Building AI Agents That Actually Work | Malte Kosub, CEO of Parloa

Malte Kosub is the cofounder and CEO of Parloa, an AI agent platform for customer service. They raised their $66M Series B led by Altimeter.

Malte's favorite book: The Qualified Sales Leader (Author: John McMahon)

(00:01) Introduction
(00:26) Overview of AI in Customer Support
(01:33) The Current Landscape of AI Agents
(02:46) Enterprise Adoption of AI Agents
(04:16) The Founding Story of Parloa
(06:25) Deciding What Goes into V1 of a Product
(07:56) Achieving 99.9999% Accuracy in AI Agents
(09:29) How to Identify Customer Needs for AI Products
(10:55) Scaling from Early Customers to the Next 10
(12:41) Growth Experiments: What Worked and What Didn’t
(14:42) Current State of Parloa: Capabilities and Scale
(16:36) Structuring Teams for AI-First Companies
(18:49) Technology Stack and Internal AI Use Cases
(21:29) How to Pitch an AI Product to Enterprises
(23:32) Essential Tools Used Inside the Company
(25:41) AI’s Role in Daily Life and Workflows
(27:54) The Future of AI Agents in Customer Support
(28:37) Can AI Agents Fully Replace Human Agents?
(29:05) Exciting AI Advancements Impacting Parloa
(30:06) Rapid Fire Round

--------
Where to find Malte Kosub: 

LinkedIn: https://www.linkedin.com/in/maltekosub/

--------
Where to find Prateek Joshi: 

Newsletter: https://prateekjoshi.substack.com 
Website: https://prateekj.com 
LinkedIn: https://www.linkedin.com/in/prateek-joshi-infinite
X: https://x.com/prateekvjoshi 

Transcript

Prateek Joshi (00:12.786)
Malte, thank you so much for joining me today.

Malte Kosub (00:24.28)
Thanks for the invitation, Pratek.

Prateek Joshi (00:26.544)
Let's start with the basics. AI is everywhere, and customer support, customer interaction is an area where it's very, very active. So can you explain what are all the things a customer support AI agent should be able to do?

Malte Kosub (00:47.406)
So obviously there are a lot of things an AI needs to do to solve customer support issues. The first thing is it should be naturally talking to a customer and be polite, be empathetic. But that's not enough, right? You need to be able to answer questions. So there should be a knowledge base connected, but you also should be able to connect to back end systems in order to solve end to end problems because most of the cases,

Customs are calling because they have an issue, so you need to solve something in the backend. So it's about having conversations, having access to knowledge and having access to backend systems. And then you should be able to talk to the customer on the phone, in a chat, via messaging and so on.

Prateek Joshi (01:33.052)
How would you describe the current landscape of AI agents for customer support?

Malte Kosub (01:39.734)
Yeah. So you have the legacy conversational AI players that are still focused on having a natural language understanding. They have workflows like if and else trees. They infuse large language models a bit, but they're not agentic first. And then they are the agentic first players still. A lot of them are focusing on chat, adding voice step after step, but I would...

position those two as the players driving automation. And then you have the agentic players that do agent assist. So they actually support humans in doing customer support.

Prateek Joshi (02:22.438)
And when a company, a large company, when they want to adopt an AI agent for this, what are the things they have to do? And also, if you look at the broader landscape, where are they in terms of adoption? Are they still in the experimentation phase? Are they fully using agents and replacing humans? Can you talk about the phase we are in right now?

Malte Kosub (02:46.892)
Yeah. So I think we saw a couple of different phases in customer support. So 20 years ago, it was primarily touched on IVRs, right? Then it started a bit with keyword-based IVRs. Then we had the classical voice bots, which had...

and our use and the workflows I mentioned, so the legacy players. And now we are moving towards agentic first where large language models actually drive the conversation, drive the reasoning. And in the last two years, a lot has changed. initially, enterprises were very careful of adopting agentic first solutions. They wanted to do it, use it internally.

like for agent assist, then it was more or less the first stage in the agentic wave. The second stage then was more to use it internally so that employees can chat with an AI agent. And what we're seeing, particularly in the last six to nine months, is that companies realize, OK, agentic,

customer-facing use cases actually can drive a lot of value. They can be safe. They can be reliable. So we see a lot of enterprise now moving towards the customer-facing agentic use cases. But that is something that happened primarily in the last six to nine, maybe 12 months.

Prateek Joshi (04:16.732)
Let's go back to the months leading up to the launch of Parlova. So let's go back to that time and what was the biggest gap or the angle of attack that you saw that made you want to launch the company?

