Synthflow Podcast

In this conversation I'm joined by Leonard Poppel and Maximillian Huber of malmachen for a chat about everything involved in building a voice AI agency. Leonard has been in the voice AI space for a long time and with Synthflow they're now building out their Agency solution to offer it to clients inside of Germany. In this conversation, we dive into everything involved: how they're structuring their offers, how they're building and deploying their assistants, and what they're moving on to next. Enjoy.

Check out malmachen here - https://malmachen.com/

SIGN UP FOR SYNTHFLOW

→ Website: https://synthflow.ai/
→ Become an Affiliate: https://synthflow.tolt.io/

LEARN MORE

→ Blog: https://blog.synthflow.ai/
→ Pricing: https://synthflow.ai/pricing
→ Docs: https://docs.synthflow.ai/
→ Twitter: https://x.com/SynthflowAI
→ LinkedIn: https://www.linkedin.com/company/synthflowai/
→ Feedback: https://feedback.synthflow.ai/
→ Brand Assets: https://synthflow.ai/brand
→ Changelog: https://feedback.synthflow.ai/changelog
→ Whitelabel: https://docs.synthflow.ai/docs/whitelabel-solution-overview
→ API Docs: https://docs.synthflow.ai/reference/getting-started-with-your-api

WE’RE HIRING

→ Careers: https://mirror-biplane-a1f.notion.site/Join-Synthflow-AI-0c714d33aa47435388a843ff59b14434

LEGAL

→ Terms of Service: https://docs.synthflow.ai/docs/terms-of-service
→ Privacy Policy: https://docs.synthflow.ai/docs/privacy-policy

Creators & Guests

Host
Tom Osman 🐦‍⬛
Growth @synthflowai
LP
Guest
Leonard Poppel
Co-founder of malmachen
MH
Guest
Maximilian Huber
malmachen

What is Synthflow Podcast?

Welcome to the Synthflow Podcast, your go-to resource for exploring the latest advancements in voice AI.

Join us as we dive into insightful discussions with industry experts, share real-world applications of AI-powered voice assistants, and uncover how cutting-edge technology is transforming communication, productivity, and customer experience.

Whether you’re a developer, entrepreneur, or tech enthusiast, this podcast will keep you at the forefront of the voice AI revolution.

Tom Osman:

Today, I'm speaking with Leonard Poppel and Maximilian Huber who are from a voice AI company called Malmakken. Now they've been building voice AI solutions for companies for a long time, so I thought it'd be good to have a chat with them to see exactly how they're doing it. Enjoy.

Maximilian Huber:

The food

Tom Osman:

We were just chatting before, so everything's going pretty well, Max. Hey?

Maximilian Huber:

Nice. Yeah. Absolutely. It's, some crazy times right now. So nice to be in voice AI.

Maximilian Huber:

I think it's the best field I can currently imagine and very exciting journey so far, for sure.

Tom Osman:

Thanks for taking the time. That's awesome. And just give a quick background on Malmachin and who you are, give the intros, and what you're doing.

Leonard Poppel:

Yeah. Sure. I can I can start? My name is Leo. I'm a cofounder of Malmachin, and, we are a voice AI agency from Germany in nearby Munich.

Leonard Poppel:

And, yeah, we've been working together now. The team was built, like, 2 years ago. We were building stuff, AI automations and voice AI systems. And now, Max joined our team around 6 to 8 months ago. And now we are working with, 5 people, and everything's going very well.

Leonard Poppel:

Lots of projects going on. And, yeah.

Maximilian Huber:

Absolutely. Yeah. For me, it's even a bit bit different. So I actually only joined the team in May this year. So a bit later, they already founded the company, but before that.

Maximilian Huber:

And I finished my degree at university, then I did an internship in IT consulting. And then only in February this year, I started to working on myself, and I started, like, as a typical AI agency, did some projects with chatbots, and then came across voice AI. And I thought, wow. This is amazing. And, like, in at this moment back in February where I first experienced some voice AI some very early voice AI callers, and I just thought, well, I'm only going to focus on this from now on.

Maximilian Huber:

And then over a friend, I met Leo and the other guys from Mymahan and then decided to join them in May this year. So, yeah, that's my a bit bit more about my background.

Tom Osman:

Nice. How did you how did you both discover voice anyway? I guess you is it traditional playing around with a bunch of tools, testing out what's cool, and voice just kinda ticked that that wow for you? Or was it different?

Leonard Poppel:

For me, it was a it's a bit more complex story. Me and Anton, one of our, teammates, we had 2 years ago while doing a master's degree, we had the idea to while ChetGPT was, coming out, the 3.5 version, that there were weren't any, like, platforms like VAPI or Sunflow. We had the idea to run ChatGPT local on a Raspberry Pi, and we had the idea with hardware initially, like, to replace call centers and everything. And, yeah, we, we went after that idea, even got some some funding for it, some small funding. But we were also stuck in the master's degree, and capacity wise, it wasn't really possible to go after that kind of project.

Leonard Poppel:

So we went after other projects built like Max, quite similar chatbots, automations. We were heavy in the AI space already. And then, yeah, we came around Sunflow and, really liked the performance and everything. And then we realized, hey, this is everything we were dreaming about, roughly said, and it gave us the opportunity to really use this platform and deliver solutions custom made to clients.

