Latest / Deep Tech Germany – AI, Robotics & Frontier Innovation by Startuprad.io™ / Voice AI Agents at Scale: Synthflow's Memory-Enabled Systems

Transcript

Voice AI Agents at Scale: Synthflow's Memory-Enabled Systems · Deep Tech Germany – AI, Robotics & Frontier Innovation by Startuprad.io™

• WEBVTT 1 00:00:00.160 --> 00:00:03.960 If you run a contact center or business process outsourcing, 2 00:00:03.960 --> 00:00:07.360 you know the pain. Customers repeat themselves, average 3 00:00:07.360 --> 00:00:11.120 handling time aht creeps up and your 4 00:00:11.360 --> 00:00:14.800 AI can't remember anything. Today's guest 5 00:00:15.280 --> 00:00:18.560 hey Cop Astabastian, CEO of 6 00:00:18.560 --> 00:00:22.040 synthflow, the Berlin voice AI platform backed by 7 00:00:22.040 --> 00:00:25.680 Excel, just launched Memory for voice agents 8 00:00:25.920 --> 00:00:29.390 so AI can keep contacts across 9 00:00:29.470 --> 00:00:33.270 calls, add sub second latency inside 10 00:00:33.270 --> 00:00:36.670 your existing CRM telephony stack. 11 00:00:36.910 --> 00:00:40.550 In the next few minutes you'll unpack how memory 12 00:00:40.550 --> 00:00:43.790 turns conversational AI into enterprise infrastructure. 13 00:00:44.350 --> 00:00:47.790 So use costs, lift FCR 14 00:00:47.950 --> 00:00:51.070 and stop making customers repeat their story. 15 00:00:57.150 --> 00:01:00.250 Welcome to Startup Rad or IO, 16 00:01:00.810 --> 00:01:04.330 your podcast and YouTube blog covering the German 17 00:01:04.330 --> 00:01:08.010 startup scene with news interviews and 18 00:01:08.010 --> 00:01:09.290 live events. 19 00:01:11.770 --> 00:01:15.570 Our guest today is Hakob Astabatsyan, co founder 20 00:01:15.570 --> 00:01:19.330 and CEO of Synthflow, the enterprise voice AI platform 21 00:01:19.330 --> 00:01:22.970 out of Berlin that's scaling like few others. The 22 00:01:22.970 --> 00:01:26.630 company powers human like phone conversation with for 1000 23 00:01:26.630 --> 00:01:30.110 plus customers, serves 100 enterprises, 24 00:01:30.510 --> 00:01:34.310 integrates with 200 tools from telephony to 25 00:01:34.310 --> 00:01:37.470 CRM and runs at sub second 26 00:01:37.630 --> 00:01:41.150 latency with 99.9% availability. 27 00:01:41.790 --> 00:01:44.630 In 2025, Synthfloww raised US$20 million 28 00:01:45.430 --> 00:01:48.990 Series A LED by Excel, cementing its position 29 00:01:49.070 --> 00:01:52.270 among the world's top AI agent platforms, 30 00:01:52.510 --> 00:01:55.550 recognized by G2 for customer satisfaction, 31 00:01:56.160 --> 00:01:59.800 fast implementation and best ROE return 32 00:01:59.800 --> 00:02:03.400 on investment. Today we dive into Synthflow's newest 33 00:02:03.400 --> 00:02:07.160 capability memory. The VAL framework lets voice 34 00:02:07.160 --> 00:02:10.800 agents remember details, preferences and 35 00:02:10.800 --> 00:02:14.600 unresolved cases across sessions, support seamless 36 00:02:14.600 --> 00:02:18.040 human handoff and even share state via 37 00:02:18.040 --> 00:02:21.800 memory groups across multiple agents. If you tried 38 00:02:21.800 --> 00:02:25.340 AI in your call center and hit the context wall, this 39 00:02:25.340 --> 00:02:28.700 episode is for you. Hey Cop will take us from 40 00:02:28.700 --> 00:02:31.740 pilots to global rollout. How to deploy, 41 00:02:32.140 --> 00:02:35.860 measure and govern memory enabled voice 42 00:02:35.860 --> 00:02:39.620 AI without adding risk. Hey Cobb, welcome to Startup 43 00:02:39.620 --> 00:02:43.020 radio. Hey John, thanks for having me here. 44 00:02:43.180 --> 00:02:46.700 Glad to be here. Before we get started, 45 00:02:47.100 --> 00:02:50.060 for a little bit older people, I was wondering, 46 00:02:51.140 --> 00:02:53.780 you remember Space Odyssey where 47 00:02:54.900 --> 00:02:58.620 the AI hell goes crazy and tries 48 00:02:58.620 --> 00:03:02.060 to kill all the people? Did you have at any 49 00:03:02.060 --> 00:03:05.820 moment the intention also to have a hellboat in 50 00:03:05.820 --> 00:03:08.900 your voice agents? Yeah, 51 00:03:09.540 --> 00:03:13.300 it's a good question actually. When we started building this in the early 52 00:03:13.380 --> 00:03:17.220 2023, everyone was talking about AGI, 53 00:03:17.300 --> 00:03:21.100 right? The first version of GPT came and everyone was saying oh now 54 00:03:21.100 --> 00:03:24.640 you're just going to click one button. AI is going to do everything. And 55 00:03:24.640 --> 00:03:28.440 then everyone, you know there was like scandals everywhere, like people were 56 00:03:28.440 --> 00:03:32.240 warning against AI AI. But I think in the last 57 00:03:32.240 --> 00:03:35.600 two years everyone came to this realization, oh 58 00:03:36.000 --> 00:03:39.760 actually the LLMs are not that smart as we 59 00:03:39.760 --> 00:03:43.120 thought, right? I mean it's an incredible progress what we had, 60 00:03:43.360 --> 00:03:46.960 right? And, but, but I think everyone realizes there is 61 00:03:46.960 --> 00:03:50.780 also as much as this technology can do and we are actually 62 00:03:51.100 --> 00:03:54.700 very far away from that AGI. So I think the 63 00:03:54.700 --> 00:03:58.540 expectations are very way lowered right 64 00:03:58.540 --> 00:04:02.300 now. And we're entering into this phase of 65 00:04:03.660 --> 00:04:07.380 soberness right after the, after the, after this excitement 66 00:04:07.380 --> 00:04:11.140 and euphoria, I think people are becoming sober in terms of being 67 00:04:11.140 --> 00:04:14.860 realistic. What can actually this AI do? And the same 68 00:04:14.860 --> 00:04:18.650 applies to the voice AI, right? I also realized 69 00:04:18.650 --> 00:04:22.330 that inside fiction there have been always the bad, the 70 00:04:22.330 --> 00:04:26.130 mean AIs think about Skynet, but I 71 00:04:26.130 --> 00:04:29.530 do believe simply for the reason they make way better stories, 72 00:04:30.970 --> 00:04:34.770 different topics. So let's take us to 73 00:04:34.770 --> 00:04:38.370 the start. Like not necessarily when you 74 00:04:38.370 --> 00:04:42.130 founded the company, but the original Spark. And when 75 00:04:42.130 --> 00:04:45.720 did, like what was the moment, you 76 00:04:45.800 --> 00:04:49.440 may remember when you conceived the 77 00:04:49.440 --> 00:04:53.160 original Spark for Synthfloww and when did it click 78 00:04:53.560 --> 00:04:56.720 that voice AI agents needed to be 79 00:04:56.720 --> 00:05:00.520 infrastructure, not just smart stuff? Yeah, very 80 00:05:00.520 --> 00:05:04.120 good question. Actually. We started 81 00:05:04.360 --> 00:05:07.932 experimenting with LLMs around like March 82 00:05:08.068 --> 00:05:11.650 2023. This is when the first, 83 00:05:11.650 --> 00:05:15.370 I would say proper version of GPT was released with 84 00:05:15.370 --> 00:05:18.970 an API that you could start working with. And everyone was 85 00:05:18.970 --> 00:05:22.610 building like text bots on this LLM, basically this wrapper 86 00:05:22.610 --> 00:05:26.050 around LLM, right? And then you would have a text bot. 87 00:05:26.930 --> 00:05:30.690 And for us the first idea was that we wanted to build a no 88 00:05:30.690 --> 00:05:34.530 code layer so that business folks could interact with it, right? 89 00:05:34.530 --> 00:05:38.150 Because. Because this was like a lot of developers and in developer forums 90 00:05:38.470 --> 00:05:41.830 everyone was trying to work with the API, but we also wanted to have this 91 00:05:41.830 --> 00:05:45.430 no code layer. And we built the first version for 92 00:05:45.430 --> 00:05:49.270 Chatbot. And I think it was around summer of 2023, 93 00:05:49.510 --> 00:05:52.550 like few months after the release of GPT that 94 00:05:53.190 --> 00:05:56.870 the first ideas of this voice orchestration started 95 00:05:57.190 --> 00:06:00.950 resurfacing on Internet, right? And the orchestration 96 00:06:00.950 --> 00:06:04.480 idea is actually very simple. You take the speech of a human, you 97 00:06:04.480 --> 00:06:08.320 convert it to text, you do some work there 98 00:06:08.480 --> 00:06:12.320 and then you send it to LLM. This text LLM gives you an 99 00:06:12.320 --> 00:06:15.960 answer back in a text form. Then you take this text, 100 00:06:15.960 --> 00:06:19.800 do a couple of things and then convert it back to speech. You do 101 00:06:19.800 --> 00:06:23.440 it and you have to do it very fast so that you play 102 00:06:23.440 --> 00:06:26.960 back the answer to the human. That's where latency comes into game. 103 00:06:27.440 --> 00:06:31.240 This was the first orchestration. Everyone, not everyone, but I would say a few 104 00:06:31.240 --> 00:06:34.800 teams actually, because it's very hard thing to do. Text is easy, 105 00:06:34.880 --> 00:06:38.290 but. But voice was very hard. And we started building this 106 00:06:38.290 --> 00:06:41.490 orchestration. And I remember around end of the summer of 107 00:06:41.490 --> 00:06:45.170 2023, we had the first version that you could talk on the phone 108 00:06:45.890 --> 00:06:49.690 and it was terrible. If I look back, it was terrible. The first 109 00:06:49.690 --> 00:06:53.410 version and it would wait for two seconds to answer. 110 00:06:53.810 --> 00:06:57.010 For example, right now when we Talk, we have 400 111 00:06:57.010 --> 00:07:00.570 milliseconds of latency humans, we are perfect because the moment you 112 00:07:00.570 --> 00:07:04.210 interrupt me, I immediately process it and stop and then vice versa. 113 00:07:04.450 --> 00:07:08.010 That's why we can have such a fluent conversation together. 114 00:07:08.250 --> 00:07:11.370 But it was not the case for AI, right? It would pose for a long 115 00:07:11.370 --> 00:07:15.170 time and then when you would start speaking as a human, it would speak 116 00:07:15.170 --> 00:07:18.810 over you. Right. At some point you just hang up as a human. 117 00:07:18.970 --> 00:07:22.490 Right. This was the very first version, but it was an 118 00:07:22.570 --> 00:07:26.410 interesting revelation for us because we realized this is 119 00:07:26.410 --> 00:07:29.970 a very hard problem to solve and it's really worth it 120 00:07:29.970 --> 00:07:33.560 investing our time because the demand was very high for 121 00:07:33.560 --> 00:07:37.400 this. And then we started working on this and we released in early 122 00:07:37.400 --> 00:07:40.960 2024. So we worked six months on this. We released the first 123 00:07:40.960 --> 00:07:44.600 version, which is way worse than what it is right now. 124 00:07:44.600 --> 00:07:48.240 Right. But it was already pretty good. And 125 00:07:48.240 --> 00:07:52.040 this is where we actually started commercializing. We had our first customers 126 00:07:52.440 --> 00:07:55.880 and we started iterating with them. And since then, 127 00:07:55.880 --> 00:07:59.550 basically it, the technology has been progressing very 128 00:07:59.550 --> 00:08:03.150 fast. Right. New LLMs have been coming out and lot of 129 00:08:03.150 --> 00:08:06.630 infrastructure around this orchestration has 130 00:08:06.630 --> 00:08:10.030 developed like Deepgram, you know, speech to text and all this kind of 131 00:08:10.030 --> 00:08:13.389 solutions. And at some point, I think in 132 00:08:13.389 --> 00:08:17.030 2025, early 2025, I think the entire voice 133 00:08:17.030 --> 00:08:20.710 industry crossed the chasm of actually 134 00:08:21.190 --> 00:08:24.630 voice. Because if you asked me in the early days, I was still not sure 135 00:08:24.630 --> 00:08:28.380 if this is actually going to be a technology that humans will 136 00:08:28.380 --> 00:08:32.140 be actually adopting on mass on scale. But 137 00:08:32.220 --> 00:08:35.260 this year we passed that chasm fortunately 138 00:08:35.740 --> 00:08:39.060 and it's just being now deployed in 139 00:08:39.060 --> 00:08:42.300 production like all around the world. Yeah, 140 00:08:43.100 --> 00:08:46.500 I also experienced that from a very different 141 00:08:46.500 --> 00:08:50.260 perspective. For example, I'm member of 142 00:08:50.260 --> 00:08:54.060 a few platforms where you can book voiceover talent. And I've 143 00:08:54.060 --> 00:08:57.660 realized that the number of requests are going way 144 00:08:57.660 --> 00:09:01.400 down and a few still pop up where you basically 145 00:09:01.400 --> 00:09:05.120 hand over your voice until eternity to 146 00:09:05.120 --> 00:09:08.840 an AI, which is not necessarily attractive 147 00:09:08.920 --> 00:09:11.240 for people who make a living. From that. 148 00:09:12.600 --> 00:09:16.200 Can you paint the reality of contact 149 00:09:16.360 --> 00:09:20.120 centers, business process, outsourcers, pre memory, 150 00:09:20.680 --> 00:09:24.320 what do aht? So the average handling 151 00:09:24.320 --> 00:09:27.980 time or the first contact resolution, also called 152 00:09:27.980 --> 00:09:31.740 FRC or the CSAT customer satisfaction course 153 00:09:32.220 --> 00:09:35.980 and call deflection, typically stall without 154 00:09:35.980 --> 00:09:39.740 the persistent contact. So. How 155 00:09:40.620 --> 00:09:44.460 did it work before your memory introduction and to 156 00:09:44.460 --> 00:09:48.140 all of our audience, yes, we have tlas Three letter acronyms. 