Malte Kosub (04:33.902)
So we launched already in 2017, right? And we initially built one of the first agencies for conversational AI in the world. In 2017, AI was not the hype topic, right? And as the name says, Future Voice, we were believing that voice and phone still is the most relevant channel when you look at customer support for the biggest enterprises in the world. And back then, everyone was talking about chats and chatbots and that

Prateek Joshi (04:47.707)
All right.

Malte Kosub (05:04.046)
that will eventually revolutionize the way how customers interact with companies. And we believed already back then that voice AI will revolutionize the customer support which is doing, which gets done over the phone. So.

We saw a huge gap in platforms that bring voice AI into enterprises on scale. That was why we initially started the company. And when we look at the agentic development over the last, let's say, three years, the biggest gap we saw is that everyone saw the potential, but just a large language model.

is just a little piece of value delivery you need in order to deploy agentic systems in the enterprise with end customers. So a company needs to do a lot of things in order to deploy AI agents globally, on scale, with the highest reliability, highest accuracy. And we wanted to close that gap. And that's what we did. And yeah.

Prateek Joshi (06:25.212)
How did you decide what goes into V1 of the product? And maybe we can talk about both back in 2017 and also more recently with all the new LLMs coming up. So if you had to relaunch or launch a new product, how do you decide in general what should go into V1 of any product?

Malte Kosub (06:47.948)
Yeah, so roughly two and a half, three years ago, we decided to completely re-architect the product based on an agentic first architecture. So build everything from scratch. Again, we build an AI for Seekers platform that scales globally with lowest latency and so on. And we built the agentic platform on top of that.

And we are always a fan of directly working with our enterprise customers and focus on what they need. So we had first use cases we identified and worked towards solving those use cases and take it from there. One of the biggest challenge at the beginning obviously was instruction following so that the AI agent is actually following what the enterprise process.

needs and that was one of the first problems we needed to solve because you can create a 75 % working demo in like two hours, that's easy, but in order to get to 99.9999%, this is the hard part.

Prateek Joshi (07:56.37)
How do you do that? Like follow making the product follow all the steps and get to 99.9999 % accuracy?

Malte Kosub (08:07.758)
Yeah, so I think there's a lot of things you need to build around the model in order to guarantee that. So first of all, you have easy guardrails through prompting. The second one, we have a compliance layer where you have actually rules that help.

the agent consider if the next step is the right step and those clear compliance layer helps you to guarantee that it's always following the right rules. If you have backends, you have the API as another layer of safety, right? Then we have a safety layer which checks if there are harms, harmful language.

or if the customer wants to trick the agent in order to create hallucination. And there are a couple of other layers on top which helps the AI agent to actually be reliable, accurate all the time. And just if you put all those things together,

you can get to that accuracy level because otherwise you're at 75 % or less or maybe a bit more but it's not enough in order to be sufficient for an enterprise.

Prateek Joshi (09:29.106)
And two and a half years ago when you were re-architecting everything and you said you like to engage with your customers to do all of this. So what questions do you ask your customers when you're doing this thing? Or in general, when you're building something, what questions do you ask to know what needs to be built?

Malte Kosub (09:51.438)
Yeah. So I think you always need to talk with the customer about the impossible. So what would be a 10 X solution for you? dream of the perfect solution? And then you get out to a state where you can think out of the box with your customer, And you might not be able to get there in six or 12 or 18 months, but at least you get out of the thinking of, how can I?

solve it in two or three months but you actually building a much bigger vision and that's how we approached it and then we needed to tackle one problem after the other which worked out pretty nicely.

Prateek Joshi (10:34.436)
And let's say you build it and then you're working with your early pilot customers. How do you acquire the next 10? The ones you didn't talk to, but they just buy your product because it's good. So what's your playbook to get those next 10 customers?

Malte Kosub (10:55.116)
Yeah. So first of all, belief is our customers needs to be our best sellers. So our customers needs to love the product and need to be ambassadors of the company. And they should

They should be champions for us. So that's, think, the first pillar. If you don't get this right, you won't be able to acquire a lot of customers in the speed you need in order to build a global company. The second part is that we have a partner first go to market. So we're working a lot together with partners because we believe this market is changing so rapidly. You need partners on your side in order to bring

a platform into the biggest enterprises on the planet at such a rapid speed. So we are closely working with partners that accelerates our go-to-market. The third, yeah, Pratik.

And I think the third one is that we have a focused enterprise go-to-market approach. So we search for the right events or the right moments where we can target specifically our target group and get into a conversation and also pluralizing with clearly stated visions about the future and showing them, how will

the enterprise of the future look like because we believe customer experience will completely transform and you need to have topics where people listen to you. So there are a lot of different things, but those are, I would say, three pillars.