Tom Osman:

Nice. Should we if anyone's, like, watching this who's complete beginner, how do we wanna, like, explain what voice AI is in terms of, like, what you're building being related to making phone calls. I'm sure some people have seen, like, chat gpt or they've spoken to chat gpt, but let's explain kind of what it is in relation to making and taking calls.

Leonard Poppel:

Maxio?

Maximilian Huber:

Absolutely. Whoever

Tom Osman:

wants to take that one?

Maximilian Huber:

Yeah. Sure. So voice AI, it's I always explain it to people. It's like you can talk to JetGPT. That's actually basically just what it is.

Maximilian Huber:

So we've got a prompt for the assistant. We can give it a name, a personality, a knowledge base, so it just knows what's going on. If it's like an inbound assistant for a company, we simply create the knowledge base with the company and then connect it to SynthFlow, for example, create the prompt to some voice settings, choose gender, voice, and so on. And then it's connected to a phone number, of course. We use Twilio for that most of the time.

Maximilian Huber:

Yep. And then you have a number that you can call, or it can also make Outbound calls itself, and then you can talk to that thing. It also then has function calling, which is also a very, very big part, so you can connect it to make live connections to systems like CRMs, for example. And then, like, you can basically build whatever you like. For starting, it's quite simple to just create an assistant with knowledge base that can answer some basic questions.

Maximilian Huber:

But then if it's supposed to be really valuable for the company, you really have to, like, dig a bit deeper, create those connections, really find out what's most valuable, and then you can really create some amazing systems with Synflow and make.com, for example. So, yeah, that's hope that's that explains it well.

Tom Osman:

Yeah. That's good. So what would be, like, a a really simple example of a build that you do for a client? How does how does a voice AI bot do you create in SynthFlow connect into a client for just they talk about one simple use case so people can get an idea about it.

Leonard Poppel:

Yeah. For example, a lot of companies, they are running paid ads, for example, through Facebook or LinkedIn or Instagram. And clients or leads, they sign in through a lead formula, and people are struggling heavy to reach out to these leads fast because leads submit the form oftentimes in the evening where no human is working anymore. And if you're not fast enough reaching out, if your speed to lead isn't fast enough, you will lose the lead because in that time, you wait with the contact. He can see an ad from a competitor with maybe a different or better offer and, he will go to the competitor.

Leonard Poppel:

And if you have an AI, for example, who is after or behind this formula, AI knows who is calling or which person the AI is calling. The AI qualifies the lead, asks some questions like, hey, I saw you signed up through one of our one of our ads. What are your interests? Or what are you looking for? Qualifies the leads and sends the data directly into the CRM of our customer.

Leonard Poppel:

And the customer knows already. Okay, this is a qualified lead. It isn't a qualified lead. The AI is also able to book appointments, and, yeah, that's one use case, for example.

Tom Osman:

Nice. So then So Go on. Sorry, Matthew.

Maximilian Huber:

Sorry about that. So that's one. That's like what we're currently focusing on heavily because we found that to just be an amazing use case, this speed to lead because it's just so so difficult. Like, if you if you don't want to spend all your day in the CRM and then manually call or message every lead that comes in, then this is, like, a super, super valuable solution that actually really wouldn't be possible without AI without maybe hiring 5 people to do that full time. But I think one one simpler thing that we're we've just built recently is a simple FAQ bot.

Maximilian Huber:

So it's for a dance studio here in, in Bavaria close to where we are. So we just created, like, the owner just told us, yeah, I would like to, my customers to be able to call this number. And then he just gave us some info. We structured that info into a good prompt, gave it a voice and a personality, and it can then answer questions and also book appointments into their calendar, which is like a quite fairly simple build, I'd say.

Tom Osman:

Wow. That's pretty cool. Did you there was one crazy stat we just saw you give on a presentation before about the the multiplier of somebody getting a call immediately from an AI when a lead comes through within the first few minutes. Was it, like, a 100 x more likely to convert for something like that?

Leonard Poppel:

It was, 100 times x to make contact to them, like, that you reach them, and 20 21 times or 21 x to convert them into a real paying client. If you contact them after 5 minutes or in the first 5 minutes after submitting the form, and let's say study from the MIT in Massachusetts, professor Aldroyd. And, yeah, it's really crazy if you think about it. 5 minutes, that's really nothing. It's not even possible as a human.

Leonard Poppel:

Like, how can it be that fast if you don't hang around in the CRM all day? So it's you don't have any other option apart from using an AI that qualifies the lead. So, yeah, it's really, really a no brainer. And if the moment we deliver this, the sales pitch to one of our leads or clients, They know the pain. And this product also it's like regarding the background.

Leonard Poppel:

This product was built because we were facing the struggles. We were already running ads for inbound AI assistance, and we had amazing lead numbers, crazy lead stats. But we weren't able to convert them because we weren't fast enough reaching out to them. And, then I built the system. Like, I had the idea.

Leonard Poppel:

Okay. We have we just use an AI outbound caller, put them into our CRM. This worked out so well, and we realized, hey, this is a crazy product. A lot of people need this. Then we defined the product, created some pricing, run ran ads for it, and now the product literally sells itself because we have ads for this outbound color for people who are interested in this in this product.

Leonard Poppel:

The people get called by our outbound caller, and the product tells the lead why this product is really good. And the people that show up to our appointments are mad qualified and are really, really interested in buying our service. So it's really a no no brainer. The product buys itself. Yeah.

Leonard Poppel:

Sells itself.

Maximilian Huber:

Does itself.

