies is important. It's going to be it's
all a big trial, right? Like people are going to

screw up and try something, they're going to mess
it up. There's going to be bad publicity about it,

but that's just part of the game with technology.
Like you got to go through those growing pains.

Yeah. And it's a real, I think it's a challenge and
it's heartening to me to think about it is the

people who work in these industries that will be
the choosers of whether to use this technology or

not. I don't think the regulation is necessarily
going to come from the government or it may not look

the way that we want. I don't know if it's going to
come from the CEOs of companies. I think it's going

to be the kind of rank and file employees that look
at the ways that we're rolling out AI and either

say, yes, this is something I feel comfortable
doing or no, I don't and trying to work within their

companies to either not deploy technologies when
it is an example like the cropping of photos or

choosing how and when it's used to make it more
ethical. Benjamin, who's using Replicant right

now, do you have like a count on how many companies
are using it or like what types of companies are

more attracted to your technology? Yeah, it's
mostly in the enterprise. We have like tens of

customers. And so that's really exciting, kind of
more joining every year. I would say industries

like insurance are really big for us. Food
ordering, starting to get into more retail

customers. We've avoided places like financial
services just because there's a lot of regulation

that comes with dealing with banks. Also the sale
cycle on banks is quite long. I could see us moving

there now that we've got our GDPR certification.
That was quite a large lift for us. And so excited

about moving into more areas where this can be
used. But what's surprising to me is across

industries, people say we work with, let's say,
pest control companies. So this is actually a

really great usage for LLMs. Like I got scorpions
in the den again and kind of understanding what

you've got and where you've gotten all these kind
of problems. But they say, well, how many pest

control companies have you worked with before?
And at the end of the day, people are people. They

talk mostly in the same way, the same kind of
problems we had to solve with food ordering kind of

come back when it's pest control or insurance use
cases. And so now having done a couple hundred

million calls, we've really got the data sets.
We've got kind of the understanding of not only how

most people talk, but also how do we deal with folks
who might have a particular accent who deal with

English as a second language. And so the time it
takes to bring on a new industry is not really so

high because all of the conversational smarts are
kind of built into the system. So you really just

focus on where is my industry different and what
are those kind of jargon keywords? And is there

anything, scheduling a pest control appointment
looks like scheduling a dentist appointment

looks like scheduling an MRI, just swapping out
those kind of back-end APIs. And then in terms of,

have you seen any kind of general percentage like
from the companies you've worked with that use

your product? Do they say, hey, this has saved us
50% of our cost to pay a human to do the same job? Is

there any kind of benchmark there that you've seen
from your customers? Yeah, I think the metrics

that move are, one use case might be a customer who
does, imagine you move to a new location, but

you've got to send back your Comcast modem or
something like that. And so they're in the

business of reclaiming these modems that people
can't even use anymore. You just got to send it in.

And so it's usually an outbound call or an SMS or
something like this that's going to the person in

order to get them to send these back in. When they
brought in Replicant, it was like a drastic

savings, millions of dollars of savings of cost,
something that they really couldn't even reach

the number of people that they needed to because
there just weren't enough agents to call all the

people that had to return these modems. It was a big
win for the customer, the callers into the

customer as well, because they don't want to be
charged $85 or $110 for this modem. They're not

going to use it. So it was like a reminder it's
saving the money to come back in. And the return

rates went up. I think it was like an 11% increase in
returns. And so it's very rare. I've certainly

never worked at a company before where it's higher
customer satisfaction, it's higher kind of

bottom line KPIs and metrics like returns. You
save a couple million dollars and the agents in the

contact center get happy. It's really exciting to
see and kind of heartening to see that there's not,

I think in business, at least for me, you're
looking for the trade off. We saved money, but

people were angrier. We made the agent experience
better, but we had to spend a lot more money on it.

It's like rare to see when you can move all of these
metrics in the correct direction, but you don't

really have to have a downside. It's almost like
your biggest problem is the sales end. I think this

can solve a huge problem for a lot of people, but
people are scared to use it. And once their eyes are

open up a little bit, I think a lot of people are
going to be reaching out to you guys to implement

something like this on their end. Yeah, I think the
message is that the time is now. There's going to be

this, I would call it like a K shaped recovery where
the people who choose to use AI now are just going to

continue to get better and better and better at
that technology. And the people who delay are

going to fall further and further behind. And so
it's not just bringing in this technology, it's

getting your own data in order. We can tap into any
knowledge base, we can tap into any FAQ, but if your

FAQs disagree with one another or they're not kept
up to date, then that's going to be a problem for the

machine in the same way that it's probably a
problem for your agents. But maybe a seasoned

agent is going to be like, oh, yeah, well, I read
that document, but it's not right, you've got to do

the other thing over here. And so there's kind of,
in addition to bringing in this technology,

there's a little bit of improvement you need to
make within the organization of getting

knowledge bases up and running, making sure that
you've got API access. We want to do everything

that an agent can do, but if you don't have an API for
it, then it becomes difficult for us to do. So

sometimes we work with RPA or like other
technologies that can get us access to those APIs.

But yeah, I think if the question was, should I wait
for next year's GPT model or should I go and start to

do this today, I think now is the time to really
move. Definitely. And if you're unsure, just take

a look at there, replicant.com and look at their
website in the video. I mean, it is pretty stunning

what can be done right this second. For something
like customer service and these calls, it's

awesome. You've got to check it out. Yeah. And I
would also say that a lot of customers I talk to say,

well, let's run a POC. Let's run like a trial of
this. And the problem with running POCs is that

it's kind of the same lift to do a POC as it is to do an
actual deployment. It's not like we can cut

corners, because you've got to make sure that the
conversation design is there. You've got to make

sure that the APIs are in place. And when you're
running a POC, I think you're often looking for

well, this didn't quite work and that didn't quite
work. And so you're coming up with reasons to not

quite take the plunge. And so regardless of the
vendor that you go with, I think choosing to kind of

do something fully, and you don't have to redo your
whole contact center, you can say, hey, we're

going to take DigiPress number four off our IVR and
just do billing. We're going to take DigiPress

number two and just do returns. Take the plunge,
jump in, start to understand what this technology

is good for, and you're going to be part of the
people who are getting richer using this

technology rather than those that kind of delayed
and fall behind. What would you like to promote,

Benjamin? I know replicant.com is the main thing,
but anything else that you want to put out there,

we'll put all the links in the comments below.
Yeah, I think replicant.com is a great place to go

to. You go to try.replicant .com to give the
thinking machine a spin. We can put my email out

there if people want to get in touch. I'm happy to
talk about the technology, how it's used. I also

really care about the ethics that are behind it.
That's a big piece for me. We just had our Resolve

conference, and so we're doing Resolve Rewind,
which will be some videos from that. We have a lot of

great speakers. It's not just salespeople
talking from replicant. It's like actual people

who have used the technology talking about how and
why and where they used it. Hopefully, we can lift

this industry up together. Something that
surprised me is that in the context center, you can

still be working a phone, and then you become a
manager, and then you become the head of the

company. It's like you can still move up through
that organization. The people really matter. The

people that stick with the organization are
really important. Being able to hear from those

experts and being able to hear from people who've
already taken the plunge is great. Cool. Then be

sure to subscribe to Ryan and I's newsletter
fry-ai .com. We have three top stories a day along

with some cool tools and community engagement
things as well. Just thank you so much, Benjamin,

for coming on today. We really appreciate it.
Thank you so much for having me. Great questions

and really enlightening.