#271 How aiOla Turns Natural, Multilingual Speech into Workflow-Ready Data

0:00If I can help a supermarket chain, which is not very advanced technology-wise, but never invests in AI, to actually leverage the own generated data to make better decisions, and I have such an ROI, I can do it everywhere else.

0:18SlatorPod Host:Hey everyone, and welcome back to SlatorPod. So we want to continue to explore multilingual speech tech and AI. So we asked Amir Haramaty to join the podcast. So Amir is the co-founder and president of AIOLA. So AIOLA is a deep tech voice, speech, and conversational AI company. And I'll say a few more things about the company. So they work on enterprise workflows. The problem they tackle is standard speech tech fails in real business environments. And we really want to talk about that challenge. and yeah so they focus on must-have industries like food, pharma, logistics, energy, semiconductors and I'm sure Amir is going to tell us a lot more so Amir thanks so much for joining today.

1:05Thank you Florian thank you for having me. Where in the world are you recording this from today? It's challenging because I was in Frankfurt yesterday, Madrid the day before I'm actually in Tel Aviv today and I'll be back in Europe on Sunday but you know especially now when speech is no longer It's no longer a secret. It seems like it's one of the hottest topics and I cannot run fast enough.

1:30SlatorPod Host:100%. Yeah, that's why we also wanted to get you on the podcast. And I have to confess, I listened to a few of your podcasts and it's super, super interesting. So I'm very excited to get this podcast on the road today. So first, what does AIOLA stand for? Actually, AIOLA is AI Operating Layer. That's what it stands for. AI Operating Layer. What a great name. So take us back to the time first before AIOla, you worked with and found multiple AI and data-focused ventures before. So just kind of tell us the trajectory, where this came from, and then also the company in a nutshell, where the idea or motivation originated.

2:07Absolutely. So some may claim that I'm a serial entrepreneur. I don't like that word anymore. I think it's been washed down completely. The way I like to define myself as a serial problem solver is to find a real problem, build an incredible team, bring amazing technology that can generate value. That's a simple blueprint. I've been lucky enough to do it now for the sixth time. I've been multiple success stories with amazing outcomes. The last one got me deep into AI 10 years ago, long before AI and I became the most used and abused letters in the English language at present time. And actually, with my previous company, we were the AI platform of choice of one of the leading management consulting companies globally.

2:52We had this privilege to be with them, to serve with them in about more than a triple-digit number of engagements with Fortune 1000, which is a rare privilege to have a courtside seat, if you will, to see what's working, what's not working, where the big promise of AI is meeting traditional enterprises. and I cannot sugarcoat it. It was not a pretty sight. Many takeaways from that, but the three main ones is the fact that with everything that happened with AI, which is incredible, I think collectively we got a bit confused. AI tools are amazing, but despite the fact that AI is still just tools, which really polarized and emphasized the fact it's no longer an AI challenge, it's truly a data challenge because the majority of data is still not part of the game because it's uncaptured and unstructured.

3:46We built this incredible monster, the AI tools, we need to feed them. So the challenge right now is absolutely data challenge. Now, there are very few things that people hate more than entering data. Data entry is a chore. They treat it as such. Therefore, it's reflected in a quality, quantity, timing. And therefore, Zurich, you know, we got a challenge. We got a problem here. It's all about data. Most of the data is still not part of the game. Data entry is a challenge. So how can we fix that? And then speech is the most natural form of communication. It's three times faster than anything else.

4:23Actually, in one third of the time, you're capturing 3x amount of data. And eventually, any human-machine interface will be voice first. There's no question in my mind. However, when you put it in a real setting, in a real world setting, the combination of languages, accents, challenging acoustic environments, because we are not working in lab. It can be production floor or a busy airport. And above all, we learned, Florian, that every industry and every company and every location has a very specific jargon, that in all fairness, there's no ASR, automatic speech recognition platform, off the shelf that was built to handle it.

5:00Hence, the biggest challenge, when you put the concept, I would love to use voice, but the setting of language, accent, acoustic environment, and jargon, the performance is subpar and unacceptable. So first, we view it as a triple challenge. One, how can we get it to be close to 100 % in those settings? Even when we do so, we're not interested to be a passive listener or anything like that. We actually want to automatically separate the noise from the signal, the width from the shaft, or whatever you want to call it, but to be able to extract only what's relevant for that specific process or workflow.