Tom Osman:

That's a very use case. Has it been like any, like pushback with clients or putting it in, or is the soon as they skip school at first, and then when they see it working, they're blown away? Or how how does that how does that work when you send it to a new client?

Leonard Poppel:

X, maybe you.

Maximilian Huber:

Absolutely. That's just a very good question. It's it's, of course, necessary to to refine it a lot because, like, the client was gonna tell us what exactly it is that they want. And then we built the product as good as possible, and then most of the time, they still have, like, I would have to this change, maybe a different voice, this response a bit too quickly or a bit too slowly, maybe. So it's, like, it's a lot of back and forth, usually with quality and also the choice of the voices.

Maximilian Huber:

But then it's, like, we've made some very we've had some very great experiences with it. Of course, it's, like, important to mention that if you come if you literally compare it to human to humans doing the outreach, like, most likely, the human is still going to be bet be a bit better. So if the agency hires a human and then that human is only there, like, chilling in the the CRM all the time and calling all the leads manually, and if they happen to find a very, very good human that's going to call them and qualify them, like, as of right now, that's still going to be a little bit better. However, it's not that accessible. So, like, small businesses, they can't afford to hire that person.

Maximilian Huber:

There's just too much effort, and that's where our AI comes into play play. And we've even seen that over the last month, the AI has gotten so much better. And we're, of course, like, assuming that this rate of improvement is going to continue even with the real time API and all that stuff that we're seeing. So we're very sure that in the near future, AI is going to be indistinguishable from real humans. And then, like, that's going to be very interesting then.

Tom Osman:

What's what's the difference currently between one that's like why why would a human be better than AI? What are the what are the things that a human can currently do better than, like, a voice assistant with a tech we've currently got?

Maximilian Huber:

I'd say it's mostly that the human can really comprehend the situation and really also comprehend the emotions. Because as of right now, the systems we're using, they're still with so with the human the, like, the the lead talks, and then it gets transcribed to text. And then the AI gets the text, creates a response, and then the text is read with a voice model. But then we're losing the emotions. So if the human is, like, you can with his tone of voice, he's, like, not very interested or super enthusiastic.

Maximilian Huber:

The AI doesn't have this info because it's only like, the model we're using is only seeing the text and responding based on that. So I think that's quite a big big part that's still like, we humans still have an advantage right now, but we're already seeing, like, systems that can actually really get the emotions. Yeah. So and I'm sure that we're going to get to this point in the future, and then there's still going to be improvement. However, in most cases, the AI works very, very well because, like, most of the time, the lead is simply interested and in a good mood, and then, like, all the stuff he hears and gets is actually very good.

Tom Osman:

Is there some tools like Hume, etcetera, which are focused on, like, emotional understanding, like text and facial expressions, but then there's, like, trying to navigate the use cases within like the EU AI act. So now it shows a bit of a spanner in how you can use emotions in AI. Is that how are you navigating that? Is there any blockers with what you can and can't do currently when building, building agents?

Leonard Poppel:

One, one main thing is regarding cold calls. We are not able to do cold calls because, we have to say in the beginning that this this call gets, recorded. And if you're doing cold outreach and you say in the beginning, yeah, this call gets recorded, then the person will immediately will will stop the call. And, because at the moment, we are Sunflow isn't able to, like, not record the call. There is always a transcript, and that's what, yeah, regulates us a bit.

Leonard Poppel:

But, yeah, in Germany in general, we have a very, late adopter market and also very technologically adverse market. In general, people are very adverse to new stuff. And always like, yeah, I will do everything by myself. Don't leave me alone with this technology stuff, with this new stuff. But the harsh truth is that most of these people, they won't survive in the long term because people or businesses who use automation systems either way or the other, they will survive in the long term.

Leonard Poppel:

And it are oftentimes the bigger companies that are contacting us because they are in the need more for this. Also, regarding, it's hard to find a lot of good employees these days in Germany, so everything comes together into the direction. Hey. We need AI automations.

Tom Osman:

Got it. What are some specifics that you have to take into account for building voices in Germany? I know things like voices, transcriptions. What are the little, nuances that you found when building for that market?

Maximilian Huber:

That's that's a very a very interesting question. So we found out that in English, voice AI currently is a bit it's a bit easier because it has all been just works a bit better in English. For example, voices, in 11 Labs, there are so many English some very, very good English voices. With German voices, it's a bit worse, I'd say. There are some very good ones, but just not as many.

Maximilian Huber:

But there are still some very good voices. Also, transcriber. Like, transcriber is a challenge in German because in English, it's just much, much better. But we know that, also, you guys are central. You're currently working on improving the transcriber for German language, which we're also looking forward to.

Maximilian Huber:

But, like, all in all, I think we're at the, like, the best level you can be at with with German language and voice AI right now, and it's only going to become better.

Tom Osman:

Is there any, like, funny kinks with different, words that get transcribed wrongly, which then kick out like a crazy response?

Leonard Poppel:

In general, it's rather it's rather the thing that the assistant thinks that there is nothing spoken or doesn't really understand anything. Like if I answer it, for example, and it's a really clear answer, I'm having the phone here at my ear, and it's not like on speaker as a clear German answer. Like, everyone would understand it. Like, every human, the AI, the transcriber is so wrong. It's it doesn't understand any context.

Leonard Poppel:

So it's like, hey. Are you still there? Like It

Tom Osman:

just completely breaks the transcriber?