157 00:09:51.670 --> 00:09:55.110 Yeah, exactly. Yeah, it's a, it's a 158 00:09:55.110 --> 00:09:58.550 specific industry with specific acronyms. Look, 159 00:09:59.750 --> 00:10:03.390 so I think before, before LLMs, like the 160 00:10:03.390 --> 00:10:07.190 way all these conversations were handled is through IVR 161 00:10:07.190 --> 00:10:10.550 trees, right? So you would call 162 00:10:10.950 --> 00:10:14.670 like, I don't know, I used to experience that a lot in Germany. I would 163 00:10:14.670 --> 00:10:18.500 call like insurance company or whatever and they would say, if 164 00:10:18.500 --> 00:10:22.300 you want to ask question about your financial thing, press 165 00:10:22.300 --> 00:10:25.900 one. If you want to talk to customer support, press two. 166 00:10:25.900 --> 00:10:29.100 Right? So you would have to go and then you would go into wait line 167 00:10:29.100 --> 00:10:32.540 and wait forever. So it was actually a terrible experience for 168 00:10:32.540 --> 00:10:36.340 humans, but sometimes you just didn't have a choice, right? And I 169 00:10:36.340 --> 00:10:39.420 vividly remember they gave you, for example, when you 170 00:10:40.140 --> 00:10:43.740 talk to an insurance hotline, in the beginning, they gave you 171 00:10:43.740 --> 00:10:47.340 options where you had totally no clue where you fit in and you had to 172 00:10:47.340 --> 00:10:51.000 try through all the menus. Yeah, was really bad 173 00:10:51.000 --> 00:10:54.840 experience. And then it asked you, would you still like to stay 174 00:10:54.840 --> 00:10:58.440 online? And you yelled, yes. And it said, I understood. No, 175 00:10:58.440 --> 00:11:01.720 I'll hang up now. There was level of experience, right? 176 00:11:01.960 --> 00:11:05.720 Yes, exactly. And it was like a lot of NLP 177 00:11:05.720 --> 00:11:09.520 NLU technology involved there, sometimes capturing information. I mean, it 178 00:11:09.520 --> 00:11:13.360 was a very basic technology, right? And the 179 00:11:13.360 --> 00:11:17.080 human experience, it was not built for human experience. It was 180 00:11:17.080 --> 00:11:20.680 built for just doesn't matter. Your experience doesn't matter. 181 00:11:21.080 --> 00:11:24.560 But at least we will get something solved rather than never 182 00:11:24.560 --> 00:11:28.280 solved. So it was, the mantra was 183 00:11:28.280 --> 00:11:30.280 better late than never, sort of. 184 00:11:32.440 --> 00:11:36.000 But it was the state of the technology. And 185 00:11:36.000 --> 00:11:39.640 now. With this conversational AI, 186 00:11:39.880 --> 00:11:43.560 the thing is that now it's all about experience 187 00:11:43.720 --> 00:11:47.240 as well, right? Results, but also experience, right? 188 00:11:47.640 --> 00:11:51.360 It's not like, oh, I just reduced my deflection rate or I just 189 00:11:51.360 --> 00:11:55.200 increased my resolution time, but what you just mentioned, csat, 190 00:11:55.200 --> 00:11:58.600 right? The satisfaction score, this is very, very important, right? 191 00:11:59.000 --> 00:12:02.760 So when, when we started building Synthfloww, actually 192 00:12:02.760 --> 00:12:06.120 for us in the very, in the beginning it was all about 193 00:12:06.280 --> 00:12:10.120 customer experience. So this is the only thing we cared about, 194 00:12:10.280 --> 00:12:14.000 right? Like how is your experience with AI, right? 195 00:12:14.000 --> 00:12:17.780 Are you, are you having a good con? And 196 00:12:17.780 --> 00:12:21.460 that's why, you know, the entire industry tries to build this empathy into 197 00:12:21.460 --> 00:12:25.100 AI, right? Better voices. So the early feedback 198 00:12:25.100 --> 00:12:28.260 was many voices are robotic, right. And then you 199 00:12:28.740 --> 00:12:32.420 try to build more empathetic voice. Right? And all these things is all about 200 00:12:32.660 --> 00:12:36.180 human experience. Ultimately it's all about humans here, right? 201 00:12:36.260 --> 00:12:39.980 And then the second stage of the technology, when we started working 202 00:12:39.980 --> 00:12:43.720 with contact centers, of course they have to have clear, as you 203 00:12:43.720 --> 00:12:47.520 mentioned earlier, roe, return on investment. And this is where 204 00:12:47.840 --> 00:12:51.640 is the second part of the conversation? Right? When you have a conversation with 205 00:12:51.640 --> 00:12:54.720 voice AI, the first part is the experience, the voice experience. 206 00:12:55.280 --> 00:12:59.000 The second part is the rpa, right? This business process 207 00:12:59.000 --> 00:13:02.800 automation, like what, what is the work that AI does, right? Does 208 00:13:02.800 --> 00:13:06.560 it, does it answer your question? Does it route you to 209 00:13:06.560 --> 00:13:10.160 a human? Does it send you an SMS with some information? 210 00:13:10.480 --> 00:13:14.190 Whatever the job of the AI is, it has to be done right because only 211 00:13:14.190 --> 00:13:17.510 that way the contact centers can measure the success. 212 00:13:18.230 --> 00:13:21.230 And this is what, what we measure at the sink flow. For us, the most 213 00:13:21.230 --> 00:13:24.790 important KPI for our customers is that, that they get their 214 00:13:24.790 --> 00:13:28.590 results, roe but also the end, end customer, right? 215 00:13:28.590 --> 00:13:32.070 The person on the phone has to have an amazing 216 00:13:32.150 --> 00:13:35.830 great experience because the entire purpose of this technology 217 00:13:35.990 --> 00:13:39.270 is in vain if people don't have a great experience with this. 218 00:13:41.600 --> 00:13:43.840 Totally. Yeah, exactly. 219 00:13:47.120 --> 00:13:50.880 I was wondering what feedback, what signals. 220 00:13:51.040 --> 00:13:54.000 Or. Any other 221 00:13:54.720 --> 00:13:58.240 input did you receive from 222 00:13:58.320 --> 00:14:01.640 your customers that you had to 223 00:14:01.640 --> 00:14:04.800 prioritize this multi turn 224 00:14:05.410 --> 00:14:08.770 memory over adding yet another feature? Because 225 00:14:09.010 --> 00:14:12.610 looking back now, it does make sense. But if you do 226 00:14:12.610 --> 00:14:16.450 have to make the decision, let's do A, B, C, D, 227 00:14:16.450 --> 00:14:20.210 E. How did you decide? And 228 00:14:20.690 --> 00:14:24.450 decide based on what? So 229 00:14:24.690 --> 00:14:28.290 the way we build our product is always based 230 00:14:28.370 --> 00:14:32.170 on our customer feedback, right? We talk to our 231 00:14:32.170 --> 00:14:35.890 customers and we try to understand what is important for them. 232 00:14:36.290 --> 00:14:40.130 That's one. And second, we always try to understand 233 00:14:40.850 --> 00:14:44.690 is it feasible, right? Because as I mentioned earlier, right? When this 234 00:14:44.690 --> 00:14:48.290 technology came, everyone was expecting AGI, right? 235 00:14:48.450 --> 00:14:52.290 Like general intelligence and you just click one button and everything 236 00:14:52.290 --> 00:14:56.130 gets done. And then I think suddenly everyone 237 00:14:56.210 --> 00:14:59.930 woke up to the reality that the context is everything. What is 238 00:14:59.930 --> 00:15:03.770 the context, right, of the agent take like where is the data? What 239 00:15:03.770 --> 00:15:07.330 does it have to do? And this is for our customers, right? 240 00:15:08.160 --> 00:15:11.560 Being like in a B2B context, right? Basically you can think of 241 00:15:11.560 --> 00:15:14.240 Synthlow like an AI contact center, right? 242 00:15:15.120 --> 00:15:18.920 And the whole purpose of doing a work end to 243 00:15:18.920 --> 00:15:22.480 end, right? AI has to do a work end to end. What I mean by 244 00:15:22.480 --> 00:15:25.680 that? Think about if someone calls and 245 00:15:25.840 --> 00:15:26.880 wants to get. 246 00:15:29.520 --> 00:15:33.040 Some number sent, right? Or some information sent per 247 00:15:33.040 --> 00:15:36.760 sms, right? If the AI has a good conversation but doesn't 248 00:15:36.760 --> 00:15:40.540 send that SMS to the customer, then the job is not done, right? 249 00:15:40.540 --> 00:15:43.980 The action is not executed. So this is where we started 250 00:15:44.140 --> 00:15:47.420 working, right? And you need like memory and you need, you need 251 00:15:47.420 --> 00:15:51.180 context, right? For AI to be able to do 252 00:15:51.420 --> 00:15:55.180 the job and for large companies, enterprises, very often 253 00:15:55.180 --> 00:15:58.980 it needs that Synthfloww agents have to be integrated into 254 00:15:58.980 --> 00:16:02.620 their existing tech stack ecosystem, right? 255 00:16:02.780 --> 00:16:06.340 Maybe they use one telephony system Right. They have their 256 00:16:06.340 --> 00:16:10.180 own CRM and they have whatever they have that 257 00:16:10.500 --> 00:16:14.340 agent takes inflow agents have to interact with. Right. 258 00:16:14.820 --> 00:16:18.580 So that they can actually resolve customers problem end to 259 00:16:18.580 --> 00:16:22.300 end was. Was the key. Right. Otherwise you don't 260 00:16:22.300 --> 00:16:26.060 have that ROE we just mentioned. And this was the reason. They're 261 00:16:26.060 --> 00:16:29.860 just an interaction layer. They make the data from 262 00:16:29.860 --> 00:16:33.580 your customer, meaning their tech stack, their problems, 263 00:16:33.580 --> 00:16:37.220 their interaction. They make it feasible. They just a layer of 264 00:16:37.220 --> 00:16:40.920 interacting instead of sitting there trying to navigate through 265 00:16:40.920 --> 00:16:44.720 the systems. Exactly. But it also does, it also 266 00:16:44.720 --> 00:16:47.360 navigates. Right. So you can think of it like. 267 00:16:48.320 --> 00:16:51.920 Basically it's like a knowledge base. Right. Like existing knowledge 268 00:16:51.920 --> 00:16:55.480 base. Like you connect it to the agents and when the agent 269 00:16:55.480 --> 00:16:59.200 interacts with the customer, it has the information that it needs 270 00:16:59.920 --> 00:17:03.560 to answer customers questions or send information or even 271 00:17:03.560 --> 00:17:07.120 capture. Sometimes it's the other way around. Right. The agent captures an 272 00:17:07.120 --> 00:17:10.890 information from customer and sends it to the CRM. 273 00:17:11.210 --> 00:17:14.890 Right. To the existing CRM and updates, let's say the database 274 00:17:14.890 --> 00:17:18.730 with the data. Right. So it can go both way. Either it 275 00:17:19.050 --> 00:17:22.770 shares information with the customer or it gets the information from the customer 276 00:17:22.770 --> 00:17:26.410 and saves in the existing database. Right. 