5:35and specifically, by the way, with all the ultra sensitivity and rightly so to privacy, it's like I'm going to take the data that you want to share with me, only when you share with me, and I'll know automatically how to clean the chit-chat from the key data point I want to extract. So one is accuracy. Two is separating the noise from the signal. And third, I think is the most important one, we're not talking about speech to text. We're not speaking about speech to transcription. Speech for us is highly unstructured data that we turn to highly structured data. So it's speech to data, speech to schema, speech to workflow.

6:13And that's the vision we had in mind when we initiated IOLA. And actually, I do many things well, but I'm not a speech expert. And I wanted to get some stamp of approval in the right direction. So I was searching for one of the best experts in the world to give me that blessing. I found one of them here in Israel, Professor Joseph Keshet, which at the time was leading the research of Alexa for Amazon, and some claimed that Siri was founded in his laboratory. And after spending several hours with Professor Keshet, he said, Amir, you want to do something really big here? You haven't offered me a thing, but I already accepted it.

6:50Tell me that you want me and I'll join you as a chief scientist. And he did. And so that's the genesis. That's how IOLA came about. But then we took it to the next level and decided how are we going to attack it in a completely different way. But that's a completely different story.

7:04SlatorPod Host:Yeah, we want to touch on all of that. So I want to start with the kind of the question, like we have these giant multipurpose LLMs. They've been trained, you know, on I don't know how many gigawatts are these days. And, you know, you need to build nuclear power plants to power them and trillions of dollars are pouring into it. But I think in a previous podcast, you said instead of a huge general LLM, you're building a small, very kind of workflow-specific language model that is tuned to a single process. So would that be a fair description? And tell us more about this. It's accurate enough. And the main thing there, actually, it was the discussions we had.

7:39And I realized what you described is absolutely true about the big LLMs. if I try to oversimplify, they're constantly using brute force to boil the ocean and then to try to narrow it down to your specific use case. And we were talking and said, why do we need to boil the ocean in order to make coffee? All we need is that much water. With that process in mind, think about, and it's a real use case, by the way, a chicken nuggets manufacturer for one of the largest chains, global chains with golden arches worldwide. I cannot mention the name. We'll let the imagination run wild. Ah, yes, very wild.

8:17But in any case, they still have to meet some very strict procedures and very strict guidance. And they have to complete the process, and it's been done in Thailand. And Thai is a very challenging language. It's a tongue language. You probably know that very well from your past. You know, you can say the same word that has seven different meanings depends on the tongue. and none of us, unfortunately, speaks Thai. And all we need from them is to take that one little form in Thai, a list of keywords. We don't need to retrain the data. And everything in zero shot for that specific workflow or process or inspection form, we're going to build a language model that can be deployed on any Android or iOS device.

9:03They're going to turn in accuracy close to 100 % in those specific settings of acoustic jargon, etc., etc., then now they can walk a talk hands-free and the data will be captured. And a process that previously took two hours, now it dropped to 34 minutes. So first, it's an unbelievable efficiency against, especially in these razor-thin margin industries. All of a sudden, give them an hour plus of production. It's very dramatic. But this is just the first wave. The second wave, we now created data data that previously was not captured and not structured, which now becomes a raw material that allow you to connect the dots beyond what the human eye can see and the human brain can process and constantly bring evergreen, ever-learning insights and intelligence and identify trends before the form.

9:51So there are two elements here, and then we can talk about the importance of agents, etc. on top of it. But that's the story here. Now, the most important part at the end of the day, And I'm really glad that three months ago, four months ago, in July, MIT came with an incredible, interesting report that claimed that 95 % of AI pilots at enterprises have failed to demonstrate value, impact, and ROI. Because at the end of the day, you know, there's a lot of discussions right now about the AI bubble, et cetera. Well, breaking news, we all know there's a bubble. Absolutely, there is a bubble. But underneath it, I think very quickly we'll be able to separate the pretenders from the contenders.

10:35And basically say, at the end of the day, it doesn't matter how small or big you are, you still need to do more with less. How you demonstrate an ROI. Let me give you a very fresh example that starts, and I have multiple use cases. Every single use case we do, even before we start, we're simulating what's going to be the return of investment, what's going to be the impact. So right now, for example, we're dealing with a Fortune 50 sales organization that's conducting sales meetings at client sites. And they're using one of the leading CRMs, Salesforce in this case, as the backbone. And again, the first challenge is people treat it as a chore.

11:13So the data of the quality, the quantity of the real-time data entry struggles. and second is, so first we were able to take it from 7.5 minutes to 1.5 minutes to enter the data. Much more accurate, perfect data, no mistakes, real-time data. But by doing so, we actually freed up 120 minutes a week per salesperson that he or she now will be able to do two more meetings per week. So first, you drive efficiency and data quality. But second, you're generating more opportunities to do more with less with the same amount of time. So everything we do, and we have those examples from supermarket chains to airlines to manufacturing to pharmaceuticals.