Leonard Poppel:

Completely. Like, that's the thing. There's not like it's I say something and it completely mistranscribes it that it's the context changes. Like, the context doesn't exist, like, because it's so wrong. So it's not like it's giving out some complete random stuff.

Leonard Poppel:

Yeah. Got it.

Maximilian Huber:

Transcribers definitely change, but what we find one interesting thing that I found out is that the AI model, like, even though the transcriber sometimes gets words wrong, the AI model, context. Yeah. Like, it's it's hard to name some some examples because they would all be in German. But a few times, then the model actually understands what you're saying. But when we look at the transcription, it's actually a like, an incorrect word, but the AI then still gets, oh, it must have been this word or he must have meant this, but then still gets it right, which shows, like, some very good AI capabilities.

Maximilian Huber:

And, also, if we know if we, like, we know the words that we're ex that the client is probably going to say, and as simple, we can simply add those to the to a to the transcriber. So we can put in those transcriber keywords, and that makes it much easier. So if we let the client choose between different locations, for example, then it's very, very helpful to add those locations as transcriber keywords, and that makes it makes it much better. So it's challenging, but we're finding good ways to still work with it very well.

Tom Osman:

Nice. If, if someone's looking to you building a voice agency around this now, if someone's watching it and wants to build something similar, what would you think is either, like, the biggest opportunity for new people coming in to build, and, like, what should they what should they look out for? So if they're an agency owner, they wanna come in, they wanna use Synflow and sell agents, what have you what cheat codes could you give somebody who can, like, skip forward a few steps?

Maximilian Huber:

German market and, like, special focus on the German market. No?

Tom Osman:

Any. We'll go general with this one.

Maximilian Huber:

That's a that's a a very good question. I'd say definitely be careful with regulatory regulatory compliance because, like, that's also the way I start. So I started, like, as a solo solo entrepreneur, and then I also worked in a few voice AI systems. And it's can be quite, like, what you don't what you don't want to happen is that you're on a sales call telling them all this amazing stuff, but then say, oh, we are actually not even allowed to do this, and we're not even allowed to record calls. So definitely read into the, into the regulations.

Maximilian Huber:

It also depends very depends on the country. And it's also depending on the country, whether people actually care, because I've had a few conversations in the US, and there basically, no one really cares, and they just implement all this stuff, and it's alright. In Germany, people are much much more afraid. So business owners, like, they themselves don't really care about those regulations, but they just don't want to get fined for doing something wrong. So I'd say definitely be aware of those regulatory things that you have to comply with, but it's definitely possible.

Maximilian Huber:

So you just have to know when you're allowed to record calls and when not, also where data is processed. And that's good about Synthro because data processing happens in the EU only, which is very good for us. So I'd say mainly this.

Tom Osman:

That's it. Gotcha. And, agency wise, there's a lot of different ways to price like products and offerings. How are you doing it at my mark? Because we have a few things.

Tom Osman:

Some people use that retainers, some people use usage, some people do products, say like number of assistants, etcetera. How are you putting together your offer at Mannmarken?

Leonard Poppel:

Yeah. It's a very we have different products. We have, for example, the lead caller and, let's say the inbound assistant. In general, we have a one time setup fee. And, the setup fee can range very, very heavily.

Leonard Poppel:

In the beginning, we were also doing to get knowledge and to get clients and to get, for example, testimonials. We were that's also a tip. Start out low. And do maybe a project for free. Tell the client if we if you have someone, if you know someone, you have a friend or a business owner, go to them.

Leonard Poppel:

Tell them, hey, I will build a system for you for free. If you give me a a video testimonial or some testimonial in some way because this is really massive, it can really help you out for getting new clients and maybe in the same niche. Because then you have already a product in that niche, and that's really gold worthy. Yeah. Regarding pricing, now it's a bit more high price, I would say, because, we are getting a lot of offers.

Leonard Poppel:

But in general, we have a one time setup fee, and we have, retainers plus the usage fee.

Tom Osman:

And where's, what's been, the, like, sweet spot client wise? How in terms of, like so education, when you you take a new product like voice AI to somebody and you have maybe the white label offering with SynthFlow where you can give a client a login. How are you seeing that sweet spot between giving them control and you having control? And maybe we could chat more about that, the white label platform

Maximilian Huber:

too. Absolutely. So we found the white label to be very, very useful. And, like, there are so many different types of clients. That's also one more hint that I'd give is to actually find clients that are really enthusiastic and really love those AI solutions.

Maximilian Huber:

And that's, like, the clients that we're currently getting, they come from our ads. And our ad literally says, call your leads within 5 minutes using AI. And then we get the people that are like, I really have to use AI. I really want to implement it. I don't wanna be left behind.

Maximilian Huber:

So those are some very good clients to work with because they simply they they're probably already informed, and they know what's going on. It's much harder with clients that are not like they don't really care about AI, don't really know what's going on. They're just they're just much harder. So, yeah, it's definitely very useful to get those AI enthusiastic clients, and you can definitely find them in Germany. There are maybe not as many as in other countries, but they definitely exist.

Maximilian Huber:

And with the so and those clients that really love also to experiment with AI, for them, we let them edit lots of stuff in the white label solution because, generally, the white label solution then we we create a white label account for them that they can log into, and then we can control what they can edit and what they cannot edit. So, like, if we know the client just wants the product to work and doesn't really care that much about what we do, then we basically just deactivate everything. So they don't see the prompt. They don't see the actions, the info extractors, everything. But many people really want to experiment with this stuff.