277 00:17:26.730 --> 00:17:30.490 And without the context. Right. All this ecosystem, the agent is 278 00:17:30.490 --> 00:17:34.130 basically it's a demo. Right. It will not be able to do an end to 279 00:17:34.130 --> 00:17:36.480 end work that humans do nowadays. Yeah. 280 00:17:38.560 --> 00:17:42.000 We, we talked about the latency that it's below 281 00:17:42.480 --> 00:17:46.320 a second sub. Second latency and like 282 00:17:46.320 --> 00:17:50.120 having really persistent sessions. Not that it's 283 00:17:50.120 --> 00:17:53.440 interrupted and it's, it's not useful anymore. 284 00:17:53.680 --> 00:17:56.960 So how, what was the. 285 00:17:59.120 --> 00:18:02.440 What was the hardest challenge there from an engineering 286 00:18:02.440 --> 00:18:05.420 standpoint to get, to get this workable? 287 00:18:06.620 --> 00:18:09.980 Yeah, it's a very good. So actually latency was the 288 00:18:10.140 --> 00:18:13.900 number one problem when Fox started 289 00:18:13.900 --> 00:18:17.620 building voice AI because by 290 00:18:17.620 --> 00:18:21.380 following that orchestration, right. Like the speech to text and text to 291 00:18:21.380 --> 00:18:22.820 speech, right. You had to do. 292 00:18:25.140 --> 00:18:28.860 Was very rudimentary. The first version that was built, 293 00:18:28.860 --> 00:18:30.870 it was very basic. Right. 294 00:18:33.190 --> 00:18:36.630 And also like everyone was building, it was like almost for demos. 295 00:18:36.790 --> 00:18:40.630 So you were just trying to optimize it for one call, etc. 296 00:18:40.870 --> 00:18:44.550 But the problem became really complicated. Right. When we started 297 00:18:45.030 --> 00:18:48.630 growing very fast. We started growing very fast. And 298 00:18:48.790 --> 00:18:52.390 imagine like processing millions of calls every month or 299 00:18:52.470 --> 00:18:56.260 hundreds of thousands, like tens of thousands per day and 300 00:18:56.660 --> 00:18:59.860 all these calls together. You have to make sure that your entire 301 00:19:00.020 --> 00:19:03.140 infrastructure. Right. Can support that. And 302 00:19:04.020 --> 00:19:07.140 there are also so many factors that can affect 303 00:19:07.140 --> 00:19:10.980 latency. Right. It could be related to telephony, it could be 304 00:19:11.140 --> 00:19:14.660 related to connectivity, it could be 305 00:19:14.660 --> 00:19:18.220 related to even how you set up the 306 00:19:18.220 --> 00:19:21.940 agent. Right. And it was basically. But we realized 307 00:19:21.940 --> 00:19:25.700 that this is the number one challenge to solve without this, like nothing 308 00:19:25.700 --> 00:19:29.060 really matters in this business business. And what we did is we 309 00:19:29.220 --> 00:19:32.940 hired engineers who are, who have been specialized 310 00:19:32.940 --> 00:19:36.580 in this area and we have a team in Synthfloww 311 00:19:36.820 --> 00:19:40.340 that's called like it's basically our infrastructure team 312 00:19:40.420 --> 00:19:44.220 that only works on these topics, right? They work on reducing 313 00:19:44.220 --> 00:19:46.340 the latency, they work on improving the 314 00:19:47.780 --> 00:19:51.100 interruptions and then some point. It's basically two 315 00:19:51.100 --> 00:19:54.940 components that affect this one. It's an empirical work, right? You 316 00:19:54.940 --> 00:19:58.540 constantly collect data, you constantly improve your system 317 00:19:58.540 --> 00:20:02.160 and infrastructure. And the second is like how you 318 00:20:02.160 --> 00:20:05.880 build your architecture, right? Which vendors do you use, which LLM do you 319 00:20:05.880 --> 00:20:09.360 use? Right? How do you send the data between A and B? 320 00:20:09.520 --> 00:20:13.240 So there's a lot of architectural decisions you have to make so 321 00:20:13.240 --> 00:20:16.720 that when you start scaling this, you don't. It doesn't break. 322 00:20:17.120 --> 00:20:20.800 Because everyone that works in voice AI knows how 323 00:20:20.800 --> 00:20:24.160 hard it is to ensure reliability when you process 324 00:20:24.240 --> 00:20:27.350 millions of calls. I always bring this example, I say it's. 325 00:20:28.460 --> 00:20:31.980 It's similar or analogical anecdotal 326 00:20:31.980 --> 00:20:35.700 example could be for folks who work at the airport and have 327 00:20:35.700 --> 00:20:39.020 to ensure that the planes land safely, right. One 328 00:20:39.900 --> 00:20:43.739 coordinators. And of course if you do this job well, 329 00:20:44.300 --> 00:20:47.900 no one talks about this, but if you do one mistake, it's a huge 330 00:20:47.900 --> 00:20:51.420 disaster, right? Everyone will. And the stakes are 331 00:20:51.660 --> 00:20:55.460 very high in our business as well, right? You can process million calls 332 00:20:55.460 --> 00:20:58.380 and it's expected. But let's say 100 calls 333 00:20:59.150 --> 00:21:02.910 break, right? Something happens suddenly people or contact 334 00:21:02.910 --> 00:21:06.430 centers panic, right? Because I don't know, there's something 335 00:21:06.430 --> 00:21:09.950 about calls that people get really angry when 336 00:21:10.350 --> 00:21:13.870 these things don't go through, right? It's not the case with text bots. 337 00:21:13.950 --> 00:21:17.790 Sometimes people get frustrated and close it and go. But there's something about 338 00:21:17.870 --> 00:21:21.150 calls that people make huge 339 00:21:22.030 --> 00:21:25.870 buzz around this. They close down operations and this. And 340 00:21:25.870 --> 00:21:29.690 that's why reliability is one is basically 341 00:21:30.250 --> 00:21:33.210 we print it out and put it on the wall at the office. For our 342 00:21:33.210 --> 00:21:36.810 engineering team in the early days, this is 343 00:21:37.290 --> 00:21:40.650 the most important thing we have to solve in the very beginning. Yeah. 344 00:21:41.130 --> 00:21:44.770 Did you ever talk to psychologists why it is 345 00:21:44.770 --> 00:21:47.850 that voice and calling is so special here? 346 00:21:49.290 --> 00:21:52.170 I haven't talked to, but I have read about 347 00:21:53.530 --> 00:21:57.160 similar kind of concept. It's like it's basically 348 00:21:57.160 --> 00:22:00.400 psychologists call it anthropomorphism, which means 349 00:22:00.720 --> 00:22:04.160 that humans basically like humans have this 350 00:22:04.160 --> 00:22:07.600 tendency to attribute human qualities to 351 00:22:07.600 --> 00:22:11.200 machines or non. Non living things, right? Also you see it like in 352 00:22:11.200 --> 00:22:14.800 children, books, etc, when animals, animals talk, robots talk. 353 00:22:14.800 --> 00:22:18.480 Robots became humans friends and there is an empathy, 354 00:22:18.480 --> 00:22:21.840 right? So there is definitely an anthropomorphism, which means 355 00:22:22.370 --> 00:22:26.050 humans in the beginning have this tendency to be creeped out 356 00:22:26.530 --> 00:22:30.210 when the technology machine behaves like human. It scares 357 00:22:30.210 --> 00:22:33.970 humans. Right? So there is a, There is an adoption. There is an 358 00:22:33.970 --> 00:22:37.690 adoption window for humans actually to get used to it. 359 00:22:37.690 --> 00:22:40.610 Right? So it has been with many technologies. 360 00:22:41.490 --> 00:22:44.930 Like for us, everything we use is normal, right? The phones and 361 00:22:44.930 --> 00:22:47.570 laptops and everything. But 362 00:22:48.810 --> 00:22:52.490 the folks that saw it for the first time in their lives, there is always 363 00:22:52.570 --> 00:22:54.170 some, I would say. 364 00:22:55.770 --> 00:22:59.290 Adoption time needed to get used to this technology. It was same thing with voice 365 00:22:59.290 --> 00:23:02.330 AI, right? People thought, this is dangerous. 366 00:23:04.330 --> 00:23:07.770 I don't know, it's weird that the AI speaks like human, 367 00:23:07.850 --> 00:23:11.330 right? But that's why I was saying this year we crossed the 368 00:23:11.330 --> 00:23:15.170 chasm, because right now there's a large amount 369 00:23:15.170 --> 00:23:18.690 of humans that are actually used to it or already had their first 370 00:23:18.690 --> 00:23:21.990 experience talking to AI. And there were so many. Like, 371 00:23:22.550 --> 00:23:26.230 I always bring this example of this movie called her, 372 00:23:26.390 --> 00:23:30.190 right? When, When Joaquin Fenix is talking to the, To, 373 00:23:30.190 --> 00:23:33.990 To. To the AI, right? All the times, basically, it's like a human. 374 00:23:34.550 --> 00:23:38.230 And, and I think, I think humans have also 375 00:23:38.230 --> 00:23:42.070 been prepared for this a little bit through, through popular media, 376 00:23:42.070 --> 00:23:45.670 et cetera. But it's still like in the early days, I remember it was 377 00:23:45.670 --> 00:23:49.230 like it was very hard to make certain type of calls. Yeah, 378 00:23:50.510 --> 00:23:54.270 those, those people who warn about something. I vividly 379 00:23:54.270 --> 00:23:57.950 remember that I once read there were doctors recommending not 380 00:23:57.950 --> 00:24:01.550 to travel by train because they thought it would be 381 00:24:01.710 --> 00:24:05.470 harmful to the human body to move faster than a horse 382 00:24:05.470 --> 00:24:08.030 ride. Yes, yes, exactly. 383 00:24:09.150 --> 00:24:12.750 Yeah, so we, we've seen that. Let's be 384 00:24:13.110 --> 00:24:16.950 a tiny bit geeky here. Talk about here. What did you try 385 00:24:17.190 --> 00:24:20.750 that didn't work? When building memory prompt only 386 00:24:20.750 --> 00:24:24.070 hacks external key value stores, native 387 00:24:24.950 --> 00:24:28.750 retrieval, augmented generation, and what 388 00:24:28.750 --> 00:24:32.070 replaced it. So. 389 00:24:33.910 --> 00:24:37.030 Look, so the very first I would say approach 390 00:24:37.750 --> 00:24:40.630 was put, put everything. So 391 00:24:41.350 --> 00:24:43.910 if we go back to the very beginning, 392 00:24:45.090 --> 00:24:48.730 two chatbots, right? So it was the, this was the rug, right? 393 00:24:48.730 --> 00:24:52.370 Like you connect a knowledge database, right? But 394 00:24:52.370 --> 00:24:56.130 in, in, in some situations, you know, in some context, I 395 00:24:56.130 --> 00:24:59.250 would say it's for, for, for. Especially in the Enterprise 396 00:24:59.330 --> 00:25:02.290 situations, each situation was different, 397 00:25:03.170 --> 00:25:07.010 right? Each situation was different, each document was different that you 398 00:25:07.010 --> 00:25:09.810 had to reference, right? So there was some 399 00:25:10.290 --> 00:25:14.040 complexity involved to make it work for each, especially for 400 00:25:14.040 --> 00:25:17.800 the Enterprise. And then I think there was a 401 00:25:17.800 --> 00:25:21.640 moment where we solved this by having the prompt, right? You have the 402 00:25:21.640 --> 00:25:25.320 prompt field and everything that agent needs to know is 403 00:25:25.320 --> 00:25:29.040 within the prompt, the open prompt way, right? And 404 00:25:29.280 --> 00:25:32.720 you just paste it there and the information is there. But of course, there is 405 00:25:32.720 --> 00:25:36.240 also a limit there, how much you can put. Because if you put too much, 406 00:25:36.480 --> 00:25:39.960 right, if the text is huge, that also increases the 407 00:25:39.960 --> 00:25:43.620 probability of hallucination, right? Basically, think about it. 408 00:25:43.620 --> 00:25:46.900 The more the AI, the more information you give to AI, 409 00:25:47.540 --> 00:25:51.140 there are more paths that the AI can go in the conversation, 410 00:25:51.380 --> 00:25:54.740 right? So it increases basically the probability 411 00:25:55.060 --> 00:25:58.820 that it can deviate or it can get confused from the main thing 412 00:25:59.060 --> 00:26:01.460 that you set the AI to do, right? 