Klarna CEO: We’re Giving AI More Customer Service Work, Not Less

Channel: Alex Kantrowitz

Published at: 2025-05-16

YouTube video id: pNd4KlccWAY

Source: https://www.youtube.com/watch?v=pNd4KlccWAY

CLA is in the headlines for walking back
its ambitious plans to automate customer
service with AI, but its CEO Sebastian
Shimi Andovsky is here today to say the
company is actually accelerating those
efforts and push back against the
headlines. Sebastian, we last spoke a
year ago. It's great to see you again.
Thank you for coming in for a quick
check-in in a bonus episode to talk
about what's actually going on with
Clara's AI efforts. Good to see you.
Good to see you. It's like funny like
pushing back against my own claims just
like it feels like uh I don't know. It's
funny though. Well, yeah. So, I think
that the reason why I wanted to do this
interview is because we spoke a year ago
about your effort to automate what was
then said what you then said was 700 uh
customer service jobs with AI. Now, it
was part of a broader AI initiative
within the company. And there have been
some recent headlines that have talked
about how CLA is now pulling back from
that effort. Like our conversation was
all about is this real? You convinced me
that it was real. And now there are
headlines that Clara is pulling back.
But in the time since those headlines
have come out, you've actually gone on
uh to Twitter and talked about how uh
actually CLA is not pulling back from
this program. Uh and you wrote we are
actually accelerating it. So what is the
truth? Is there going to be uh the is
this the end of handing customer service
over to AI at CLO or is it the
beginning? Well, we can start with like
uh where does the confusion come from?
So I did an article with Bloomberg um
where we talked about a number of
aspects uh to AI and customer service
and actually if you read the original uh
Bloomberg article it's quite balanced
and uh describing quite accurately what
what I said and what happened but then
what tends to happen in classic media is
that they take just the headline of that
and then somebody just quotes the
headline without reading the actual
article and then that gets a lot of
errors are built up on errors and then
suddenly uh for whatever reason this
seemed to have been a very interesting
topic in the world of that I'm surprised
how many people got engaged with this
but no but the but so let me let me give
you uh what's actually going on u the
truth and the truth is that um we two
things can be true at the same time so
uh which my wife always says and I think
it's a smart saying helps a lot when you
quarrel with your wife sometimes that
like you realize that like there isn't
actually anything to quarrel all about
but um the point being is that um uh
there are two things right like
cla is very much leaning into AI and
obviously
um customer service is one area where
that's the case our agent that we
launched about a year ago is still uh
dealing with I think it is about 1.3
million errands per month which is the
equivalent of previously was on by about
800 uh people on a monthly basis and um
so that is obviously and it's actually
gone up a little bit. It used to be
about 700, now it's about 800. Now with
that said, you know, again, media has
tendency to kind of simplify things and
it's like the the message we always said
and we've always been in in podcasts
like this, I've been very honest about
the fact that like some of those errands
were very simplistic, right? they could
have been uh you know hey did I pay to
cla have you received my payment the
answer was yes and then that was the end
of the errand so um I think from a like
if you are doing like let's say advanced
merchant support with technical
integrations that is obviously something
that is more challenging to offer over
an AI than the kind of question that I
just described right so um AI to us
initially was clon was not necessarily
the best company in the world when it
comes to like IVRs and FAQs and so
forth. A lot of companies implement
these things. Yeah. Exactly. And I think
that like so at some point of time it
was predominantly like level one support
as we would describe it, right? And then
uh but what has happened since then
since you and I spoke uh you know a year
ago is definitely the quality and the
ability of it to deal with more complex
errands has increased. So it's kind of
eating up into like level two support
etc. uh as we're going
um this is true. What is also true is
that throughout this process and having
the experience of this, one of the
things we started reflecting on is that
there is a strong negative sentiment
around AI because people feel threatened
uh for their jobs which is natural. We
understand why. And also if you kind of
reflect on historical events where
machines took the jobs of humans, it
actually usually led to a higher
appreciation for what was humanly
produced. So if you take like you know a
lot of things that were manufactured in
factories or you know assembly lines
like even 100 years ago whatever even