Prateek Joshi (12:41.906)
And now you're at a bigger scale. In the last two and a half years, you've grown. During this time, what growth experiments did you run? And more importantly, maybe pick one that worked really well and one that you thought would work really well but just didn't do anything.

Malte Kosub (13:04.684)
Yeah.

And again, we are very much focused on enterprises. For us, it's not product-led growth. It's talking to a few, but very dedicated people. So we need to come up with a marketing that is tailored and very high quality. So we created, for example, a lot of...

tried to get into conferences with very clearly defined keynotes and those keynotes had more or less nothing to do directly with our product. Those messages were clearly defined of this market will completely revolutionize and we pictured how an enterprise should look like in five years and we made companies curious to talk to us because

A lot of those things they heard probably the first time and we get them the feeding. are someone that can help them in order to get there.

And that worked out very, very nicely. Another way of approaching our potential customers is obviously via partners. So we do a lot of campaigns with our partners, very vertical specific, very use case specific, and tailor a lot of content around that and go to our customers.

Malte Kosub (14:28.526)
with our partners. But again, it's not the product-led growth motion. So it's very tailored, very industry specific to a very targeted audience.

Prateek Joshi (14:42.706)
Can you paint a picture of where the company is today? Meaning, what are all the things the product can do? And also in terms of scale, whatever you can disclose, number of customers, revenue, like where are you at today?

Malte Kosub (14:58.062)
Yeah, we're not talking about numbers, but the last years were very successful. We raised our seed round in 22, series A in 23, series B in 2024. So every 12 months we did a round and we are growing as projected and that we're very ambitious in our plans. When we talk about the product, what can it do?

Basically, we call our platform an AI agent management platform. So companies need to deploy agents for a lot of different use cases for a lot of different regions. they should create, we always call it personal AI agents. So if a customer has 100 million end customers, they need to create 100 million personal AI agents. And they need to have a platform in order to manage them, in order to deploy them. And with our platform, they can

manage, build and deploy them end-to-end, deploy them across the globe with the lowest latency with the highest accuracy.

Yeah, a lot of use cases are already available. We enable them via a UI so they can actually build the AI agents within our front end. And the interesting thing is that they are actually building the AI agents by briefing them, by telling them the use cases, the patterns, how they should behave. And it's a bit like briefing a human agent, but this time you brief an AI agent.

Prateek Joshi (16:36.21)
I wanna shift the conversation to company building. And an important part of company building is structuring a team. And there are infinite number of ways in which you can tinker with it, you can do anything you want. But at this scale, pass cities be, how have you structured your team? And when I say structure, it means, sales, marketing, product, engineering, like what do you have and relatively, how are you staffed?

those teams in terms of percentage of people in the company.

Malte Kosub (17:10.198)
Yeah, so

Obviously at the beginning we were very, very engineering product heavy. So we had like 80 % from engineering people in order to build the product. And at a certain point where we found product market fit, we heavily scaled our go-to-market team in Europe, but also in the US. as we are focused on enterprise customers, you need to have the people in order to build business

cases in order to build a tailored solution for an enterprise. those are the teams consisting of an account executive, an engineer that helps the account executive to architect the solution. And then a team that even helps either helps the partner to implement the solution or implement the solution directly with the customer.

The interesting thing is that with AI, lot of the common structures you know from software as a service and high growth companies needs to be redesigned because AI can take over a lot of things and you can create way more efficient processes, way more efficient teams. And we have been doing that already in a lot of areas and we are continuing to do so because a lot of efficiency games, we sell energetic

products, we need to be an agentic first company ourselves. And that's what we're living every single day.

Prateek Joshi (18:49.522)
Actually, that's a good segue into my next question is, what is your own technology stack? Like what have you used to build the product? That's part A. And part B, even outside of that, how do you use AI internally at the company? Could be a product that's available out there or a tool you built yourself. But basically, where does AI make an appearance?

in the way you run your business.

Malte Kosub (19:19.84)
Yeah, so first to answer the question on the product. we are building this end-to-end platform that helps an enterprise launch AI agents on scale. So we are not building our own large language models or our own text-to-speech or end-to-end audio model.

We don't want to reinvent the wheel. We bring all that technology to an enterprise so they are equipped to actually deploy it. And there are a lot of things you need to build around the actual model in order to do that.