413 00:26:02.740 --> 00:26:06.580 And I would say honestly, the answer is it depends 414 00:26:06.740 --> 00:26:10.540 because right now we have that like the open prompt thing, but we 415 00:26:10.540 --> 00:26:14.160 also just released the Bell framework and one of that 416 00:26:14.160 --> 00:26:17.680 Bell framework, the first one called Build, is the Flow designer. 417 00:26:18.240 --> 00:26:21.920 So the Flow designer allows you like, it's a visual graph that 418 00:26:21.920 --> 00:26:25.760 you can build the conversational paths, right? And 419 00:26:25.760 --> 00:26:29.279 with that we want to address the cases where you really 420 00:26:29.279 --> 00:26:33.000 know, you really know what the AI needs to do, right? Imagine you just 421 00:26:33.000 --> 00:26:36.240 hired an intern, right? You don't want that intern to go 422 00:26:36.240 --> 00:26:40.080 freestyle, right? You know your business, you just sit down and explain one 423 00:26:40.080 --> 00:26:43.840 hour and say to the intern, look, this is the script, 424 00:26:43.840 --> 00:26:47.240 this is what you say in the call, right? And this is what you don't 425 00:26:47.240 --> 00:26:50.800 say in the call, right? And I want you exactly to follow this, 426 00:26:50.960 --> 00:26:54.680 right? And this is like. And you can create it. The benefit of 427 00:26:54.680 --> 00:26:58.200 this, it's very deterministic, it reduces the 428 00:26:58.200 --> 00:27:01.680 hallucinations, it puts guardrails in the place. And 429 00:27:01.840 --> 00:27:04.760 basically AI does the job that it has to do 430 00:27:05.790 --> 00:27:08.910 if. For the open prompt 431 00:27:09.070 --> 00:27:12.830 version, right? When you have, you basically just paste the 432 00:27:12.830 --> 00:27:16.150 same way you prompt GPT, right? You don't build the graphs, but you just prompt 433 00:27:16.150 --> 00:27:19.990 it for. In that this is better if you want to keep it more 434 00:27:19.990 --> 00:27:23.670 open ended the conversation. If you want to keep the conversation 435 00:27:23.670 --> 00:27:27.230 more open ended, right? You can give it some instructions. 436 00:27:27.310 --> 00:27:30.910 But you also hope that AI would freestyle a little 437 00:27:30.910 --> 00:27:34.710 bit, right? The same way you spend some time with GPT and you just wanted 438 00:27:34.710 --> 00:27:37.550 to give also creative answers, right? So I think 439 00:27:38.430 --> 00:27:42.190 having that, that's why we've built both, right? So that 440 00:27:43.150 --> 00:27:46.990 our customers can kind of have more deterministic for more 441 00:27:46.990 --> 00:27:50.750 specific cases, especially in B2B context. Determinism is 442 00:27:50.750 --> 00:27:54.550 actually very much appreciated nowadays. And of course you 443 00:27:54.550 --> 00:27:57.550 can also leave it a bit more open ended if your use case is such. 444 00:27:57.550 --> 00:28:01.370 You don't to want it to do exactly that way. And, and this 445 00:28:01.370 --> 00:28:04.970 is like the spectrum that we leave actually our customers to choose how 446 00:28:04.970 --> 00:28:08.810 they Kind of create that, that instructions and memory. 447 00:28:08.810 --> 00:28:12.170 Yeah. You just talked about leaving 448 00:28:12.330 --> 00:28:15.210 your customers choices. 449 00:28:15.930 --> 00:28:19.690 Where do they get burned by a or 450 00:28:19.690 --> 00:28:23.450 did get burned by AI voice, promising promises, 451 00:28:23.530 --> 00:28:27.210 God rays, evaluation, handoffs or security. And 452 00:28:27.210 --> 00:28:30.680 how do you de risk them? Yes, 453 00:28:30.840 --> 00:28:34.280 that's an excellent question. Right. It's basically 454 00:28:34.760 --> 00:28:38.360 for enterprises and contact centers in B2B context. 455 00:28:38.360 --> 00:28:41.880 Right. Where ROE is everything. Risk 456 00:28:41.960 --> 00:28:44.920 has been the biggest challenge. And basically 457 00:28:45.880 --> 00:28:49.640 think about early days of agentic AI is you build the agent 458 00:28:49.880 --> 00:28:53.240 and you hope it will work. You hope it will do 459 00:28:53.890 --> 00:28:57.610 what you want it to do. Right. It's, it's a hope. You just launch 460 00:28:57.610 --> 00:29:00.850 it and then suddenly you come and see, didn't do what you wanted to do 461 00:29:00.850 --> 00:29:04.410 and then you iterate with it. Right. It's very similar. When you chat with 462 00:29:04.410 --> 00:29:08.130 GPT, you ask a question, it doesn't answer and you go back and 463 00:29:08.130 --> 00:29:11.770 say dear chat GPT. My question was this and this, 464 00:29:11.770 --> 00:29:15.530 please be very right. You just, you, you push back a couple of times 465 00:29:15.530 --> 00:29:19.250 until you get exactly what you need to do. And it was very similar. Right. 466 00:29:20.200 --> 00:29:22.680 And we said look, we have to solve this. 467 00:29:26.520 --> 00:29:30.280 We can't leave it to chance these things. And 468 00:29:30.280 --> 00:29:33.960 that's where we started working on Bell Framework, which 469 00:29:33.960 --> 00:29:37.400 is like build, evaluate, launch and learn. 470 00:29:38.120 --> 00:29:41.840 The entire idea of that framework is we 471 00:29:41.840 --> 00:29:44.600 wanted to derisk enterprise deployments. Right. 472 00:29:45.710 --> 00:29:49.070 It means you can build more predictable conversational 473 00:29:49.150 --> 00:29:52.870 paths with this designer. For evaluate part 474 00:29:52.870 --> 00:29:56.590 we build simulations. It means you don't have to go and 475 00:29:56.590 --> 00:30:00.390 call real people to see if AI does what you want to do. You can 476 00:30:00.390 --> 00:30:04.110 do it with synthetic data. You can do a simulated in the same environment 477 00:30:04.430 --> 00:30:07.870 before you actually release it. So it's way faster and way 478 00:30:07.870 --> 00:30:11.470 safer. Then there is the versioning part. 479 00:30:11.710 --> 00:30:15.560 Maybe you want to have a couple of versions that you want to test 480 00:30:16.120 --> 00:30:19.560 and lastly you want to do some auto QA and analytics. 481 00:30:20.680 --> 00:30:23.920 So basically we created an environment within 482 00:30:23.920 --> 00:30:27.480 Synthflow for customers, enterprises where they can go 483 00:30:28.280 --> 00:30:31.640 and they can do all the work there in a very safe environment, 484 00:30:32.200 --> 00:30:35.640 make sure that the agent does what they actually intended to do 485 00:30:35.960 --> 00:30:39.680 and only then they release it to the public. And 486 00:30:39.680 --> 00:30:43.280 we have been working for this for many, many months and we're 487 00:30:43.280 --> 00:30:46.980 releasing right now. And that's the entire purpose, risking and 488 00:30:46.980 --> 00:30:50.180 accelerating actually the deployment of these agents. 489 00:30:50.340 --> 00:30:50.900 Yeah. 490 00:30:54.980 --> 00:30:58.500 We also going to record a Founders Vault 491 00:30:58.500 --> 00:31:02.340 episode after this one and they'll tell us about the toughest 492 00:31:02.340 --> 00:31:06.020 enterprise deployment and where you failed initially, what 493 00:31:06.020 --> 00:31:09.820 broke in production and who escalated and who you salvaged 494 00:31:09.820 --> 00:31:13.550 it. Let's talk 495 00:31:13.710 --> 00:31:17.310 about your current playbook and how to move from 496 00:31:17.310 --> 00:31:21.030 a pilot to production in weeks, team 497 00:31:21.030 --> 00:31:24.830 setup, no code versus API. And the first three 498 00:31:25.230 --> 00:31:28.830 workflows you always target. Yes. 499 00:31:29.470 --> 00:31:32.830 So look, basically the 500 00:31:32.910 --> 00:31:36.350 older technology required a lot of 501 00:31:36.510 --> 00:31:40.090 human, human touch, right? And 502 00:31:40.730 --> 00:31:44.530 like especially the NLP NLU technology, people would have 503 00:31:44.530 --> 00:31:48.330 to go understand, map out the conversation 504 00:31:48.730 --> 00:31:52.570 and then go and build it, right? With Synthfloww 505 00:31:52.570 --> 00:31:56.210 approach, we build this layer, no code layer that 506 00:31:56.210 --> 00:32:00.010 allows customers to basically follow like three, 507 00:32:00.010 --> 00:32:03.730 four steps and you actually have your agent, right? So 508 00:32:03.730 --> 00:32:07.410 integrations is a bit different thing. This is where our forward deployed 509 00:32:07.410 --> 00:32:11.130 engineers work with the customer to integrate, especially on 510 00:32:11.130 --> 00:32:14.930 the telephony side and other tech 511 00:32:14.930 --> 00:32:18.770 stack like CRMs, et cetera. We integrate with an API or with 512 00:32:18.770 --> 00:32:22.570 SIP trunking basically. But once that work done, and this 513 00:32:22.570 --> 00:32:26.250 can be done very quickly basically within few 514 00:32:26.250 --> 00:32:29.850 days or weeks, depending on the case. But the agent 515 00:32:29.850 --> 00:32:33.610 creation part is very straightforward. So we have reduced it 516 00:32:33.610 --> 00:32:37.260 to an absolute, absolute minimum and basics, right? You have 517 00:32:37.260 --> 00:32:40.980 to prompt or with a visual graph builder, you 518 00:32:40.980 --> 00:32:44.820 just have to build a conversation and then you have to choose 519 00:32:44.820 --> 00:32:48.460 an action that AI has to do. You can also build a custom action if 520 00:32:48.460 --> 00:32:52.100 you don't have that. And then of course you can 521 00:32:52.100 --> 00:32:55.700 always leverage the Bell framework for that to test 522 00:32:55.700 --> 00:32:59.220 everything, etc. Before you deploy and then you can actually go 523 00:32:59.220 --> 00:33:01.740 live. So we have simplified it. 524 00:33:03.930 --> 00:33:06.970 I'll say to the limit, right? And 525 00:33:07.450 --> 00:33:10.970 so it's very easy actually to do and we 526 00:33:10.970 --> 00:33:14.490 continuously building new things around that so that 527 00:33:14.490 --> 00:33:17.130 we take out this, I would say 528 00:33:18.170 --> 00:33:21.810 this complex part from agents, 529 00:33:21.810 --> 00:33:25.490 right? So that it becomes almost like a deterministic 530 00:33:25.490 --> 00:33:28.850 thing you do, you just go, you build it and it works, right? That has 531 00:33:28.850 --> 00:33:32.550 been the, the whole idea. And in terms of use cases, 532 00:33:33.190 --> 00:33:35.350 the most common ones or workflows. 533 00:33:36.870 --> 00:33:40.150 That we see across the board, we see a lot of 534 00:33:40.150 --> 00:33:43.750 transfer case where AI is the first line of defense 535 00:33:44.310 --> 00:33:48.150 and then it weeds out the irrelevant calls 536 00:33:48.470 --> 00:33:51.670 and forwards the important ones, which actually 537 00:33:51.750 --> 00:33:55.590 reduces the workload. With humans we see a lot of 538 00:33:56.300 --> 00:34:00.060 customer support, or in other words, I love to call it faq, right? 