before you know when there was entirely
people were sitting in factories but
they were doing all the work by hand. Uh
once machines came along, a lot of these
products at first were low quality then
become higher quality and then machines
could produce at a fairly high quality
and then people stopped doing a lot of
that, right? And machines do shoes today
or you know help do clothes, not all of
it but part of the process. And so what
you're saying is basically that a human
customer service agent is the quality of
is the equivalent of a artisally
produced good. Um I think partially
maybe not so much that what because what
I wanted to get to is that once that
happened we started suddenly appreciate
crafted items right so today if you buy
a piece of furniture that is done by
artisan or uh an artist or somebody you
know is done like a humanly crafted we
actually pay a higher price for that
than we pay for something that comes out
of a standard factory in some you know
uh it's just like manufactured by a
machine. And so our conclusion
reflecting on these things was that like
the human connection matters. People
appreciate talking to a human. They feel
a human connection. There's an emotional
connection. And we believe that like
this means that there will be a higher
appreciation. And if a company wants to
be competitive, it it will actually be a
competitive edge to offer a human
connection. And so, but obviously that's
kind of a different type of human
connection than maybe some of the
customer service we offered historically
because this will have to be, you know,
high quality uh
um skilled people that are very familiar
with clon and understand clon and that
was not always the case, right? We
relied a lot on like outsourced agents.
They maybe would come in, they barely
knew our product. they were just like
asked according to some very strict
template to answer like have you paid or
not paid and that was it right and now
we want to build the customer service
that are people that love the product
understand the quality that was true for
some of our agents as well but now we
want to make sure that those are the
ones that we premiere in the interaction
with the customers so we actually think
two things can be true at the same time
the more simplistic straightforward
stuff will be dealt with AI but there
will be a premium to talking into a
human and also you know some of the more
complex aspects
um should be dealt with with people that
really love our products, engage with
our products and you know appreciate our
products and could be those kind of uh
amazing support people. Okay, so that is
a really good explanation of what's
actually happening. But before we blame
the media for everything, uh let's let's
let's try to figure out why that is.
because um if you look at a lot of your
past statements, I think headline
writers could be forgiven for saying
that this is a backtrack and uh in some
ways it seems like it is to me. So for
instance, in our conversation last year,
you mentioned that customer uh
satisfaction was equal to a human in
many cases when people were dealing with
AI. You also told uh Bloomberg in
February that you're of the opinion that
AI uh can already do all the jobs that
we as humans do. So, do you So, do you
think that you were a little bit
too confident in AI's abilities? I I'm
I'm You could definitely place me in the
Elon Musk box of like saying
everything's going to happen in 3 months
when it probably will take three years.
Like, you can definitely like I'm happy
to, you know, I'm guilty as charge on
that one. I think though that those two
statements I would argue are still
fairly accurate in the sense that
um if we take the the layer one first
which is around like can AI do our jobs.
Now obviously today like to me it's a
little bit like the LLM in my opinion
has reasoning capabilities. A lot of
people will like argue is it reasoning
non-reasoning is it just using what it's
been trained on and so forth. But I
would argue that it that it has
reasoning capabilities and most of the
tasks of knowledge work that we do may
be very complex and
sophisticated on a higher abstraction
layer but
basically can also be divided in smaller
uh reasoning tasks that are uh more
basic. Right? And so the point is that
like if you if you ask an LLM like give
me the best way to recruit people,
right? Like you might give very high
level answer. It might not actually be
so perfect or whatever. Um but if you
give it a very small concrete challenge
and say you know tell me what you would
expect to see at the front of a house it
will be like doors and a window and
there will be no hallucinations. it will
always be accurate in that in that
reasoning uh exercise. And so my point
with stating that is just that like the
challenge right now for the technology
is how do you take large complex
problems and break them down in smaller
reasoning challenges that an LLM is
already capable of performing. And I
think to some degree if you look at deep
research how that actually technically
works it is basically what it's doing.