To answer the second question on which processes have we redesigned on AI, maybe to give you one. So in the past, our engineers in the sales process, they obviously build demos, they build individual demos for our customers because they're a big company. So it makes also sense to build individualized demos. And that always took a lot of time. So you need to research the right use cases. You need to build a mock

a backend which the AI agent is talking to you to come up with a knowledge base, you need to set up the whole system. So the tech, took some time and we created actually an agentic system that we just need to trigger and say, this is the customer, we have the meeting in two days and we are talking to that specific department, please create a working demo and then everything is being built automatically.

And the hours the engineering in sales needs to put into building a demo went down from probably eight hours in order to build a very sophisticated demo down to 20 to 30 minutes. And that is just one example of how you can.

Malte Kosub (21:12.64)
increase the output of the teams significantly. Another one is answering RFPs and large documents, right? In the past, you need to fill out everything and now we do that way, way faster.

Prateek Joshi (21:29.01)
Obviously, you've pitched many, many customers, you've closed many of them. Now today, as you're sitting here, what goes into that? Meaning, how do you pitch an AI product to an enterprise customer? And if you were to guide maybe an earlier stage founder on the do's and don'ts, how would you guide them? What should you do when pitching an AI product? And more importantly, what should you not do during that pitch?

Malte Kosub (21:55.584)
Yeah, and I believe you need to.

talk about the value the customer gets, right? Not about the technical details and what nice technical problems you solve. You want to talk about the value you are delivering. And obviously at the end, also depends on the buyer. If the buyer already is completely sold on the value and just needs to have a product that solves a problem, then you can directly start about that. But in general, would say always talk about the value and the transformational change you can drive.

And we with customer support are completely changing how companies interact with customers in the future, right? We're not just changing customer support, we're changing customer experience as a whole. And you need to bring this message to the customer and then at the end say, hey, and we can help you to get there and focus more on the value. That's the first thing. I think the second thing is

AI is impressive, creating demos tailored to your customers in our case helps a lot. Obviously you should.

first start to understand what is actually the pain the customer wants to solve. Do they really want to do something or are they just looking around? But as soon as you have the feeling, they're really interested, then not just talk about the value, but also actually showing the technology.

Prateek Joshi (23:32.146)
And as you look inside your own company, obviously there are so many products that you'll use. There's cloud, CRM, email, marketing, so many things. What are the three most important or most useful products that you use inside your company, something that you can't live without?

Malte Kosub (23:54.99)
That is a very good question. So obviously it's a lot about communication. So Slack is our internal hub of communicating. Then we're using a tool to leverage large language models in order to write texts all the time and everyone is using it like every day for.

a lot of time because it increases your speed. And then I think you need to go into the different departments in engineering. It's something like Coercer, which is used heavily in sales. It's agentic systems we're using in order to create the demos or RFPs. So then it comes down to the different departments where we leverage different tools in order to improve the processes on AI. So I would assume if you

ask one department it would give you a different answer than another department.

Prateek Joshi (24:52.912)
Right. That's amazing and that's fair. Maybe one final question on this thread. You personally, like not Malta the CEO, Malta the person to do your work, to live your life. What's an AI tool that you've embraced that perhaps you weren't using five years ago or three years ago?

Malte Kosub (25:14.55)
Yeah, this is probably very straightforward. I use Chachiviti like day in and day out all the time. And I get a lot of productivity out of that. And it also changed my search behavior significantly. So my search went down probably 95 % on Google versus now talking to Chachiviti.

Prateek Joshi (25:21.595)
Yeah.

Prateek Joshi (25:37.521)
Right.

Prateek Joshi (25:41.266)
Yeah. All right. Coming back to AI agents. Now we are in this midst of, in the middle of this amazing wave. AI is making a huge impact. How do you see AI agents evolve in the next two years?

Malte Kosub (25:57.196)
Yeah, so.

If you look at customer support, a lot of companies think about efficiency gains, right? And I think this is just one little part of what they actually drive. I think they drive customer experience. just imagine, you want to call United for a flight and then normally you might wait to get someone and then every time another agent picks up, right? And now imagine you call and it's your personal AI agent you can speak

to and this person actually says, Pratik, how are you? you decided if you want to upgrade your flight next week? If no, no worries, just call me back in five minutes. Then in five minutes you could call back. And after one second, the same personal agent picks up same voice, knows everything about the past conversations and say, Hey, have you decided? And if you don't want to call, you could pick up your phone and you could write on WhatsApp or iMessage to have that conversation. And you could say you want to upgrade. And if you don't chat or call on Monday,

this person AI agent might call you and says, critique, there's just one seat left in the class you wanted to upgrade to. Do you want to have that seat? Otherwise I would give them to someone else. So it's a completely different relationship between the customer and the company. And if United has 120 million customers, United has 120 million personal AI agents. And this is the start of

building relationships, one-to-one relationships. So it's way more about efficiency gain. It's about building a relationship with each and every customer. So I think that's the broader theme that we see. think if we look at the next 12 months, the way that customers engage with AI agents will be way more natural. So if you think about the movie Her,

Malte Kosub (27:54.316)
I deeply believe we're getting there just in the naturality how a conversation feels.