539 00:34:00.300 --> 00:34:03.820 AI answering a lot of questions, right? For customers, 540 00:34:04.380 --> 00:34:07.940 which are very basic, so humans don't have to waste time on that. AI can 541 00:34:07.940 --> 00:34:11.380 answer all this, it's very solid. There is also a lot of 542 00:34:11.380 --> 00:34:15.060 qualification is happening. AI reaching out to 543 00:34:15.060 --> 00:34:18.300 humans, asking couple of things and qualifying and 544 00:34:18.460 --> 00:34:21.660 booking next steps. Basically these are very common, 545 00:34:21.660 --> 00:34:24.930 widespread cases we see in our, right. B2B 546 00:34:25.250 --> 00:34:29.050 AI contact center case, mainly on the phone, 547 00:34:29.050 --> 00:34:32.530 right? Yeah. Talked about 548 00:34:32.530 --> 00:34:36.130 those KPIs. Can you share some 549 00:34:36.130 --> 00:34:39.330 specific numbers? Typical containment, the 550 00:34:39.650 --> 00:34:42.770 Reduction in average handling time, the 551 00:34:43.730 --> 00:34:47.450 lift in first call resolutions, meaning the customers 552 00:34:47.450 --> 00:34:50.930 only need to call once or human hour savings 553 00:34:51.090 --> 00:34:54.500 that you comfortable stating by 554 00:34:54.500 --> 00:34:58.340 vertical. Look, there are 555 00:34:58.340 --> 00:35:02.180 no single numbers, right? It very much depends on the 556 00:35:02.180 --> 00:35:05.980 customer and the use case, right? So for example, we 557 00:35:05.980 --> 00:35:09.660 have how complicated the case is, right? For example, we have one 558 00:35:09.660 --> 00:35:12.300 customer in healthcare and 559 00:35:13.180 --> 00:35:16.900 they basically benchmark AI with 560 00:35:16.900 --> 00:35:20.580 the work that human does today, right? In terms of deflection 561 00:35:20.580 --> 00:35:24.380 rate, right? So, and I don't remember 562 00:35:24.380 --> 00:35:27.620 what the exact number was, let's say 17 or 18%, 563 00:35:27.940 --> 00:35:31.620 whatever the number is, basically that customer 564 00:35:31.700 --> 00:35:35.420 says if AI even like 1, 2, 3 565 00:35:35.420 --> 00:35:39.140 percentage points better than this, that's already a big 566 00:35:39.140 --> 00:35:42.660 win, right? Because humans, humans time is more 567 00:35:42.660 --> 00:35:46.420 valuable than AI's time, right? AI is 24, 7, you know, 568 00:35:46.420 --> 00:35:49.960 it's cheap, you can just put it there and it can do the 569 00:35:49.960 --> 00:35:53.680 work, right? The only reason you wouldn't put AI, if 570 00:35:53.680 --> 00:35:57.320 it's significantly worse in terms of ROE than humans, 571 00:35:57.640 --> 00:36:01.400 then you still justify. But if AI is even slightly better. So 572 00:36:01.400 --> 00:36:05.080 for example, in that particular case they were targeting 20% plus 573 00:36:05.160 --> 00:36:08.800 rate, then they went with AI and the 574 00:36:08.800 --> 00:36:12.520 human basically moves one step back and takes important 575 00:36:12.600 --> 00:36:16.430 calls, calls that are a bit more complicated, maybe they are 576 00:36:16.430 --> 00:36:20.270 more of an emotional nature or there is a complicated things that they have 577 00:36:20.270 --> 00:36:23.990 to do afterwards, right? So that's what we're 578 00:36:23.990 --> 00:36:27.550 seeing, right? In other cases, for example, if someone does an outbound 579 00:36:27.550 --> 00:36:31.270 campaign and does like thousand calls, 580 00:36:31.430 --> 00:36:35.110 even if you get, I don't know, five, six calls answered, 581 00:36:35.110 --> 00:36:38.710 it could be a big win, depending what's your case, right? Maybe each 582 00:36:38.710 --> 00:36:42.210 call can result in a lead that can result in 583 00:36:42.690 --> 00:36:45.650 tens of thousands contracts for you afterwards, right? 584 00:36:46.210 --> 00:36:49.730 It's a big win. But maybe you call 585 00:36:49.730 --> 00:36:53.410 thousand times and everything less than 500 586 00:36:53.410 --> 00:36:57.129 answers is already bad, right? And then in some 587 00:36:57.129 --> 00:37:00.490 cases AI works, in some cases it doesn't make sense, right? The 588 00:37:00.490 --> 00:37:04.210 roe, the simplest things to measure are 589 00:37:04.450 --> 00:37:08.170 more like, I would say the cases that I mentioned, 590 00:37:08.170 --> 00:37:11.850 right? Let's say the transfer and et cetera, these things are very easy to 591 00:37:11.850 --> 00:37:15.250 measure. And in some cases you see 592 00:37:15.810 --> 00:37:19.650 marginal improvement, right? AI is several percentage points better. 593 00:37:19.650 --> 00:37:22.970 And it's already a good case to put AI because of the cost 594 00:37:22.970 --> 00:37:26.570 savings. But sometimes we also have customers who 595 00:37:26.570 --> 00:37:30.250 never had anyone there, so they didn't even have resources. And just 596 00:37:30.250 --> 00:37:34.050 putting AI results in 80% improvement in 597 00:37:34.620 --> 00:37:38.460 time reduction, for example, right? Like some crazy numbers, because 598 00:37:38.620 --> 00:37:41.980 they never had that, right? They didn't have the resource and now they just put 599 00:37:41.980 --> 00:37:45.660 AI and it does something that was never being Done. 600 00:37:45.980 --> 00:37:49.700 And there you see like very like large numbers as 601 00:37:49.700 --> 00:37:50.380 an improvement. 602 00:37:53.340 --> 00:37:56.660 For our audience. If you run a CX or 603 00:37:56.660 --> 00:38:00.060 bpo, what's the one place you'd want 604 00:38:00.730 --> 00:38:04.170 voice AI memory like tomorrow, Bookings, 605 00:38:04.330 --> 00:38:08.130 claims, returns, tech support. Drop it in the 606 00:38:08.130 --> 00:38:11.570 comments. By the way, if there's anything interesting, 607 00:38:11.570 --> 00:38:14.330 Jacob, of course we'll share it with you. And 608 00:38:15.210 --> 00:38:18.810 let's talk a little bit about the framework. What's your 609 00:38:18.970 --> 00:38:22.810 memory read, write policy framework, what 610 00:38:22.890 --> 00:38:26.520 to store, for how long with what consent and 611 00:38:26.520 --> 00:38:30.280 what controls. This is a very important question, right? 612 00:38:31.800 --> 00:38:35.560 We touched upon anecdotally before like why are phone 613 00:38:35.560 --> 00:38:39.240 calls so dramatic, right? Have such a dramatic impact on humans? 614 00:38:39.400 --> 00:38:42.760 And it's also same in terms of compliance, right? 615 00:38:43.400 --> 00:38:47.080 So for example, there is a HIPAA compliance in the US 616 00:38:47.160 --> 00:38:50.960 or GDPR compliance in Europe. And with a Synthfloww 617 00:38:50.960 --> 00:38:54.690 very early on in our journey, invested significant resources 618 00:38:54.690 --> 00:38:58.050 of our company actually to undergo all these compliances, right? 619 00:38:58.050 --> 00:39:01.730 SoC2, type 2, GDPR, HIPAA. And 620 00:39:01.730 --> 00:39:05.250 for example, with HIPAA compliance, right, there is the toggle 621 00:39:05.490 --> 00:39:08.690 bar in our tool that our customers 622 00:39:09.330 --> 00:39:12.850 disable so that the calls don't get recorded. Because 623 00:39:12.850 --> 00:39:16.450 certain types of calls in certain regions you're not allowed to record, 624 00:39:16.770 --> 00:39:20.530 right? So in gdpr, right, for example, the data 625 00:39:20.610 --> 00:39:23.670 hosting is a very important, important thing, right? And 626 00:39:24.870 --> 00:39:27.350 so calls are very sensitive and we always 627 00:39:28.790 --> 00:39:32.430 work very closely with our customers, right? So that we ensure we sign 628 00:39:32.430 --> 00:39:36.150 like baas with our customers, right? Business associate agreements 629 00:39:36.390 --> 00:39:39.950 we have basically right now on the Synthfloww website there is a trust vault that 630 00:39:39.950 --> 00:39:43.190 you can access and we have all sub processing agreements 631 00:39:43.190 --> 00:39:46.870 baas. We take this topic very seriously because 632 00:39:47.630 --> 00:39:51.270 in some industries, and I would say in many cases like the 633 00:39:51.270 --> 00:39:54.550 data storage and recordings is a very sensitive 634 00:39:54.550 --> 00:39:57.870 topic and it has to be 100% compliant. 635 00:39:58.590 --> 00:40:02.150 This is very important in our market. I was wondering for 636 00:40:02.150 --> 00:40:05.150 callers who get interrupted, switch or something, 637 00:40:05.630 --> 00:40:08.910 how do you maintain your context 638 00:40:08.910 --> 00:40:12.590 coherence and avoid hallucinated calls? 639 00:40:12.590 --> 00:40:16.390 For example, you call for your VW Golf and 640 00:40:16.390 --> 00:40:19.850 actually they tell you yes, you're bantling is ready. 641 00:40:23.530 --> 00:40:27.130 Yeah, that's a very good question, right? And this is 642 00:40:27.290 --> 00:40:30.970 comes back to the to the point, right? It's also like 643 00:40:30.970 --> 00:40:34.810 kind of the idea of Bell framework is we want to 644 00:40:35.050 --> 00:40:38.690 make sure, right, the customers can build 645 00:40:38.690 --> 00:40:42.490 more deterministic agents if its topic is sensitive, right? So 646 00:40:42.490 --> 00:40:46.170 for example, you can leave the AI agent a bit open 647 00:40:46.840 --> 00:40:50.640 if it has to answer like I don't know, some low stakes 648 00:40:50.640 --> 00:40:54.440 situations, let's say like hypothetically has to say 649 00:40:54.440 --> 00:40:57.640 what's the weather or something like that. Right. But 650 00:40:58.280 --> 00:41:02.080 if it's about the order or Golf and Bentley, 651 00:41:02.080 --> 00:41:05.480 right. You want to use the visual graph 652 00:41:05.480 --> 00:41:09.000 builder that we built because that way you completely 653 00:41:09.000 --> 00:41:12.680 eliminate the hallucinations. Right? Because that's what we've 654 00:41:12.680 --> 00:41:16.480 built. It's a very deterministic environment. Right. That 655 00:41:16.480 --> 00:41:19.840 doesn't allow AI to deviate from your. If you know, 656 00:41:20.320 --> 00:41:24.080 as a Volkswagen. Basically too that this 657 00:41:24.080 --> 00:41:27.680 is exactly what customer is going to ask and this is exactly 658 00:41:28.080 --> 00:41:31.440 what the AI should answer. This is 659 00:41:31.680 --> 00:41:35.320 the right way to go. Right. And then of course you just simulate this, you 660 00:41:35.320 --> 00:41:38.560 create this before actually releasing this. This is very important. 661 00:41:38.800 --> 00:41:42.360 Right. And we have invested lot of resources in building the Bell 662 00:41:42.360 --> 00:41:46.110 framework so that our customers, these enterprises can 663 00:41:46.110 --> 00:41:49.550 actually serve these use cases. And I can go one step 664 00:41:50.750 --> 00:41:54.430 deeper and tell you it's even more important in financial industry, 665 00:41:54.590 --> 00:41:58.270 for example, right. You cannot quote the wrong percentage or etc. 666 00:41:58.270 --> 00:42:02.110 Right. So you have to adhere to the things. 667 00:42:02.190 --> 00:42:05.950 And the short answer, basically you reduce the hallucinations through 668 00:42:06.190 --> 00:42:09.900 graph builder and of course the simulations and versioning in 669 00:42:09.900 --> 00:42:13.500 qa. That is great guys. 670 00:42:13.660 --> 00:42:17.380 Right after the ad break, hey cop will unpack the deployment 671 00:42:17.380 --> 00:42:21.220 blueprint. ThinFlow users to get memory enabled 672 00:42:21.220 --> 00:42:24.300 agents live in under 30 days plus one 673 00:42:24.300 --> 00:42:26.780 mistake that kills ROI. 674 00:42:32.140 --> 00:42:35.340 Guys, welcome back after our short ad break. 