it's just taking more time to think and
it's doing more smaller tasks that then
it aggregates up to answer and that's
why we get higher quality responses than
we would get with an instant prompt
response historically. So I think it's
just a continuation of that as you kind
of extend that you could argue that the
core technology in place already is
capable of doing a lot of our jobs. just
that we haven't figured out exactly yet
how to set it up and configure it and
how to build like the tools around it to
allow it to perform those task uh
entirely. Right. I just want to make
sure that I concretely understand uh
again the what's going on with the AI
customer service roll out. So when we
last spoke last year you said se it was
doing the equivalent of 700 full-time
employees. Yep. Now, even though uh
you're gonna basically hand over some of
the tougher tasks to humans, you're
saying that you've expanded that to do
the work of 800 full-time employees. Do
you see that? So, can you just confirm
that? And do you see that number of
full-time uh employee equivalent work
going up? I think it will continue to go
up because what happening basically is
that and that's also when you ask me
like the customer satisfaction is equal.
But the point is was it equal on the
most basic tasks or was it equal on the
most complex task that our customer
service would get and it was obviously
the average of that. So to some degree
if you think about it a chat a person
working in a customer service make it
very concrete a person working customer
service I think all of us have had that
experience you're talking you're
chatting with somebody let's say you
chat with somebody in customer service
for efficiency as much as any company
may claim it's not true that agent will
have five parallel conversations going
because you as a customer may also be
slow and you may be busy with other
things you're not going to write all the
time so that agent is not going to sit
dedicated to your chat alone he or she
may deal with five six in at the same
point of But that also is going to slow
it down because when you actually do
write to the agent, they may be busy in
writing an answer to somebody else at
that point of time and that slows it
down. So that's why the average time is
11 minutes or used to be 11 minutes for
us. It was to a large degree due to
that. But it makes sense like I'm not
surprised that all companies do it that
way and it makes sense for the our
customer service agents as well. Now
when you have an LM it can answer
instantaneously. So if you ask a very
simple question like you know have I
paid my uh tlana that question will get
an instant answer. Our average time for
such errands dropped to two you know one
two minutes and obviously that customer
I would argue is more satisfied with the
AI experience than with the human agent
experience for that very simplistic
task. But when you go to more complex
questions like very complex questions
where AI may not be capable of answering
I would argue the the satisfaction is
higher with a human agent than it is but
the average may be the same across
these. Right. So what we're seeing is
but the numbers Yeah. Exactly. But the
AI So the numbers are rising. The AI is
becoming thanks to we are enabling it
with better and better tooling. A very
good example of that would be one of the
most complex things to answer is when a
customer gets rejected by CLA because
it's a sensitive matter. Why was I not
given credit? What is the reasoning for
you rejecting me and so forth our a year
ago our AI could not answer that
question uh because it didn't have full
access to all the data. It couldn't take
into consideration the sensitivity of
how to deal with such information in
order to help answer to a customer such
a question. Today we've enabled it to do
so and it's actually you know people are
uh you know it's it's allowed us to
answer much more complex errands. So
that's example of something that it
couldn't do and that's why that number
of the job equivalent is going to rise.
At the same point of time we think every
customer of ours should always be able
to click a button that says always human
and would they have a preference to
speak to a human? They should have the
right to do so. Okay. I want to ask you
a couple questions before we leave. I
want to ask you about this Chimath
tweet. I think we'll have to talk about
it because a lot of people have inferred
this uh move to be that um we're now
seeing proof that customer service
cannot be done by AI and it's just a
nature of the technology itself. Here's
Chimath. Not nearly enough people are
talking about the implications of CLA
rolling back some of their AI bets. Not
knowing any of the details, I can guess
why. replacing determinism or humans
with probabilistic code is fraught with
edge cases and require many new ways of
software development uh and process
engineering that aren't well solved yet.
The implication to an entire generation
of AI apps will be severe as more
companies come to terms with the
difficulty in getting products to work
reliably in production with AI in the
loop. Customer service may be the first
funeral signpost. Basically, he's saying
AI is probabilistic, not deterministic.
you can't rely on it for customer
service. And Clara learned the hard way.
Now, you responded and told Chimath is
not really true. Um, but I want to hear
your response here on the show. My
response is not true. Oh, yeah. Yeah.