Prateek Joshi (28:03.212)
And if you look looking forward, will we reach a point where AI agents for a specific use case or a specific function, can they fully replace a human agent? Like for example, calling an airline, I'm happy if an agent knows that has 100 % of the context and just answers my question and they're like a million agents so you don't have to wait on the phone. So will there be a point where we can reliably

have AI agents addressing all these calls.

Malte Kosub (28:37.71)
I would say 100 % of the calls that will take some time, but that AI agents replace a lot of the use cases to 100%, yes, I believe this is possible in the near future.

Prateek Joshi (28:52.778)
And maybe one last question before we go to the rapid fire round. What advancements in AI are most exciting to you as it pertains to Parlova?

Malte Kosub (29:05.912)
So I think that different areas, one part is multimodality.

so that you can talk, you can chat, you can actually share your video and talk to an AI agent, which opens up a lot of different use cases for companies, right? If you talk to an insurance company and you want to file a claim, you get up your phone, you talk to the AI agent, but you also share your camera because something is broken in your house. And then you have a conversation with the agent showing actually what is broken and the agent talks to you, okay, please show it from another side, okay, understood?

So multimodality, think, is a very, very interesting and I think important topic in order to improve experience. And the second thing is reasoning models because at the end we have improved the quality and the capability to also solve very complex use cases. So those are two very exciting developments.

Prateek Joshi (30:06.642)
Amazing. With that, we're at the rapid fire round. I'll ask a series of questions and would love to hear your answers in 15 seconds or less. You ready?

Malte Kosub (30:15.81)
I'm

Prateek Joshi (30:17.266)
Alright, question number one. What's your favorite book?

Malte Kosub (30:21.358)
If you look at building an enterprise company, I would say I would recommend the qualified sales leader.

Prateek Joshi (30:33.126)
what has been an important but overlooked technology trend in the last 12 months.

Malte Kosub (30:41.358)
that AI agents talk to each other. So you have an AI agent in that domain to talk to an AI agent in another domain and to clearly define protocols how those AI agents can engage.

Prateek Joshi (30:57.028)
What company do you admire the most and why?

Malte Kosub (31:01.688)
So there are a lot of great companies out there. Maybe to name one, Adyen, another European company. There's a lot of similarities. They work with the biggest enterprises in the world. They are business critical software. They are coming from Europe. Their claim is engineered for ambition, which is very much aligned with how we approach the product. So maybe that's one company to mention. Yeah.

Prateek Joshi (31:26.719)
It's a great company. What's the one thing about AI agents that most people don't get?

Malte Kosub (31:33.326)
One very interesting discussion is always latency, yeah, because you need to wait one second, two seconds to get a response and it feels like a human is faster. But if you look at the whole customer support conversation, actually it's way faster because they can talk to API is way faster. actually latency, if you look at the whole conversation is actually not a topic, but it's always the main topic to talk.

Prateek Joshi (31:55.25)
Hmm. All right.

Prateek Joshi (32:00.708)
Yeah, that's a very interesting, very interesting point. All right, next question. What separates great products from the merely good ones?

Malte Kosub (32:10.446)
Creating a nice demo is the one thing. Creating millions of AI agents globally with the highest accuracy possible is a completely different thing.

Prateek Joshi (32:26.522)
What have you changed your mind on recently?

Malte Kosub (32:30.99)
How we do research internally. think deep research has significantly changed how we approach thinking about guiding deep insights into topics and that changed the way we work in some areas significantly.

Prateek Joshi (32:48.71)
What's your wildest prediction for the next 12 months?

Malte Kosub (32:53.806)
I think there will be the first companies that just have a couple of people and drive millions of revenue, maybe even just one person. So you have a lot of those agentic systems and those agentic systems are doing the work, but you still generate a lot of revenue.

Prateek Joshi (33:14.738)
All right, final question. What's your number one advice to founders who are starting out today?

Malte Kosub (33:21.89)
And I think it's connected to the thing I said before. Redefine, redesign every process AI first and don't trust common practices in software.

Prateek Joshi (33:36.53)
Amazing. Malta, this has been a great discussion. Loved your insights. And clearly, you a company and rapid growth in the last three years has been fantastic. So thank you so much for taking the time and sharing your insights.

Malte Kosub (33:49.192)
Thanks a lot.