675 00:42:36.140 --> 00:42:39.410 Hey cop, can you walk us through through the 30 day 676 00:42:39.410 --> 00:42:42.810 deployment blueprint from data mapping and 677 00:42:42.810 --> 00:42:46.370 integration selection to evaluating metrics and 678 00:42:46.610 --> 00:42:50.330 blue green rollout. Definitely. So the way we 679 00:42:50.330 --> 00:42:53.970 work with our customers is in the very first step 680 00:42:53.970 --> 00:42:57.490 we try to understand the use case and 681 00:42:57.650 --> 00:43:01.290 the value that the customers seek to get out of this 682 00:43:01.290 --> 00:43:04.990 technology. Right. Because there is a lot of noise around AI 683 00:43:05.460 --> 00:43:08.660 and we want to make sure that our customers have the right 684 00:43:08.660 --> 00:43:12.340 expectations from this technology and we can actually 685 00:43:12.740 --> 00:43:16.260 deliver them in a promised time frame. 686 00:43:17.060 --> 00:43:20.580 Once we have done this part, we actually get 687 00:43:20.580 --> 00:43:24.260 into a. Period working 688 00:43:24.260 --> 00:43:28.100 period with customers where we help them with integrations. 689 00:43:28.260 --> 00:43:31.580 Right. We advise them on a lot of topics. Our forward 690 00:43:31.580 --> 00:43:35.200 deployed engineers stay very close to the customers and 691 00:43:35.200 --> 00:43:38.640 we advise them on a lot of things like for example on the agentic part 692 00:43:38.800 --> 00:43:42.560 we demo them the Bell framework so that they can deploy 693 00:43:42.560 --> 00:43:46.320 this in a de risked way. But also we help them 694 00:43:46.800 --> 00:43:50.479 integrating this into their existing ecosystem. Right. Especially many 695 00:43:50.479 --> 00:43:54.160 customers have their own telephony system and we 696 00:43:54.160 --> 00:43:58.000 have to integrate into that with the SIP trunking and there are a 697 00:43:58.000 --> 00:44:01.840 lot of nitty gritty details sometimes and our forward deployed engineers 698 00:44:01.840 --> 00:44:05.590 work with them to make that happen. And generally 699 00:44:05.590 --> 00:44:09.390 during this period we do these parts and the 700 00:44:09.390 --> 00:44:11.910 customers would go live and start testing. Right. 701 00:44:13.830 --> 00:44:17.270 And Then they would get their KPIs they want. Right? And 702 00:44:17.510 --> 00:44:20.790 this would be like a success. But right now with the Bell framework, 703 00:44:20.950 --> 00:44:24.550 that's not even needed to go to live. 704 00:44:24.550 --> 00:44:28.350 They can even do it like in a more controlled environment. And we 705 00:44:28.350 --> 00:44:31.040 even expect to reduce this more. Right. 706 00:44:32.800 --> 00:44:36.600 But that's generally like roughly how we work with our customers to 707 00:44:36.600 --> 00:44:39.840 make sure actually they get out of this technology what they want to. 708 00:44:41.760 --> 00:44:45.560 You've been talking about the nitty gritty details and that's usually where you 709 00:44:45.560 --> 00:44:49.040 win or lose in the startup game. Walk us through 710 00:44:49.280 --> 00:44:52.960 growth Methodology. Bell methodology. 711 00:44:52.960 --> 00:44:56.640 Bell framework. How do you scale from 1 use case to 5? 712 00:44:57.050 --> 00:45:00.890 What's the sequence that compounds return on investment? Yes. 713 00:45:01.050 --> 00:45:04.730 So I think the very first step for us 714 00:45:05.370 --> 00:45:09.210 scaling the business, the first challenge we started facing once we scaled 715 00:45:09.210 --> 00:45:13.009 was the volume, the infrastructure. Right. Basically we 716 00:45:13.009 --> 00:45:16.330 were a startup of couple of folks and then 717 00:45:17.130 --> 00:45:20.650 basically in few months we started processing millions of calls on the 718 00:45:20.650 --> 00:45:24.470 platform. And it was very unexpected. So we had to 719 00:45:24.470 --> 00:45:28.110 do a lot of infrastructure investments and expansions to be actually 720 00:45:28.110 --> 00:45:31.670 able to process this. And the way we scale 721 00:45:31.670 --> 00:45:35.350 this is we generally are very 722 00:45:35.350 --> 00:45:39.190 selective with whom we work. Right. That's why 723 00:45:39.190 --> 00:45:42.910 we do this. Step one, what I mentioned earlier is we 724 00:45:43.390 --> 00:45:47.150 always try to understand the customer's expectations from this 725 00:45:47.150 --> 00:45:50.940 technology and kind of have the same 726 00:45:50.940 --> 00:45:54.740 expectations, right. So that we can actually help customer 727 00:45:54.740 --> 00:45:58.420 to get there. But the work on our side is 728 00:45:59.060 --> 00:46:02.580 more like consultative and being as a sparring 729 00:46:02.580 --> 00:46:05.380 partner for the customer. So it doesn't require 730 00:46:06.260 --> 00:46:10.060 large headcount on our side to work with our customers. And we 731 00:46:10.060 --> 00:46:13.820 have built the product in a very scalable way and we continue doing that. Right. 732 00:46:13.820 --> 00:46:17.480 We release the Bell, but then we also are cooking a couple 733 00:46:17.480 --> 00:46:21.040 of very advanced features which I'll not talk about today, 734 00:46:21.440 --> 00:46:24.960 but we are going to release soon in the future. 735 00:46:25.840 --> 00:46:29.600 And basically we'll scale it with the product. Right. So there is no 736 00:46:29.600 --> 00:46:32.800 like an inherent, I would say limit 737 00:46:33.440 --> 00:46:36.800 to why our platform at this moment. Right. 738 00:46:36.880 --> 00:46:40.480 Cannot process hundreds of millions calls instead of millions of calls. 739 00:46:40.560 --> 00:46:44.300 Right. And, and this is how 740 00:46:44.300 --> 00:46:47.980 we want to continue at this moment. Of course we have to grow our team. 741 00:46:48.140 --> 00:46:51.540 It's natural as a company as we grow and the revenue grows and the 742 00:46:51.540 --> 00:46:55.180 usage, we have to build different departments and be 743 00:46:55.180 --> 00:46:57.900 better, also offer better service to our customers. 744 00:46:59.900 --> 00:47:03.700 But there is no, I would say a fundamental blocker to 745 00:47:03.700 --> 00:47:07.100 scale this business for us at this moment. I've seen 746 00:47:07.980 --> 00:47:11.500 some logic sprawl in such 747 00:47:11.500 --> 00:47:15.280 projects. How do you recommend no code ownership 748 00:47:15.280 --> 00:47:18.560 vs ops vs API for engineering? And. 749 00:47:21.440 --> 00:47:24.880 Would governance Prevents this logic 750 00:47:24.880 --> 00:47:28.640 sprawl. Yeah. So I would 751 00:47:28.640 --> 00:47:32.320 say of course the API gives more flexibility. 752 00:47:32.960 --> 00:47:36.800 Right. And I would say for 753 00:47:36.800 --> 00:47:40.640 smaller cases, right. So we have to, 754 00:47:40.640 --> 00:47:44.340 I would say maybe you can say like two types of customers that 755 00:47:44.340 --> 00:47:47.860 one type would prefer the no code, the other type would prefer the API. 756 00:47:48.420 --> 00:47:52.100 So let's say one of our customers, a very large contact center 757 00:47:52.980 --> 00:47:56.100 bpo, they basically use the API 758 00:47:56.500 --> 00:47:59.780 because they integrate synth flow into their 759 00:47:59.860 --> 00:48:03.700 operations, basically into their entire tech stack so 760 00:48:03.700 --> 00:48:07.460 that they can offer voice AI to their customers. Right. So 761 00:48:07.460 --> 00:48:11.100 it's like a white label integration there. And of course the 762 00:48:11.100 --> 00:48:14.940 API is the way to go there because they want to expand on 763 00:48:14.940 --> 00:48:18.660 this massively. But if there is, let's say 764 00:48:18.660 --> 00:48:22.500 for smaller customers, right. If you have one particular 765 00:48:22.580 --> 00:48:26.420 use case in mind and you want to quickly build this, it's very 766 00:48:26.420 --> 00:48:29.980 easy to go and build with the no 767 00:48:29.980 --> 00:48:33.060 code layer you have there because it's just click, click, click right 768 00:48:33.700 --> 00:48:37.380 and you have the case. But we also have 769 00:48:37.380 --> 00:48:41.100 built some sophisticated features there. It's called like multi 770 00:48:41.100 --> 00:48:43.300 tenant, basically. 771 00:48:45.620 --> 00:48:49.339 Portal where you can become the owner and 772 00:48:49.339 --> 00:48:52.940 you can have many users invite your team members 773 00:48:52.940 --> 00:48:56.420 and have sub accounts. Right. For different use cases. 774 00:48:56.980 --> 00:49:00.520 And we have many customers who just went chose this way. Right. 775 00:49:00.520 --> 00:49:03.800 Because it's simpler, it requires less engineering resources 776 00:49:04.280 --> 00:49:07.960 and it's more than enough for their particular case. Yeah. 777 00:49:09.720 --> 00:49:13.360 You do have. Some 778 00:49:13.360 --> 00:49:16.840 case studies on your website, Med Bell, SmartCat, 779 00:49:18.200 --> 00:49:21.400 what ingredients made those outcomes work? 780 00:49:22.840 --> 00:49:26.600 Yeah, so I think so. One 781 00:49:26.600 --> 00:49:30.200 important thing is that this initial period that 782 00:49:30.200 --> 00:49:33.880 we actually are close to our customers. 783 00:49:37.560 --> 00:49:40.680 We set their expectations right, we help them define the 784 00:49:41.000 --> 00:49:44.640 outcomes that they want to get. Very often customers know it 785 00:49:44.640 --> 00:49:48.360 themselves, but if they don't, they're more explorative. We help them 786 00:49:48.360 --> 00:49:51.760 understand that. And once it's done, 787 00:49:52.320 --> 00:49:55.280 then the rest is all about the right integrations 788 00:49:56.240 --> 00:50:00.000 and the right agentic part, the right agent setup. 789 00:50:00.080 --> 00:50:03.440 Right. So the agent setup we have simplified to 790 00:50:03.440 --> 00:50:07.240 minimum. But there still could be some questions. Do I 791 00:50:07.240 --> 00:50:11.000 go with the visual graph builder or do I prompt it myself? But of 792 00:50:11.000 --> 00:50:14.400 course our customer support is always there to answer customers 793 00:50:14.400 --> 00:50:17.280 questions. But the integrations part, 794 00:50:18.140 --> 00:50:21.940 this is something that often we sometimes have 795 00:50:21.940 --> 00:50:25.700 to share something like our documentation or etc so that 796 00:50:25.700 --> 00:50:29.380 they can integrate into their telephony system. Or if they 797 00:50:29.380 --> 00:50:31.580 don't have telephony system, they use us. Right. 798 00:50:33.500 --> 00:50:36.820 Or maybe they have another tool stack like a 799 00:50:36.820 --> 00:50:40.180 CRM Salesforce or something like that they need to 800 00:50:40.180 --> 00:50:43.900 integrate and these are the main components. If this is done well. 801 00:50:46.620 --> 00:50:50.420 It should start actually delivering the value. You can go to 802 00:50:50.420 --> 00:50:53.740 production. I was wondering a little bit why you talked 803 00:50:54.140 --> 00:50:57.980 what breaks at 1000 parallel calls, 804 00:50:58.300 --> 00:51:01.980 speech quality, latency spikes, rate limits, 805 00:51:01.980 --> 00:51:05.540 integration quotas. And how do you harden for 806 00:51:05.540 --> 00:51:09.100 peak season? I mean especially if 807 00:51:09.100 --> 00:51:12.220 you are into some kind of 808 00:51:13.620 --> 00:51:17.060 business that has a rush towards the end of a year 809 00:51:17.060 --> 00:51:20.860 towards Black Friday, e commerce Christmas is 810 00:51:20.860 --> 00:51:24.100 what I had in mind. How do you prepare for something like that? 811 00:51:24.420 --> 00:51:28.140 That's a very, very good question. And this was honestly at the 812 00:51:28.140 --> 00:51:31.900 very beginning this was our main problem. This was 813 00:51:31.900 --> 00:51:35.580 like when we started scaling because and thousand 814 00:51:35.580 --> 00:51:38.740 calls are fine, it's nothing. We're talking about millions. 