That we didn't roll it back. Yeah,
that's right. No. Uh, well, he is
obviously an very very smart man and he
understands this at a very deep level
which is uh, you know, very impressive.
Um, uh, this is obviously a challenge
that we faced. Um and my impression
because I've been trying to be quite
open with what we've been doing. I've
been describing what we've been doing
also uh partly I've been describing like
how do we model data and how to what are
the internal workings that we try to
apply to kind of accelerate us to the
next level of using AI. I have
highlighted these in some tweets. Um and
the challenge between probabilistic and
deterministic is very very true and this
is exactly the problem that we
internally are tackling right now and I
we've been working on that for over a
year because we had that insight already
a year ago as we've been testing and
we're very early leaning into the
software we saw that exact challenge and
we've been working for a long period of
time of how we are going to um solve
that when we look internally
We believe we have a solution uh that in
a prototyping and demos work but now
needs to go into full production and it
seems to be solving this specific
problem that he's describing and I hope
that we'll be able to bring it in in to
market very soon to prove that we
believe that it's been partially solved.
So what you're saying basically is that
yes this is a real issue that what a
generative AI is doing is making
probabilistic sort of guesses as opposed
to deterministic pulling from you know a
logic tree and that can be an issue with
customer service. However you're saying
that this is a solvable problem. Am I
capturing that right? It is a solable
problem. Now the question again we can
come back to is it an Elon Musk
prediction to say we will launch
something in a few months or is it a and
no like that that's a different topic
but yes it's definitely a solvable
problem. It's just that it I think it
the what we see clearly as we've been
hypothesizing and we believe found a
solution to that problem that the
challenge with it is it requires a new
way of thinking for the people building
the systems even if more and more is
coded by cursor and other parts. You
still need the humans that are working
on this to adopt new ways of thinking
and how they build systems, how they
build data models, how they build all
these things and that in itself is
actually one of the biggest challenges
because right it is very hard because pe
all of us are kind of you know based of
20 years of working in a particular way
and now suddenly all of us needs to
think differently. Okay, I definitely
want to have you back to talk about that
in particular. All right, let let's wrap
with this. Um, artificial intelligence,
your company has, you've said you wanted
to be the guinea pig for basically
generative AI uh in all sorts of
applications. People have seized on this
moment and said, "Aha, Clara, I can't
make it work as uh what did Chimath call
it? A funeral post." Um, so I'm just
curious from your perspective, why do
you think so many people are so eager to
find data points that show that this
generative AI moment is just not going
to work?
I think there are probably multiple
reasons I've seen there are like you
know I think at the core of it it is
fear right and I have also tried to be
honest with it myself in some my posts
I'm also you know I also look at AI with
some fear right like I also think but I
think it's just a question of like when
you face something unknown or something
that is a little bit scary you have two
options of how you apply yourself as a
human being. Either you kind of lean
away from it and say that's scary. I
don't want to know of it. I don't want
to use it. I don't want like you know
and I I I cherish if somebody tells me
that it's not going to work or that
scary thing is going to go away. I'm
going to be all over those posts and
cherish that. Ah, exactly what I said.
It's scary and it's bad and I don't know
what it is. My approach however and what
we've been encouraging people to do at
Clona is to lean in rather say what is
scary is what we don't understand. So
let's try to understand this as good as
we can. Let's learn what this can be
used for and how and by doing so I think
it has reduced the amount of fear. It's
not that I don't have concerns of what
the implications for society may be or
you know future state but I I leave
those to you know the more uh futuristic
philosophy conversations for a dinner
and at work I try to focus at the
practical applications of this in our
industry and what it means for how we
can serve our customers and in that case
I think leaning in is the right thing.
You have to lean in. This is a
technology that every company has to
lean into and learn and then you it
becomes less scary as well. Well,
Sebastian, I always enjoy our
conversations. Thanks so much for coming
on and clearing all this up. Thank you,
Alex. Good to have you. Thanks for
having me. Yeah, and we definitely got
to get you on to talk about Google's
valuation because we've been talking
about your tweets on the show every week
as you go back and forth trying to
figure it out. So, uh, come back soon
and and thanks again for making the
time. Thanks everybody for listening and
we'll see you next time on Big
Technology