Maximilian Huber:

So we can activate the prompt for them. They can make changes to the prompt, also create different, assistance. So it really depends. But you have to be careful because what is not so nice if a client's client changes the prompt, and then suddenly it doesn't work very well anymore. So you also have to to communicate it correctly.

Leonard Poppel:

On that side, well, also one tip or insight maybe, it gives you the subaccount, or it gives you the opportunity to work with consulting clients. For example, if you have someone that reaches out to you and has already some kind of knowledge, and, then you have the the plan, the agency plan. You have the ability to give them accounts, send them the link. Hey. This is your I will give you account.

Leonard Poppel:

And then you can really guide them through everything. So you they have their own account. You can really consult them well, help them set up everything. That's also a good thing and maybe a good thing to start out with, if you have someone, to do that for you.

Tom Osman:

Yeah. Nice. One thing I think, is good with clients is is prompting and just education around prompting. How have you do you share the prompts you write with clients or do you is that kind of secret sauce, propriety knowledge for you guys? Or what do you and don't you share?

Leonard Poppel:

Next, maybe you.

Maximilian Huber:

We have we've worked on our prompts a lot because it's super important the way you prompt, and there are, like, definitely some some little secrets that are just super, super useful. Process.

Leonard Poppel:

Some

Maximilian Huber:

Definitely. I think I could like, one that's been been very good for us, I think. I can I can just just give it out? It's, like, if the assistant doesn't understand the user very well, then you don't want them to say, oh, oh, can you maybe repeat that? Or I didn't understand.

Maximilian Huber:

But what we do is we simply tell the assistant, if you don't understand the user very well, then just say, oh, I think connection was bad. Can you please repeat that? Because then it's like It's like yeah. It's like a bit bit simpler. So that's one one secret sauce, and there are, like, so many parts in our prompt that we really worked on a lot.

Maximilian Huber:

But still, with the clients we have, if they really want to see the prompt, then we let them see it. So yeah. Gotcha.

Leonard Poppel:

And do

Tom Osman:

you do you manage all your clients through SynthFlow?

Maximilian Huber:

Do you

Tom Osman:

use, any other tools in, to combine, like Make or Zapier? Are you doing additional automations? How how do you have that set up?

Leonard Poppel:

Yeah. It's a very individual. That was also, yeah, that was also one hurdle or learning, I would say. We had the idea to really productize down, like, and create a fixed product. But that's that will come, and we will we are all still working on that.

Leonard Poppel:

But everyone has a different CRM. Every a lot of people have already used make n 8 n Zapier, and they just want a kind of system or automation that you build it for them in their account. For example, they send you login data or they create a subaccount or a account for you, and you build the system on on their system. And, yeah, we as we give the client data direction. If he really has no idea, let's say he has he wants an AI solution, but he doesn't use any platforms, then we come up with the system.

Leonard Poppel:

We pitch it to them, and then we do it like that. But a lot of times, clients already have something existing, and, they just want to use that. And then you adapt to that situation. For example, if you have a new CRM client has, for example, what is it? Lead table.

Leonard Poppel:

It was one of the newest ones. We're currently haven't built a system with that. And, then you dive into it. And but in the in the end, they always work the same, and you can adapt to that really, really, really

Maximilian Huber:

Absolutely. One thing I'd like to add is that so far we've been using make.com a lot because it's quite simple. It has all the integrations. So make is really great. However, make is not so good with scheduling because if we, for example, we make an outbound call and then we don't reach the person because voice mail answers or they just decline the call, then we want to call them again because we just don't want to say, okay, didn't reach them.

Maximilian Huber:

Now they're lost. We just want to call them again. For that, we have to use scheduling. So maybe we call them 2 minutes afterwards, or we call them at 5 PM on the next day. And make.com is not so great for that because you only have the time the delay module, which is, I think, 3 minutes max.

Maximilian Huber:

And, also, if you do, like, if you then write them into a Google Sheet and then check if the current time is the same as the time you want to call them, like, that will use a lot of operations. And for that, we found n 8n to be very, very good because n 8n offers Kron. So with Kron, you can do, like, perfect scheduling. So you can do whatever scheduling you like. You can do crazy scheduling, like, trigger every 3rd day of each 2nd month or some stuff like that.

Maximilian Huber:

So and it ends, I think, really, really nice.

Tom Osman:

Nice. It's like a little, little technical hack for all the automation kings out there, and it and there's a way to go for is that what you do with, people who are submitting leads through to get the instant calling, and and over Make or Zapier?

Leonard Poppel:

No. At the moment, we are still using Make for the instant call. You can, also a bit of insight, but, we can, we have the instant form. And then in make, you can go with the so called sleep module. And, you can do, 5 minutes per module, and you can do, I think, up to 45 minutes.

Leonard Poppel:

So it's like, how many is it then? 9 sleeping modules, and then the execution or the scenario runs out. And that's if you're starting out with make. That's what you can do. Then you can also build a custom logic behind it without n a then, and that's what we currently have.

Leonard Poppel:

But, we will because our processes are very, we have lots of stuff going on at the moment. So we are currently thinking about maybe moving everything to n a then and leaving one one thing behind. I already talked to that with a client before this meeting. It's easier said than done because in the end, like I already said, you have a lot of clients who really are using make. That's just the reality.

Leonard Poppel:

And they want you to build on their platform on their account, same with Zapier.