815 00:51:39.220 --> 00:51:42.950 This is okay, totally fine with 816 00:51:42.950 --> 00:51:46.790 me. Let's talk about millions. That's where real challenges 817 00:51:46.790 --> 00:51:50.190 start happening. And honestly 818 00:51:50.990 --> 00:51:53.950 one of the first challenges we had 819 00:51:54.910 --> 00:51:58.270 was the capacity. So we had to. 820 00:52:00.110 --> 00:52:03.870 Get more and more hosting capacity 821 00:52:04.350 --> 00:52:08.180 so that we could digest all this, all this processing, 822 00:52:08.260 --> 00:52:11.980 right? And so this is one thing you have to make sure 823 00:52:11.980 --> 00:52:15.460 that your infrastructure, right, the cloud and the hosting, 824 00:52:16.580 --> 00:52:20.300 you already have that capacity. It's set for scale, right? And for 825 00:52:20.300 --> 00:52:23.860 us, because we grew so fast, we were not very much prepared in the 826 00:52:23.860 --> 00:52:27.300 beginning. We never thought we would go in few months basically from 827 00:52:27.300 --> 00:52:30.860 hundreds to hundreds, thousands or 828 00:52:30.860 --> 00:52:34.190 meeting calls. And this 829 00:52:34.510 --> 00:52:37.790 was the very first challenge, like architectural challenge, right? 830 00:52:39.230 --> 00:52:42.510 So the number of calls per se 831 00:52:43.470 --> 00:52:47.310 didn't affect in our case like the latency directly. But if something 832 00:52:47.310 --> 00:52:49.470 else affected latency 833 00:52:50.590 --> 00:52:54.350 reversely that would mean that million calls were affected by that, 834 00:52:54.430 --> 00:52:58.270 right? So the stakes are becoming very, very high. 835 00:52:58.270 --> 00:53:01.710 So you cannot work operate as a startup anymore, 836 00:53:02.030 --> 00:53:05.830 right? By pushing Things and Version 1 you 837 00:53:05.830 --> 00:53:09.030 don't care if anything broken. So suddenly you have to 838 00:53:09.030 --> 00:53:12.870 graduate very, very quickly to actually having a 839 00:53:12.870 --> 00:53:16.270 high performing, high reliability like triple 840 00:53:16.270 --> 00:53:19.950 9s 4nines at some.59 reliability 841 00:53:20.030 --> 00:53:23.590 company which is actually to be very frank, is a huge 842 00:53:23.590 --> 00:53:27.390 challenge because you just have an engineering team of like eight people 843 00:53:28.190 --> 00:53:31.630 and suddenly and now we are like 30 plus right? Engineering 844 00:53:31.710 --> 00:53:34.910 team and you have to professionalize like overnight. 845 00:53:35.950 --> 00:53:39.750 And that was like a very big challenge for us. And it's very hard. 846 00:53:39.750 --> 00:53:43.430 This is something, this is something. Even 847 00:53:43.430 --> 00:53:47.150 in the era of AI hasn't changed. It takes time to hire your 848 00:53:47.150 --> 00:53:50.670 team, it takes time to build your culture, it takes time 849 00:53:50.750 --> 00:53:54.580 to have quality people. And that's kind of was 850 00:53:54.580 --> 00:53:57.500 a big challenge to get there. So 851 00:53:58.380 --> 00:54:02.140 to summarize, basically one thing is you have to make sure that the architecture is 852 00:54:02.140 --> 00:54:05.780 there, but you also have to build a very world class team because 853 00:54:05.780 --> 00:54:09.260 your business processes, your engineering processes 854 00:54:09.420 --> 00:54:12.220 are going to affect if you push something 855 00:54:13.420 --> 00:54:16.780 to deployment and it affects even 856 00:54:17.100 --> 00:54:20.700 minimally one latency or et 857 00:54:20.700 --> 00:54:23.800 cetera. It means 4 million calls at the same time 858 00:54:24.590 --> 00:54:28.310 and then you will have too many fires you will not be able 859 00:54:28.310 --> 00:54:31.710 to extinguish. Right. So you have to 860 00:54:32.350 --> 00:54:35.990 build a very strong system to have to ensure 861 00:54:35.990 --> 00:54:36.910 that reliability. 862 00:54:40.670 --> 00:54:43.950 Getting professional, getting a big team very fast. 863 00:54:44.910 --> 00:54:48.190 Always a challenge. As you said, 864 00:54:48.510 --> 00:54:50.910 you can change it simply by applying AI. 865 00:54:52.270 --> 00:54:56.110 In the Founders Vault you'll share some internal evaluation 866 00:54:56.190 --> 00:54:59.990 dashboard used for memory co KPIs, guardrails and so 867 00:54:59.990 --> 00:55:03.510 on and so forth. Find the links down here in the show notes to 868 00:55:03.510 --> 00:55:07.110 subscribe to our premium offering the Founders Vault. Just five 869 00:55:07.110 --> 00:55:10.590 bucks a month. Let us go 870 00:55:10.750 --> 00:55:14.110 a little bit from what is now working into the future 871 00:55:14.110 --> 00:55:17.790 outlook. What's next for voice 872 00:55:17.790 --> 00:55:21.340 to voice LLMs reasoning and 873 00:55:22.140 --> 00:55:25.340 on device on edge. And where does memory 874 00:55:25.980 --> 00:55:29.260 evolve like agent teams, shared state 875 00:55:29.500 --> 00:55:33.260 auto summaries. So I would say like the near 876 00:55:33.260 --> 00:55:36.220 future in terms of like roadmap is clear. 877 00:55:38.620 --> 00:55:41.820 Longer than that. It's always hard in AI. It's always 878 00:55:42.620 --> 00:55:46.300 like basically I think humans have been historically terrible 879 00:55:46.300 --> 00:55:50.100 at predicting the future. Right. I 880 00:55:50.100 --> 00:55:53.920 usually say in such cases in our interviews, I know predictions are 881 00:55:53.920 --> 00:55:57.600 hard, especially concerning the future. Right. But, but, but you're 882 00:55:57.600 --> 00:56:01.160 here and we'll give it a shot. Yes, no, no, I'm happy to share 883 00:56:01.320 --> 00:56:04.880 actually. I am one of the guys who always has an 884 00:56:04.880 --> 00:56:08.600 opinion on these things. But humans are generally. And 885 00:56:08.680 --> 00:56:12.520 like for example, everyone 2023 was saying AGI et 886 00:56:12.520 --> 00:56:16.280 cetera and now it's like very, very different. So look, 887 00:56:16.920 --> 00:56:20.720 so that the near future, I think at least for our market, right. 888 00:56:21.030 --> 00:56:24.310 For our product, it's all about basically 889 00:56:24.550 --> 00:56:26.710 reliability, quality. 890 00:56:28.630 --> 00:56:32.230 Ensuring outcomes like deterministic outcomes. Right. 891 00:56:32.310 --> 00:56:36.070 So that reducing the, I would say the chance, 892 00:56:36.470 --> 00:56:40.230 right? Reducing the chance in the agentic space. Right. 893 00:56:40.230 --> 00:56:43.990 So that the ROE is evident. And there are many 894 00:56:43.990 --> 00:56:47.480 things, right? It's there. You build architectural stuff around your infrastructure 895 00:56:47.630 --> 00:56:51.390 structure, but you also build features. This is the Bell framework 896 00:56:51.390 --> 00:56:53.630 which we will announce very soon. 897 00:56:55.470 --> 00:56:58.830 And many similar features, right. You built now 898 00:56:59.230 --> 00:56:59.630 and. 899 00:57:02.910 --> 00:57:05.630 On the more I would say down the road 900 00:57:06.670 --> 00:57:10.110 in the future. I think I'm really curious 901 00:57:10.270 --> 00:57:13.940 kind of what's the next step from LLMs? Because basically, basically 902 00:57:13.940 --> 00:57:17.620 products like Synthfloww very much depend on 903 00:57:17.620 --> 00:57:21.340 that, right? Like what's next in that, what's next in speech to 904 00:57:21.340 --> 00:57:25.100 text. And there is also like the voice to voice model from OpenAI. Right. 905 00:57:25.100 --> 00:57:28.660 The reason we're not using that because it's like 906 00:57:28.660 --> 00:57:32.420 commercially not viable yet. It's very expensive and it 907 00:57:32.420 --> 00:57:36.020 also has more limitations when it comes to 908 00:57:36.020 --> 00:57:39.420 actually agent doing tasks. Right. The 909 00:57:39.420 --> 00:57:43.040 RPA part. So it will be I think there 910 00:57:43.040 --> 00:57:46.760 soon. So that's one of my predictions is that voice to voice 911 00:57:46.760 --> 00:57:50.440 and the voice to voice is amazing experience. Right. And you have 912 00:57:50.600 --> 00:57:53.480 many of the problems that you have with the current 913 00:57:54.120 --> 00:57:57.960 orchestration, you don't have it with there. So I'm pretty sure this will be 914 00:57:58.680 --> 00:58:01.720 also state of the art very soon. 915 00:58:04.520 --> 00:58:08.100 I think that's an interesting part. I think we'll see also like 916 00:58:08.100 --> 00:58:11.900 huge progress in terms of long 917 00:58:11.900 --> 00:58:15.380 format. So right now AI is very good for shorter 918 00:58:15.380 --> 00:58:19.100 conversations, right. And finishing something and doing a 919 00:58:19.100 --> 00:58:22.420 job like I don't know, one minute conversation, customer support. Right, 920 00:58:22.740 --> 00:58:25.620 but what about one hour conversations? Right. 921 00:58:28.340 --> 00:58:32.180 Because the longer these conversations go, the harder it 922 00:58:32.180 --> 00:58:36.030 actually for AI to stay on track. So I 923 00:58:36.030 --> 00:58:39.630 think that's an interesting to see like what's on the long format part 924 00:58:39.950 --> 00:58:43.590 and of course like the complexity of the tasks. The 925 00:58:43.590 --> 00:58:47.230 tasks that AI does nowadays are relatively simple or straightforward, 926 00:58:47.230 --> 00:58:50.910 right. Like talk, capture something, share something, 927 00:58:51.390 --> 00:58:55.030 update something somewhere. Right. But going 928 00:58:55.030 --> 00:58:58.750 forward maybe it can like do something end to end, 929 00:58:58.750 --> 00:59:02.170 right. Go and book and I don't know, follow and 930 00:59:02.810 --> 00:59:06.170 really complicated tasks. I would say 931 00:59:06.570 --> 00:59:09.210 this is what I would expect this to develop. 932 00:59:10.490 --> 00:59:14.090 However, it's not going to be an overnight 933 00:59:14.090 --> 00:59:17.930 transition. It's going to take few years at least. 934 00:59:19.450 --> 00:59:23.210 Not only that, people don't shy away from 935 00:59:23.210 --> 00:59:27.050 interacting when somebody calls, hey, this is a booking 936 00:59:27.050 --> 00:59:30.460 service and you call a restaurant. In 937 00:59:30.460 --> 00:59:34.100 24 months. What percentage 938 00:59:34.100 --> 00:59:37.940 of customer calls will AI handle end 939 00:59:37.940 --> 00:59:41.700 to end in regulated industries? And what will 940 00:59:41.700 --> 00:59:45.380 still require human? Look, 941 00:59:45.940 --> 00:59:49.620 I don't think AI is going to completely replace humans. 942 00:59:50.500 --> 00:59:54.220 I keep saying this everywhere, that AI 943 00:59:54.220 --> 00:59:58.030 is going to augment humans. I don't think humans will 944 00:59:58.030 --> 01:00:01.710 be out of the loop in the near future. That's not 945 01:00:01.710 --> 01:00:05.030 going to happen because I don't see humans 946 01:00:05.030 --> 01:00:07.470 trusting. Like I would say. 947 01:00:09.790 --> 01:00:13.550 Not only trusting but AI. It just doesn't have 948 01:00:13.550 --> 01:00:17.150 that capacity to do very complicated things. Right. 949 01:00:17.150 --> 01:00:19.630 And very long, I would say 950 01:00:20.910 --> 01:00:24.000 RPAs automations. Right. So. 951 01:00:25.520 --> 01:00:29.280 I think in the next five, so look for some 952 01:00:29.280 --> 01:00:33.000 cases it's going to be 100% to answer your question, it's going to be 953 01:00:33.000 --> 01:00:36.800 100% of the calls. Like for example for transfers, routings, 954 01:00:37.680 --> 01:00:41.200 FAQ, first line of defense. I think it's going to be 955 01:00:41.200 --> 01:00:44.000 100% is going to be done by AI. 956 01:00:45.680 --> 01:00:48.840 For some cases, more complicated ones. Right. 