Introduction

Hello everyone, I'm Mark Janji. I'm basically an AI lead at Kaizo.

About Kaizo

Kaizo, we are a small startup that actually have our office downstairs, first floor. 15 people, actually 17 now. We hired two people two weeks ago, so we're growing a little bit.

Kaizo's Mission and Services

Yeah, so what do we do? As a little bit of a background, we are basically a workforce management platform for customer support teams. We plug in basically to whatever the customer, the head of basic customer support uses. Let's say, for example, if there is a customer, our customers are, let's say, heads of customer support.

So it can be a footlocker. It can be a retailer. It can be anything. And we basically plug into the CRM.

We monitor the interaction between the agent and the customer. We collect a lot of metrics about that, quantify it. show it to our clients to reflect transparency and how basically the team is doing we have different parts in our product cover different basically users from the agent to the team lead to a head of customer support to a quality assurance expert

AI Implementation and Industry Changes

But the question is, yeah, why am I here? Basically, I'm here to talk a little bit about the learnings that we've had implementing AI at Kaizo and basically how things were before OpenAI and before this, let's say, watershed moment of GPT, how things shifted in the perception of the customer support world when it comes to the application of AI, and then where do we see this going?

Kaizo's Approach and Philosophy

and the funny thing is that we have a ninja here and this is kind of our approach or used to be our approach from the beginning is that every basically customer support agent is basically a superhero we don't see their faces they're always working in the background solving your issues and that was actually the focus of kaizo for a long time it's about gamification motivating the agents to work hard to meet their goals and improve over time so kind of continuous improvement

Challenges with Early AI in Customer Support

1And the reason we went with that direction is because the customer support world was actually burned by AI many years before through the promise of, let's say, chatbots that are going to basically automate everything. That never came to light, and the time to value was just too high. too much investment, too much questions about, let's say, the security of the system.

Just give us more data. Just give us this little bit more, and then we can give you the best sentiment analysis out there. We can give you the best summary out there. We can solve all your questions.

In the world of customer support, there is PII, private personal identifiable information. This is very sensitive information that basically IT SEC teams in the customer support field find it very dangerous to share with someone else, a third party. In this case, it's Kaizo. So that made the conversation of using AI very difficult. Anything that has to do with getting the ticket content and looking at it was kind of taboo before GPT.

Impact of GPT on Customer Support AI

But then basically GPT happened and the world kind of woke up to what AI can do with no data. no training data anymore. And this kind of reduced for us the time to value drastically, but we still had the same resistance.

It's like the same skepticism, the same conservatism towards like, hey, you know, this time is it going to be different? Shall I invest, you know, some time in it? And we saw that a lot with all of our customer calls.

I used to be part of them and the head of customer support would say, yeah, this is all great. Just get back to me in a year and tell me if it works. So we had that.

But as time progressed, as people saw the value of these models on their own time, they started to see, oh, wait, this is something I can actually use. This summary actually is great. I can use that to reduce my time, the time it takes for me to look at what the agent had done, and then immediately know if they've done something wrong or right, and then they did a great job or not. And this is called quality assurance in our world.

Strategic Shifts at Kaizo

So let's say overnight, we had to do like a strategic shift, if you will, from being more, let's say, centric on giving, let's say, from an AI perspective, developing analytics and models that rely on correlations to do some sort of estimation of the best thing that you should be doing as an agent or the best thing that you should be doing as a team lead to just give us a ticket and we will check the compliance we will screen it for you we will summarize it we will transcribe it we will basically check the sentiment check the empathy and just reduce the time it takes to review a ticket by let's say 20 or 30 out of the box And that was a major, let's say, shift for us from a strategic perspective, because now we can focus on the future that is also interesting for our customers.

The Future of Customer Support Agents

And basically the building block of a customer support team, which is an agent, the smallest unit, is now transforming into a little bit of a kind of a human with the symbiosis with an AI, right? but we're seeing more and more deflection with AI, but we still have huge hurdles, huge blockades to take over, which are basically integrations. No customer support ticket can be solved without actually, let's say, going to the system and doing the thing you should do. It's not as simple as just answering a question.

These we can do since years. And this is where we're basically hitting the wall, right? So everyone's saying, yeah, this is great, all nice, sentiment analysis is perfect, empathy is great, but can you just actually solve it for me? That's difficult at the moment, and this is where we see a lot of need for engineering at the moment.

Automating Compliance Analysis

There are a shit ton of systems out there, sorry for the word, that include, let's say, there's knowledge bases in whatever notion, there's knowledge base in what's solid, and they might have actually images in there. Some of our customers would go into the ticket And they cannot actually review it because they want to see what the agent has done exactly. So they watch a video of every single action that the agent had done. And that takes time and takes effort. And that's what we're focusing on automating at the moment, which is compliance analysis.

so the principle of compliance i don't know if we're doing well on time um so qa in the world of customer support is a very basically old discipline since the 80s where people sit down with a used to be a spreadsheet now it's a nicer app where they sit down they read the ticket they check some boxes the agent did follow the security verification process The agent didn't share information they shouldn't have shared, like GDPR features. It can be very diverse. It can be super specific to a customer's need.

And at the moment, it only covers 2% of the tickets that the customer support team solves. And 98% goes unnoticed, unscreened, can have violations of a compliance rule nobody would know. 1And we took the direction of actually trying to automate that process, that compliance, and scale it to cover all tickets to give the visibility to our clients.