Tom Osman:

Got it. How do you explain that to clients? So if you're trying to sell a voice AI solution, how do you sell, like, the assistant and the implementation? And how do you explain to them about, like, no code solutions like Zapier or Make? Or maybe the more nontechnical clients, say, like, a small to medium sized company who know they should be start using AI, but, maybe don't know where to start?

Maximilian Huber:

In general? Many many clients don't even really care. So if they just want a system that works, then we just build it all in our systems and, like, we can use whatever we want. We have to put it into the into the, like, write our sub, like, sub providers into the contract that we have with them, which is another regulatory thing that we had to adapt. But so there are the clients that don't really care, just just have to deliver a good system.

Maximilian Huber:

And then there are also clients who do care. Many people, like, especially the AI enthusiastic people, often want it built on their own systems. Most of them then have Make. I also worked with someone who uses Pebly, which is also nice. I still prefer Make or N8N, but it totally depends.

Maximilian Huber:

So basically just do whatever the client likes, because those automation platforms, N8N, Make, Pebly, Zapier, they all work quite similarly. So very, very flexible at the moment.

Leonard Poppel:

Also on that on that end, another another thing or pattern I realized after being in a lot of calls is that a lot of business owners at the moment, if they're not very technical, they don't have a lot of AI knowledge. They just have some pain in their business. They are they have this imagination of AI being the all around problem solver. And, they can you're in the sales calls a sales call with them, and they are telling you, yeah, we can use AI there, there, there, there, there. And then it's your job as a developer or salesperson to tell them or what we are always doing.

Leonard Poppel:

Also another insight. Hey. We start at one point. Which is the point or which is the thing that causes you the most pain at the moment? And then you have a starting point.

Leonard Poppel:

You know get to know the company. You get to know their data. You get to know their processes. And then you start with that. And then you move forward to the next process to the next process.

Leonard Poppel:

For example, you have a bigger company, bigger client who are running sales calls. And then they're also saying, yeah, we need also an inbound customer service. And you ask them, hey, okay, which problem which situation gives you more problems or headaches at the moment? Yes. The outbound calls in general.

Leonard Poppel:

Then you start with the outbound calls, do this for a couple of months, and then you iterate the process, and then you tell the client, hey, look, this is really, really going well. Maybe we should also target the inbound assistant now. And then you already have a working product with them. It's easier to convince them to do that also.

Tom Osman:

Got it. So a big bit of expectation management is needed Yeah. In selling AI solutions.

Leonard Poppel:

Definitely. I don't know how

Tom Osman:

it is.

Leonard Poppel:

In in in other countries, I can, I really don't know? But in Germany, you have a lot of people who they they see AI as this invisible dragon. And and, you really have to tell them, hey. Okay. This is possible.

Leonard Poppel:

This isn't isn't possible. And

Maximilian Huber:

Absolutely. I also recommend everyone, if you start a project and really very precisely define what you're going to do and write that into the contract. Cause what's you don't want to happen. What's happened. What has happened to us before is that you then say, oh, we're going to do this project, this and this and that.

Maximilian Huber:

And then the customer always comes up with, oh, let's still do this and this and this. But from our side, it's a lot of effort. So we're now just defining very, like in the beginning, we have a long onboarding call where we define exactly what the system is going to do. And then they if they still want an upgrade, if they want to do more than we can do that, but it's going to be another another Got it.

Tom Osman:

Yeah. That's a interesting thing about upgrades. So CDs are gonna get better over time with, like, more features available, like, better voice quality. How would you, like, approach, say, example would be OpenAI's real time voice comes out and you have a widget, which is now available that you can use as another assistant on someone's website. How would you handle that with a client?

Tom Osman:

Have you done anything like that, where they've come back and asked for something else?

Leonard Poppel:

In general, we have pricing packages for for example the lead caller. And we have 3 packages. And in the best package we have a feature called feature updates included. And every time something new pops out, the client gets immediate access. And, so when the best pricing, plan or best package, a feature update thing is included.

Leonard Poppel:

Yeah.

Tom Osman:

Like that. That feels like a good agency tip. So nice clear upsell. Have you done, anything with OpenAI's real time voice widget yet?

Maximilian Huber:

Absolutely. Yeah. So we we it's a very good question. So it's it was so amazing when it came out. We were we were very, very excited about that.

Maximilian Huber:

Then we put it on our website where we'd let people talk to it simply over the widget, and we've seen that many people like, it's surprisingly many people have had conversations with this real time OpenAI widget on our website, and the conversations were very, very good. So we've even had people on live calls then test it out and say, wow, this is amazing. However, it's currently not available over the phone yet with Simpliflo, which we're looking forward to. But real time voice, real time API from OpenEye is very good. But there are a few constraint con it's few areas where it's not so good.

Maximilian Huber:

So we found that it's currently not as clever, so to say. It's a bit more random in the way it responds. And it just feels like not as basically not as clever as, for example, GPT 4.0, which is why we're currently definitely preferring GPT 4.0 because it's better with objection handling and so on. But I'm sure that the real time API is going to improve a lot and we're going to get to that point.

Tom Osman:

What do you think you said the main difference? Because for me, it sounded like a bit more human, like straight out of the box, Whereas I think for a beginner or maybe like a more advanced user, you could get a very similar result just by tweaking a lot of settings and coming up with like a really good conversational prompt and including or excluding words and you can get really close. But for maybe a basic user just coming in and trying OpenAI real time, it's like a huge, like wow moment. That was, like, the main difference for me when I first tested it.