957 01:00:48.840 --> 01:00:52.680 Processing, et cetera, Maybe even on the governmental side, I think it's going 958 01:00:52.680 --> 01:00:56.280 to be to be way slower. Right. Like I don't Know, let's say taxes or 959 01:00:56.280 --> 01:00:59.440 this kind of areas. I don't see it like 960 01:01:00.160 --> 01:01:03.760 adopting very quickly. That said, regulated spaces. 961 01:01:03.760 --> 01:01:07.440 Also like we already see like crazy adoption in healthcare and financial 962 01:01:07.440 --> 01:01:11.200 institutions despite. You just need to be compliant, right? 963 01:01:11.440 --> 01:01:15.280 You have to have everything like you have to have HIPAA compliance, you have to 964 01:01:15.280 --> 01:01:19.080 have pen tests, right. You have to ensure there and, and, and 965 01:01:19.080 --> 01:01:22.560 you're signing SLAs and this kind of things with the customers, right? 966 01:01:23.590 --> 01:01:27.350 If the state of the technology is there that you can guarantee that, for example 967 01:01:27.430 --> 01:01:30.310 with visual graph builders, then. 968 01:01:32.950 --> 01:01:36.430 Basically if the AI works for 10% of the calls, there is no reason for 969 01:01:36.430 --> 01:01:40.030 them not to deploy it for 100% of the calls. And I think it's going 970 01:01:40.030 --> 01:01:43.710 to be very different case by case, like use case from use case. For some 971 01:01:43.710 --> 01:01:47.030 use cases going to be 100% for some probably 972 01:01:47.270 --> 01:01:50.590 very low. Not even. 1%. 973 01:01:50.750 --> 01:01:54.150 Yeah. A little bit of 974 01:01:54.150 --> 01:01:57.750 contrarian view when we go close to the end. By the way, 975 01:01:57.750 --> 01:02:01.390 you're holding up pretty well and we're recording for no more than an 976 01:02:01.390 --> 01:02:04.629 hour now. I was wondering what's an 977 01:02:04.629 --> 01:02:08.350 overrated metric or myth in an AI 978 01:02:08.350 --> 01:02:11.550 contact center deployment that teams should 979 01:02:11.630 --> 01:02:15.390 ignore or that it's still missing the return on investment? 980 01:02:17.290 --> 01:02:20.490 Yeah, it's a very good question. So 981 01:02:21.290 --> 01:02:24.890 look, I don't think that the classical ones like 982 01:02:24.890 --> 01:02:28.650 CSAT. Or 983 01:02:29.050 --> 01:02:32.370 like things like deflection rate or average 984 01:02:32.370 --> 01:02:36.090 resolution time or time to answer 985 01:02:36.650 --> 01:02:39.690 these things are actually very important, 986 01:02:40.330 --> 01:02:42.570 right? And they do matter, right. 987 01:02:46.040 --> 01:02:49.560 I think one thing that I always look at, 988 01:02:49.640 --> 01:02:53.000 but I don't care so in a sense is that 989 01:02:53.640 --> 01:02:57.360 is the subjective feedback that we get very often. For example, sometimes 990 01:02:57.360 --> 01:02:59.320 people say oh, it was too robotic. 991 01:03:01.400 --> 01:03:05.000 And of course at this stage it still could be 992 01:03:05.240 --> 01:03:08.960 some elements of that it's an AI. And also I tell to 993 01:03:08.960 --> 01:03:12.610 all my customers, customers, you know, like you have to disclose that 994 01:03:12.610 --> 01:03:16.410 it's an AI. It's not built to lie to people that 995 01:03:16.410 --> 01:03:20.250 it's human. That's a bad practice, right? So when humans talk with 996 01:03:20.250 --> 01:03:24.050 AI, they should know they are talking to AI and they should have the 997 01:03:24.050 --> 01:03:27.850 right expectations, right? If someone tells me, ah, it didn't 998 01:03:27.850 --> 01:03:31.210 feel like it was a bit robotic, I think that's fine. 999 01:03:31.770 --> 01:03:34.850 Of course, as long as you know it's an AI, as long as it does 1000 01:03:34.850 --> 01:03:38.640 the job for you, right? You have a good experience, right? Better 1001 01:03:38.640 --> 01:03:42.400 than the Ivr trees or etc of the past. We're 1002 01:03:42.400 --> 01:03:45.800 making a huge progress as a humanity in that area. 1003 01:03:46.440 --> 01:03:50.120 And this is like for me, I wouldn't care that 1004 01:03:50.120 --> 01:03:53.680 much about that we're getting that subjective feedback that 1005 01:03:53.680 --> 01:03:57.040 it's. Because first of all, it doesn't have to be exactly like 1006 01:03:57.040 --> 01:04:00.720 humans. Right. As long as your experience is good, you know, you're talking to AI 1007 01:04:00.720 --> 01:04:04.330 and your problem is solved on the phone. What's the 1008 01:04:04.330 --> 01:04:05.170 problem? Right. 1009 01:04:08.050 --> 01:04:11.690 So that's kind of what I look right now. And I'm very, very tolerant in 1010 01:04:11.690 --> 01:04:15.490 the subjectivity towards AI. And 1011 01:04:15.490 --> 01:04:19.010 I just think we just need more time as humans to actually get 1012 01:04:19.010 --> 01:04:22.130 there. To say, okay. 1013 01:04:23.650 --> 01:04:25.970 This is where we are. Right. Yeah. 1014 01:04:27.250 --> 01:04:30.970 Interesting. Yeah. Bottom line, the slower developing part is 1015 01:04:30.970 --> 01:04:34.750 not AI, it's the humans in the. Yeah. 1016 01:04:35.710 --> 01:04:39.430 For our audience, I was wondering, should AI remember 1017 01:04:39.430 --> 01:04:42.910 you across calls if it improves your service? 1018 01:04:43.470 --> 01:04:45.630 Yes or no? Tell us why in the comments. 1019 01:04:47.710 --> 01:04:50.990 An advice for founders and CX leaders. 1020 01:04:52.750 --> 01:04:56.470 What's your three step guide here? And can you 1021 01:04:56.470 --> 01:05:00.150 share some advice for our listeners who are 1022 01:05:00.150 --> 01:05:03.730 mostly founders and executives from this 1023 01:05:03.730 --> 01:05:07.250 plan? Yes. So look. 1024 01:05:09.730 --> 01:05:13.570 I would share some advice on a bit like higher level. Right. 1025 01:05:13.810 --> 01:05:17.530 So I think one thing that. It'S 1026 01:05:17.530 --> 01:05:21.330 very different right now in the AI era compared to before 1027 01:05:21.330 --> 01:05:25.050 AI is that it's very stressful. I 1028 01:05:25.050 --> 01:05:28.650 don't think that founders have been under 1029 01:05:28.650 --> 01:05:31.930 such a pressure and stress ever, 1030 01:05:32.410 --> 01:05:35.050 ever in the history of startups because 1031 01:05:36.250 --> 01:05:39.610 things are moving basically on weekly basis. 1032 01:05:40.170 --> 01:05:44.010 Like every week, every day something new is happening. Like I wake 1033 01:05:44.010 --> 01:05:47.850 up in the morning, I look into my phone. Some, someone, 1034 01:05:48.010 --> 01:05:51.650 something, someone new model is out there. Someone released that, someone 1035 01:05:51.650 --> 01:05:55.370 released this. This is changing. OpenAI is doing that and does 1036 01:05:56.210 --> 01:05:59.890 so I think like these skills, navigating skills. 1037 01:05:59.970 --> 01:06:03.370 Right. Filtering out through the noise is a 1038 01:06:03.370 --> 01:06:07.210 superpower at this stage and founders need 1039 01:06:07.210 --> 01:06:11.050 to learn actually not to get too distracted 1040 01:06:11.050 --> 01:06:14.810 and too stressed down by the noise. It's going to be 1041 01:06:14.810 --> 01:06:18.450 there for a while. It's a new technology. We're experiencing 1042 01:06:18.450 --> 01:06:21.330 an immense progress as a humanity. 1043 01:06:22.370 --> 01:06:26.140 I think this is already way bigger than Internet at. Right. What's 1044 01:06:26.140 --> 01:06:29.900 happening right now. This is, we have never seen this, this 1045 01:06:29.900 --> 01:06:33.660 type of progress. Right. And, and on a weekly basis. 1046 01:06:33.900 --> 01:06:37.740 So really understanding and focusing on value building business 1047 01:06:38.140 --> 01:06:41.860 is, is, is harder right now because you. There are many shiny things, 1048 01:06:41.860 --> 01:06:45.100 right. You always confront these decisions. 1049 01:06:45.260 --> 01:06:49.100 Oh, I know. This new model came. Shall I use this and that or 1050 01:06:49.100 --> 01:06:52.690 do I go this way? This way? What about this? Right? And, and I think 1051 01:06:52.930 --> 01:06:56.530 being very close to the customers, very, very close to the 1052 01:06:56.530 --> 01:07:00.290 customers and focusing on building long term 1053 01:07:00.290 --> 01:07:03.570 value for them for your customers is the key. 1054 01:07:05.090 --> 01:07:08.930 This is my view, this is what I will highly recommend 1055 01:07:09.330 --> 01:07:13.170 based on my own experience in the last Three years, I would 1056 01:07:13.170 --> 01:07:16.850 say. Yeah. Three more questions for you. 1057 01:07:18.380 --> 01:07:21.980 Closing Bold Insights Finish this sentence. In 1058 01:07:21.980 --> 01:07:25.780 three years, voice AI with memory will be 1059 01:07:25.780 --> 01:07:29.500 the default interface for for. 1060 01:07:30.620 --> 01:07:34.059 It will be default interface for contact 1061 01:07:34.140 --> 01:07:37.580 centers, both external and internal. It means 1062 01:07:37.580 --> 01:07:40.900 external BPOs, right? Contact center as a service and 1063 01:07:40.900 --> 01:07:43.740 internal contact centers for insurance, 1064 01:07:44.550 --> 01:07:48.270 financial institutions, companies. Because we already see 1065 01:07:48.270 --> 01:07:51.910 the ROE and the case is very clear. It's just 1066 01:07:51.910 --> 01:07:55.590 that it's just a matter of time until enterprises 1067 01:07:55.590 --> 01:07:58.950 start deploying and adopting this in their companies. 1068 01:08:01.430 --> 01:08:04.710 And two closing questions. 1069 01:08:06.390 --> 01:08:10.030 Are you open to talk to new investors? Yes, definitely. 1070 01:08:10.030 --> 01:08:13.770 Always happy to. I think the best format 1071 01:08:13.770 --> 01:08:16.610 is always I say it Coffee Chat. Yes, 1072 01:08:17.330 --> 01:08:21.130 we link your LinkedIn profile down here in the show notes and I do believe 1073 01:08:21.130 --> 01:08:24.970 every investor that has been sticking around for over an hour, they'll 1074 01:08:24.970 --> 01:08:28.530 be interested to talk to you. And of course you're looking 1075 01:08:28.690 --> 01:08:32.250 for hiring talented employees. Yes, 1076 01:08:32.250 --> 01:08:35.970 always. Right. And we're hiring on a rolling basis 1077 01:08:36.290 --> 01:08:39.899 so we are not a company that just 1078 01:08:40.379 --> 01:08:43.979 fills in gaps. We always look for exceptional talent 1079 01:08:43.979 --> 01:08:47.419 whether it's AI, engineering, voice AI 1080 01:08:47.659 --> 01:08:51.139 or also commercial roles, sales, solutions, 1081 01:08:51.139 --> 01:08:54.779 engineering, forward deployed engineers. We're always open 1082 01:08:54.859 --> 01:08:57.659 on a rolling basis for exceptional talent 1083 01:08:58.459 --> 01:09:02.099 and we guarantee you and promise you also an 1084 01:09:02.099 --> 01:09:05.720 exceptional environment to work on cool things and 1085 01:09:05.880 --> 01:09:09.640 grow as a professional. Awesome. 1086 01:09:09.640 --> 01:09:13.480 Hey Cobb, with such a pleasure having you here for 1087 01:09:13.480 --> 01:09:17.080 our premium subscribers. We'll be back shortly with 1088 01:09:17.160 --> 01:09:20.920 the Founders Vault. Thank you, thank you. 1089 01:09:20.920 --> 01:09:21.880 Thanks for having me. 1090 01:09:26.920 --> 01:09:30.530 That's all folks. Find more news, streams, 1091 01:09:30.690 --> 01:09:31.730 events and 1092 01:09:31.730 --> 01:09:36.130 interviews@www.startuprat.IO. 1093 01:09:36.530 --> 01:09:38.770 remember, sharing is caring.