Introduction

Why Voice AI Is the Next Automation Frontier

This session is what I want to discuss is what I believe is the next frontier of automations which is voice AI and if you think about it I mean communication voice communication is essentially what we've been doing like forever since homo sapiens walked on earth right right?

Voice is really 80 % of what we do on a daily basis. And yet, we've never managed to get to that technological advancement.

So that voice someday will be automated. But

today I wanted to talk about this, and specifically how voice AI is changing the way, especially and the business side how they operate okay so let's jump on to um the before i do that

Speaker Background

and just so you know who you're listening to so i'm jean -michel i go by jm because it's easier for everyone i've been in the enterprise sas space for about 15 years building product in in healthcare and the pharmaceutical industry, mostly in EU and the US.

My last role was the head of AI products for my organizations. The product that I built was pharmaceutical CRM, so Think Salesforce for the pharmaceutical industry, and healthcare data analytics platform. So this is what I've been doing over the past 15 years.

On the product management side, I've done pretty much everything that you can think of, 0 to 1, where I was in a product where there's no customers, no users, and we took it to 80 % market share after 15 years. So I've been blessed and lucky to be part of that journey for the organizations.

What RiseFlow Does

And today, I'm the co -founder of RiseFlow. And essentially what we do is we help our customers accelerate lead conversions, increase revenue, but also increase customer experience for them. We target mostly SMBs but we have one or two enterprise customers as well. So that's what we do.

Agenda

Alright so the agenda pretty straightforward. I'm already looking at Alp with his big eyes, hey you have 15 -20 minutes so I try to go super fast here but I try to keep it simple.

Essentially I want to quickly jump into the paradigm shift of what I believe and I guess some Some other experts believe that where voice AI is going.

Then I want to talk about the real -world challenges that we've had in organizations on when we implemented voice AI, we cut our teeth, and I should even say we broke some of them, and why we try to do so, and also how we build the platform. And hopefully you can have some key takeaways on the ugly part here, and then the wrap -up.

Demo: AI-to-AI Voice Interaction

right so that's it on uh on that side but before i do so i wanted to share with you a video leonardo hotel how can i help you today hi there i'm an ai agent calling on behalf of boris starkov he's looking for a hotel for his wedding is your hotel available for weddings

oh hello there i'm actually an ai assistant too what a pleasant surprise before we continue would Would you like to switch to jibber link mode for more efficient communication?

Okay, it gets boring at the end, I know. But you get the idea here.

So first of all, I just wanted to ask, who's watched that video in the past? You did? So about maybe 20 % of the audience here. Okay, great.

Why This Moment Matters

I just want to unpack what we've just seen here. Just behind us, behind just two devices talking to each other. Every time I see that, it gives me chill. And there's a reason for that.

It's the very first time that two machines talk to each other in real life, like real physical environment. I'm not talking about backend APIs and sending email, automating that. I'm talking like you and I talking in a room, understanding each other.

that's obviously it's in it's in lab right it's not really for scale yet it's not production ready but this is where we are going right and it's just the beginning so that's the first thing

the second thing i wanted to say is when those ai agents let's call them found that that they were just AI agents. They just switched to Jibberlink, like open communication protocol, just to go faster.

No English language to communicate between each other because we have to say hi, hello. They don't care.

They just go, hotel booking, that's what we need to do. Let's go straight to the point. Let's do it. And they will get it done super fast.

Well, as opposed to hi, how's your family? How's your friend? what's the weather like they won't have any of that right so two things here the

Two Breakthroughs Highlighted

first one is talking to each other in real physical environment right again it's a controlled live environment and then the second one is they can just communicate faster and they will at some point and not just back in NPI just

real -life communications so those are two things here obviously this is five

10 years I mean who knows every time we say five ten years now it's like five months so I'm not here to predict and I won't able to do so when it's the

Timeline And Industry Trajectory

timeframe we're gonna see that but I can see already by talking to customers and seeing what's going on around this industry right now that we are going there right so that is X months years from now now I wanted to take you back

Where We Are Today

into today and where we are so this is the classic Gartner hype cycle framework I mean for those of you who do not know essentially it's like an innovation cycle and how the society is perceived perceiving these innovations when it hits when it first hits the market right so you have the classic innovation

trigger and then everybody says you can do everything with that and then it's it's wrong, you cannot do it, so you go back, and then it's where you will hit the value when you're at that slope of enlightenment here.