Maximilian Huber:

Yeah. It's definitely definitely superior also in just in the way it works that you don't have those transcriber and voice stage, which is like those other systems are just much more complicated. And for that reason, because the real time real time AI, real time API just it's just so much faster because there are not all those stages with transcriber and then voice output. So it's definitely superior. And we're looking forward to the model actually being a bit stronger, a bit more clever.

Maximilian Huber:

Because one thing that we found out is with short term memory of the real time API, it's not as great because for a few use cases, we have to get the user to say a booking code, for example. So it's going to be a, let's say, 9 digit booking code. And if the user tells that to GPT 4 o, then it'll easily remember. So he says it, and the model will just say, most of the time, say all those digits correctly. But the real time API, like, when I tested it, I think 2 weeks ago, it's just it got the first few, and then it just seemed to forget the last few digits.

Maximilian Huber:

So it wasn't good enough to be used for that use case.

Tom Osman:

Wait. Where do you think this goes next? What do you what do you most looking forward to, coming which maybe it could be a capability which isn't currently doable or and something you wish for or something that you know is coming down the pipe from either OpenAI or another provider like Claude? What are you most excited about?

Leonard Poppel:

Lots of stuff. General for us, the most urgent stuff are, 1, the voices. 2nd, the transcriber. If we can get improvements there, it would make a massive change for us.

Tom Osman:

Is this Germany specific?

Leonard Poppel:

That's for us, for our agency specific. In general, I would say, yeah, g b GPT 5, really. Yeah. The also what I don't know how it is in English, but, if we are able to dynamically catch or fetch emails through the calls without, like, errors and, complex booking codes or numbers, that will also be a massive, massive change. Yeah?

Leonard Poppel:

Or massive impact. Actually, do you?

Maximilian Huber:

Yep. So, basically, I'm, like, in the future, it's we're very sure that it's going to move to those voice based AI models that just process the voice input and then create voice output directly without the step of being text based only. So we're looking forward to having stronger models there, also with more choices in the voices that you can use. But then it's like that'll be very that'll be amazing if that's like we're one 100% sure that this is going to improve a lot in the future. And then, like, with this tech, we're definitely going to reach the point where AI or voice AI will be indistinguishable from a human talking.

Leonard Poppel:

That's also another insight. If you start right now, and that's what we are we are experiencing, you are learning. You are growing as a business owner, as a developer. You are moving into the right direction. And the same is happening with the AI.

Leonard Poppel:

Everything you get better, the AI gets better, and you just both go into the same direction moving forward. And you meet there, and it's it's just massive. And you have a great product. You have a lot of you have great knowledge about the product, and then you have massive opportunities. Like, you can do everything.

Leonard Poppel:

If you focus on one niche, for example, regarding outbound calls or inbound calls, just one thing, like small thing, Businesses are facing lots of pain at the moment, especially here in Germany. And you can do you have amazing opportunities.

Tom Osman:

Got it. And overall, like, high level current, like, current capability and then limitations when you're speaking to companies. What, would you say for voice specifically that they cannot do yet, but they can do soon? So it's gonna be they want a 100 things doing for their business. Let's say somebody wants to automate every single person who picks up a phone in their company.

Tom Osman:

What do you say will be here in like 6 months? What what we can't do yet? And is there any is there any, like, big gap in expectations or is there everything is so new and incredible that they're like, wow, they're not really fussed. They're just waiting for good stuff to come out.

Leonard Poppel:

In general, you have to communicate well. That's the first thing. You have to, like I already said, they are seeing a lot of pain points in their business. And they're also having to face to explain to to their employees that, hey, we are using AI system without them getting afraid, like, hey, I will lose my job and and everything. So you can communicate, for example, hey, you have a sales team, and they are are closing the deals.

Leonard Poppel:

But you only have 10 people in the sales department, and they are also they also need to qualify the leads. So spend or waste a lot of time manually calling them down. And you can communicate, hey, okay. If we run a AI system after every ad that qualifies, the lead pre qualifies and pre selects the leads so your real sales team, like with real humans, just has better qualified leads. They are way more efficient, and they can can close more deals.

Leonard Poppel:

That's really, really a good communication strategy, for example.

Tom Osman:

Nice. Yeah.

Maximilian Huber:

And also regarding regarding your question sorry. Feature wise, we're looking forward to we're we're expecting the AI to be able to catch emails correctly because that's actually a huge thing. So we also have tried many, many different solutions and workarounds just to be able to catch emails on the phone, which is not current currently not possible, but we're quite sure that it'll be it'll be possible in the future. So that's one thing. Then also just general latency and also transcriber quality, so just how good the AI understands you.

Maximilian Huber:

So in general, we're just expecting everything to be getting much better in the future.

Tom Osman:

Nice. Gary, we've got a few minutes left. Let's talk about everyone's favorite topic, which is tools and tech stack. So what do you what are you using to run your agency? Top to bottom, a website, CRM, caller, database.

Tom Osman:

Just run through it all.

Leonard Poppel:

Yeah. We have, we pretty much, leaned it down, I would say. We are 100% regarding some individual stuff, but really most of the time, it's inflow, for the voice part. For us internal, make, dotcom with some other stuff, but also most most of the part make a CRM, Atlassian Jira. But regarding that, we are also currently looking up, for a switch.

Leonard Poppel:

And we are running our sales funnel, and that's what's really maybe also another insight. Just we had the idea at the beginning. We have to be very diverse. And here, a lot of channels. But now we just run paid ads, like, with a fixed budget, and we really niched it down or made it more lean.