Compliance Needs for Humans and AI

Because the way we see it, basically compliance is something that will be needed for both humans and AI in the future. So kind of our focus shifted from developing things only for humans to things that would work for whoever is solving the ticket. And that was kind of the biggest shift for us.

We still encounter similar issues when it comes to integrations because, again, if you want to validate that the agent did the correct security verification, you need to actually go into the database and check that the name of the customer was inserted correctly. Good luck if you don't have that integration. So there is a lot of missing data. There's a lot of gaps.

Integrations and the Road Ahead

And the future looked, at least from an AI perspective, looks brighter for customer support. But the main issues are still there. And there's a lot of, basically, need for engineering.

Conclusion and Vision for the Future

efforts to basically develop all these integrations and make the whole thing smoother and let's say the long-term vision that we see is that once we have this compliance engine let's say for the sake of the imagination and it works perfectly it can screen all the tickets understands every process that the client has or the customer has and then there's no stopping it from actually solving the ticket And this is kind of the vision that we have. So going from assistance to compliance to solution. And yeah, I hope we get there.

And I hope that basically all these issues that we're facing at the moment, these kind of the latency is too high, the context window is too small, the hallucinations are too bad, you don't support this tiny language that 100 people speak in this island. That all is going to be hopefully resolved with time, and this is kind of where we stand. I hope this was actually useful for you, and I would take any questions that you might have.

Audience Q&A

So the previous talk was about marketing.

Early Pilots and User-Driven Development

How did you guys land your first five or 10 pilots? That is a very good question. That was before I joined Kaizo.

So basically, Kaizo started off with being like any other startup as a part of an incubator at the time in Amsterdam. And then through the connections and basically the people that our CEO knew, he was able to do the networking and basically get first clients or first... users and then from there the value started increasing and then we could basically use the same client that we had for the same business let's say a retailer and then we can just go and reach out to other retailers like hey you know this guy yeah we can give you the same value and we worked very closely in the beginning with these one or two or three clients building the product based on their needs and that turned out to generalize to other basically users and we went from there so kind of a user-driven development

Wasn't it risky to build features specifically for those? It was, and we still have some pains from that because we built something thinking that, hey, this is going to be used by everyone. But it turns out to be a very niche sliver of customer support world that would like to look at this. So there's a danger there indeed.

And at the moment, things shifted. So from being, let's say, whatever you want, we're going to build to whatever brings most value to the most amount of customers and basically commercially driven development.

Marketing and Client Acquisition

And how do you reach out for new clients? We have a lot of inbound coming in due to some marketing efforts. We have also... Through the websites, through Google Ads, through several different options.

We also have our sales team, which does outbound, basically cold calls, emails, all that jazz, which I know nothing about as an AI engineer. But yeah, this is mostly how it's working.

Thank you.

Balancing Productivity with Employee Satisfaction

What is the line that you might be seeing between clients driving against effectiveness and efficiency and more productivity versus employee satisfaction, employee happiness? Everyone's overloaded, everyone's running a startup, everyone's got too much work, not enough people. Is all the drive to just be able to do more with less or is there any drive at all in the market to do more happier?

That's a very good question. And this is the case for gamification, right? This is the case for the part of the product that we didn't kill yet.

It's because when our customers come in, they do have these things for their basically agents. We do have the ninjas, we do have the gadgets that you can get and all that kind of competitive, nice aspect. But we also try to utilize the, let's say, time saving efforts for the management level. At least this is what we're basically perceiving to be the most effective in, let's say, gaining new users.

But yeah, indeed, there's always, we do have clients that are on the efficiency side. It's like chopping off whatever. I want to just trim the fat and make sure that everything is going smoothly. But we do have a lot of clients who are mostly focused on the human, the human centric.

And this is basically customer support world. If you work in the world of customer support, you meet a lot of warm people, a lot of people who are willing to help. And most of the managers there are actually like that. The cold-hearted people usually come from the more like disconnected level of head of customer support that, you know, makes sometimes the shots, but actually the people kind of call in for like, hey, I want to use this tool because it's going to save me this time and it's going to make me happy. So pay for it. And sometimes it works.

Kaizo's Integrations and Client Base

So what platform are you using for customer support to get to that point? What do we plug into? So we have two integrations at the moment with Zendesk and Salesforce. Salesforce is one of the biggest ones.

What's CRM? I mean, like Customer Relationship Manager?

We have clients as big as 300-400 agents. Our system is designed to handle a lot of scale. Most of the hiring we do is in people who write Scala, which is known to be highly scalable. Actually, one of the engineers is right there.

We're also hiring for engineers and product engineers, if you know anyone.

So one last question. Yep, in the back.

What's your level of experience and trust using the AI tools with translating into other, like you mentioned obscure languages, but how about just kind of the more common languages as well? I ask because my company, we have to do a lot of that.

And second, it goes into anything more technical. If I didn't have a local read over that, it would just look quite unprofessional.

So what we've experienced is that for low resource languages, as I mentioned before, like anything that has dialects in it, like Arabic languages are usually very dialectal. And with that, we've noticed a lot of failure in basically producing accurate translations or transcriptions for that matter.

What the analysis showed is like for 20 languages, let's say 20 European languages, the support is quite good. yeah we didn't notice any difference from a or any complaint let's say from our customers who are let's say bilingual or they have agents basically speaking in german italian and russian it works but when it comes to other let's say Korean also ends in there, but then it gets more difficult as you have less and less basically trained data these models have on.

But I can always share with you the list of basically languages that we confidently support and the other ones that we are planning to actually improve. Amazing. Thank you so much.