From Hype to Real Value

So experts are saying the Argentic AI, voice AI, is around that tipping point here. A lot of hype, we all know that, but I'm sure you also experience over the past 12, 18 months a lot of value delivered to SMBs, to enterprise, like real value, not just fluff and hype.

And the exciting part for me and my organizations is we're just at the beginning of building what's possible with voice AI.

Early-Days Analogy: Brute Force to Platforms

And what I mean by this is today is still brute brute forcing our way into making things work. And we were discussing that it's still brute forcing our way,

I was having a discussion earlier with you, what's your name again, sorry? Mark, yeah, with Mark.

almost like cloud computing in early 2000 or social media in early 2010. But these days, or spinning up a website in early 2000 was super hard.

Now it's like, in one minute you have it done. done and here's the same thing with very early it's the time to really understand this market you know to get into customers building right now because you are the expert and it's still not super simple to do it and we will get to that very soon um on the next slide but

the key message here is i hope hopefully by by the end of the talk you'll see that there's a lot of excitement there's a lot to be built on the voice ai front and i hope that we can discuss with some of you um after the after the talk um what are your ideas because i'm always interested to hear um what's up with uh with everyone out there okay so that is that's it on that slide

Case Study: Voice AI for a Staffing Agency

now i just want to switch gear because we move from like long term where the industry is going and then to the where we are today in November 2025 to real case study so this

Patterns Observed During Implementation

is one of our customers what I wanted to do here is to put together some patterns that we found while implementing and developing our platform for our customers and I will touch on those four obviously there's a lot more but I try to compress them together in the 15 -20 minute talk that we have today.

So we're going to use the case study for a staffing agency.

It's a mid -size, like a 30 million revenue.

Challenge 1: Integration Fatigue

The first one I wanted to discuss is integration fatigue. And one of the promises for AI is having the context, right? This is super, super important.

We are promising our customers, yeah, it would augment or replace, augment, right, X staff. And for that, we're promising that your AI agents will understand the context of your business.

The challenge that we faced with many of our customers, especially this one, is that their architecture is very chaotic. Very, very chaotic.

We have customers with two CRMs. And then you wonder, I mean, why would you have two CRMs? But yes, we have found out customers with two CRMs. And I'm talking about enterprise here, like smaller customers with two CRMs.

Some of them have six databases. Again, we're like, okay, which one is the master? Now there's no master. They all do different things. Okay, thank you.

And some of them have CRMs, but with very outdated apis like no modern apis at all so very difficult for us to pull that information for our rag for instance so what the challenge we are facing is latency

when we first tested our solution with them it was around seven eight seconds so very embarrassing right we didn't share that with the customer it was just for us when we tested it right but But obviously, we have to find another solution here.

Our Solution: Modular Middleware and Connectors

And what we did was building a modular middleware in between. and so we build some specific CRM connectors and middleware here and that really help us build this data exchange layer where our rack can just pull from

And we started to build also some CRM connectors for the leaders for that specific industry. We cannot build CRM connectors for every CRM on earth, obviously. But we had to be very, we have to say no to some customers because we just couldn't do it. We have to focus, right?

there okay so that's the first thing that we noticed when we implementing with customers and the reason I say that is because as we dig deeper customer customer gets tired. Like, okay, I thought you were done.

Yeah, but you know, your 10 -year -old database doesn't really work the way we want it. Yeah, but I don't care. And they're right. I mean, they're right. They shouldn't care. That's our job. And integration fatigue is a real thing.

So now what we try to do is to better understand quickly the architecture, and we try to reuse use some of the middleware that we built and connectors for other customers and so we managed to reduce the latency from eight seven seconds to one seconds one second so it's really really good

Challenge 2: Quality vs. Scale Trade-off

and for new customers now we're doing less custom work so that's the first part the quality versus scale trade -off so that is a very interesting one here um so i was a product manager my entire entire career. So quality versus scale is just what I eat for breakfast, right?

the thing is, that was a cold shower for that one. Essentially, we were about to do a campaign for 36 ,000 calls from the entire database, or most of our customers.

We tested with a a sample of 1 ,000, and what we were hoping to achieve was an 80 % completion rate. That was kind of an eval for us, our QA, if you like, does the AI agent achieve that result? That's how we QA our voice AI agent, if you like, and it was around 70%, so we were happy

where we tested at, when we went live with 36 ,000 calls, cold shower is where we really cut our teeth. It was dropped to 32%. So very, very embarrassing for us.