Leonard Poppel:

And this was really a game changer for us regarding efficiency and leads leads getting in. We really have good data on that. And, Slack, we're using Slack to communicate, Jira for the CRM, Facebook leads, Sunflow make basic basic website. Doesn't have to be that complicated. And, yeah, I think, Max, maybe you also.

Maximilian Huber:

No. I think that's it. Also Google Workspace with Drive and, email, which works very well for us.

Leonard Poppel:

Also, what also also could be

Maximilian Huber:

Yep. I think those are the most most

Tom Osman:

How do you about testing, like testing your agents? When you're deploying one for a new business, do you have a, like a set list of things that you wanna test for? Are they is there jail breaks that you try? So like a, let's just say, like a jail break is when someone manages to break the assistant, and then it talks about, they're baking cakes or something when it should be qualifying a lead. Like, how do you, how do you go about testing those guidelines?

Maximilian Huber:

Yep. So that's, like that's a big part of the prompt because our prompt, if you would look into it, it's like lots like, a lot of big percentage of the prompt is simply about staying in the staying in the poor persona. Only talk about what you can talk about. Don't give any wrong informations information, like, really only talk about the the stuff that you were you were explicitly given in the prompt or over the, over the, actions, like CRM calls, etcetera. And we found some guidelines that work very well for that.

Maximilian Huber:

We've done lots of testing. So, like, one of our cofounders once managed to jailbreak it. So it's basically sold a Porsche 911 for €1, which is not perfect, but we looked also looked into into it from a legal perspective, and, like, that's not a that's not a proper contract that has been made. So it's alright. We're trying the jailbreaking stuff.

Maximilian Huber:

We're trying it a lot, but our current guidelines work very well for that. So it's very, very hard to jailbreak it to get it to talk about anything. And also with testing, so we're doing lots of testing internally. And what, I do for every client I work with is I first create the assistant and then tell the client, okay. Please test it.

Maximilian Huber:

Like, test it a lot. Test it yourself or with your friends or customers. And then I create in Google Sheets, I create a feedback sheet where they can enter the the time stamp of the call they've had and then leave detailed feedback. So I've got one column for technical feedback, so latency or did the agent stop or did it not understand something, and then another column for conversation flow or conversation quality, meaning that it may be not like, the words used, was it not perfect? Did it didn't it that maybe not give the optimal response?

Maximilian Huber:

And then, like, there's a lot of back and forth. And then I always look at this. So, typically, the client makes 5 to 6 calls, leaves detailed feedback, then I look at it, improve it, then he makes more calls, leaves more feedback. And then after maybe 4 to 5 feedback loops, it gets to a point where it works very well.

Leonard Poppel:

This is really a game changer game changer, I would say, also for others as a tip. Take them take the clients with you on that development journey because then they're with you. They know you are doing something. They can really be part of the creating process, and you don't have the pain of delivering the product without then testing it after 6 weeks, for example. And then after you launch it, like, hey, this, this is not working.

Leonard Poppel:

This is not good. I want this and that. So you really it's just a win win for everyone for you regarding efficiency and closing and finishing the project on time, and also for the client to have a more stable product product and be part of the journey more.

Maximilian Huber:

Absolutely. And one more hint to people that are probably watching and thinking about starting this business themselves. It's important to not underestimate the project, because it might look like it's actually not that hard, and you just create an assistant with knowledge base, but there's so much that goes into it. So many feedback loops, still many things that can go wrong. So don't underestimate the time it's going to take, and then also don't underprice it.

Maximilian Huber:

Don't sell under your value because you don't want to do a cheap project and then work on on it for weeks. So be Yeah. A bit careful with that.

Tom Osman:

Nice. Do you think it's a good place to end? So you guys are a Synthrogl partner, I think. Is that right? Yes.

Tom Osman:

How can anyone get in touch with you if they see this video and want to work with you guys?

Leonard Poppel:

Yeah. Just just on our website, and, maybe Tommy will leave email address. I think, Max Max is frozen. And, just leave our email address. And, yeah, we're open open for projects.

Leonard Poppel:

Yeah. I think we are in Germany, really one of the first needle movers, I would say, and, we have a lot of knowledge already. And, so we can help a lot of people with that. And, yeah, I hope, the person listening got some good insights. I think it's really, really learning by doing a lot of stuff.

Leonard Poppel:

You can create assistance really, really fast. But the difference between having an assistant that can be hosted on the UI and just talk with you and really having a a valuable product that is really customer ready is still is still big big of a gap. But regarding, other stuff, just go for it. Massive opportunities at the moment. We have we are already at that point.

Leonard Poppel:

Like, in 5 months, we had we were at a different place. Right now, we have so much people contacting us for solutions that is already about scaling up, like hiring someone and, looking out how to scale, like, what are the tasks we can give to someone else so you can really make progress in, like, 1 year. You can really change, change everything your whole life, basically.

Tom Osman:

Sounds good. Perfect way to stop. You know, thanks for your time. Makes a minute when you watch this back. Thanks for your time too.

Tom Osman:

I think you got disconnected. But, yeah. Yeah. I'll talk.

Leonard Poppel:

Reach out

Tom Osman:

to these guys. They do a great job building voice systems, and, yeah, catch you soon. Yeah. Thank you, Tom.

Leonard Poppel:

Bye bye.

Maximilian Huber:

Thank you. Bye.