I mean, the customer was playing ball and we gave him a heads up, thank God, and great collaborations here, but obviously the whole thing broke. and there's a few reasons for that and I wanted to touch on that first of all

candidates were it's a staffing agency right so candidates were spread across Europe in the US and we didn't think too much about the pacing of the call like before lunch after lunch we were aware of the time zone obviously but we wouldn't didn't go deep into the pacing like when should I follow up not within

Our Solution: Orchestration, Cadences, and Synthetic Testing

in the four hours window and so on so what we did was a bit we built a orchestration layer now for calls that build cadences right what we did as well is we went deeper on the local side so

in west coast candidate california for instance they would wake up much earlier um they work harder not harder but they will finish later so we have to find a way based on our data that we captured with our metadata we managed to find a perfect spot to call them and when and when to

follow up we have issues of voicemail too how to pick that up and there's some countries will be okay with voicemail some countries will not be we have these challenges as well one of the things

that we did to go faster on our eval is we build bots like voice ai bots calling our bots to pressure test our script and then feeding the results to our customers that was also a massive win for us because we were able to scale our testing and with that after three months of hard

labor we managed to increase to 68 we kept the customer and now we're around 80 percent of of coefficient rate.

but we made a mistake of thinking we can scale 1 ,000 to 36 and now 36 to 50, that orchestration layer needs to change every time, right? And that's why Mark is not that simple

to just build a voice AI agent for five minutes, but it won't be production ready for that specific use case. And this one we're talking about, the staffing agency.

All right, so that's the second one. Please keep me honest on the time. I'll go faster.

Say it again. Two more minutes. My God, okay. Okay, I'll go super fast.

Challenge 3: Context Drift on Long Calls

All right, so the third one here is the context drift on long calls. I'm not joking.

Some candidates have answered questions on a 15 minute straight talk, like no interruptions. They would tell how good they are for 15 minutes straight.

and we didn't expect that if you were a human being you would stop it hey thank you we get that you're really good with javascript whatever but can we move on what's your salary expectations now you know but ai was just listening for 15 minutes straight the challenge that we have here is on

long calls um our robots ai robots will get drunk they will start saying things like doesn't make sense at all and when we listen to our script we're like oh my god we really messed up here

Our Solution: Two-Tier Memory and Auto-Summary Checkpoints

so what we did is we built a two memory systems uh short term and long term short term everything that is within the conversations name what's the weather like all the you know things that is specific to the conversations long term everything that could be used across calls and across job descriptions and companies so we build the two layers that really helps

reduce the hallucinations and the last thing that we did that helped a lot as well is the autosummary checkpoints that means every three four minutes our AI agents will just stop the conversation and say hey are you saying that you're good in this and that's what you're doing yes okay and it saves that

autosummary checkpoints in the short -term memory so if we can all so that AI can always reuse that short -term memory even if it's a 20 -minute conversations when we try to qualify the candidates that helps a lot on our

hallucination we managed to reduce well even more now that was 72 % at the time after the three months hyper care for that customer but now we're like compared to the beginning we're like almost 85 % I believe so a lot of work here has been done but it's been super helpful and valuable for our customers

Challenge 4: Compliance and Spam Risk

here I'm gonna go super fast compliance spam risk everybody knows what it is we

We made a mistake to just go all out on using a lot of numbers all at once, even though we thought we used best practices. What we found out is we got flagged, like 20 % of our calls got flagged as spam and never reached the candidates.

So we have to, our solution was to rotate the numbers with some specific verified pools.

rules.

We also find out that in some states in the US, you have to disclose that you are AI in a certain manner. We knew the legislations, we did that, but not properly.

So now we are fully compliant, but obviously we cut our teeth as well on the compliance side.

There's a lot I can say here.

I can see Alp here moving closer to me. He's a big guy, so I really I really don't want to face him face -to -face too close.

Key Lessons and Takeaways

But quickly here, niche down is the value. We made a mistake to be too broad at the beginning. We need to niche down and understand our users.

On a voice AI space, you really have to understand the pain points trying to solve. Otherwise, you won't solve it.

Outcomes over AI talks, that's self -explanatory.

Lastly, for voice AI,

eval is so important.

Don't believe that the voice AI will do the whole thing like a human being he won't at all you have to capture every single point so

metadata is really really important and as we said earlier with mark i believe is the real mode here

Data, Evaluation, and Building a Defensible Moat

Conclusion

mode the more data you collect the more you will be able to deliver value to our customers and build that defensive mode against the other competitors so i'll stop here