12:02The interesting part about us, we are dumb as it gets. We know nothing about any industry. There's no domain knowledge or expertise. We just know how to take unstructured data by speech, put it in the right place, combine automatic speech recognition with natural language understanding, and be able to generate quantifiable and measurable value.

12:22SlatorPod Host:In the industry you're seeing, or you've in the past seen kind of most rapid adoption, the Salesforce example you gave would kind of be an outlier, right? Because you guys were really strong and focused on these kind of frontline environments, which is maybe what a lot of the kind of tech people typically wouldn't even think about and think of, right? So how did this come about? The Thai example is fascinating, right? This is not something somebody in, I don't know, Palo Alto would think of when they start coding. Absolutely right. And I think this is, you know, people using big words like democratization, et cetera, et cetera, of all this technology.

12:56And I love that part because at the end of the day, when you talk about frontline workers, we're talking about Florian. We're talking about 2 billion frontline workers. That's the size of the market.

13:05SlatorPod Host:It's huge. Yes. And it doesn't matter if you're in oil rig in the middle of the North Sea or you're doing it right now in the chicken nuggets factory in Thailand. Or another example, we worked in one of the largest jewelry manufacturers out of Hong Kong in China. and they're doing visual inspection, QA inspection of the parts. And an inspection that currently taking 175 seconds to view and write by speech, it's going down to 30 seconds. So now you can do five, six X amount of pieces of jewelry in the same three minutes. So for us, at the end of the day, it's really the ability. And I love the fact that one of the things that I always talk about with all due respect to the technology, and we are very proud of our technology, Don't get me wrong.

13:48It's never about technology. It's about usage, about adoption, about the human factor. And at the end of the day, when I'm working right now with an airline baggage lady, that her job is to write the content of an orphan bag that arrives a day after you arrived. And now they're trying to match between you two. And she needs to go and write down a report. Now she opened the bag and she just describes it. Hoka shoes, deodorant, red spot, whatever. And the quote I got from that lady was, this is changing my life. So the combination of the technology efficiency and all these elements, or a different example, I just came back from Germany.

14:29We were working with automotive. And it's a very clean environment situation, robotics, and they have big screens where they're monitoring the shift. And then there's always something goes wrong, and there's always incident throughout the shift. They have to stay at the end of the shift to write a report. But when wrong, they hate it. They want to rush home. They'll do the bare minimum. Now we allow them to speak in real time during the shift. And you know, my mother always told me, you have two ears and one mouth. Try to keep that proportion. And one thing which I learned, it's amazing what you're able to hear, Florian, if you're willing to listen.

15:05Now, if you give them a voice, they all speak up. and once they'll speak up, it's amazing how much information you can extract out of it, how much data you can extract out of it.

15:15SlatorPod Host:And also one of your kind of USPs are very strong areas is jargon handling, right? That's what we learned in our research. And I think you just gave two interesting examples. I guess the suitcase example is a little less jargon because, okay, Hoka and I don't know, like sweater or pants, right? But in the automotive, it's full of jargon. So how do you guys handle this super jargon-heavy environment? And also, I think in another podcast, I got one data point that you said that each industry only has like 2 ,000 to 3 ,000 relevant words or like super jargony words. No, no. Actually, it's limited.

15:53It can be 20 ,000, 30 ,000. It doesn't matter. But it's limited. It's not endless. And, you know, it's a learning machine. So basically, we can basically swallow all those keywords. And I give you, and we can do it on a process or location or accent or whatever. You know, we're working right now with one of the stock exchanges. And they say, for example, this. What is this? This is December.

16:19SlatorPod Host:Okay. And I give you even a more challenging situation. We're working right now with one of the leading airports in Europe. And we build a language model in German because it's in Germany. And the performance was subpar to what we used to. And we wanted to understand why. It turns out that most of the users in that specific airport are mostly Turkish immigrants. That speaks a broken language. It's Turkish-German. It's not Turkish. It's not German. It's like Spanglish, okay, in the U.S. And then we, for us, the language, we don't look at it from a linguistic perspective. The language is just the carrier of the data on top.

16:59So we use actually, I'm making it up right now for illustration purposes only, let's say, instead of saying good to Morgan, you say good at Jorgen. The Iola jargon engine identified Jorgen as a jargon word for Morgan. So even though we build now a language model for a language that doesn't exist linguistically, because we understand how to identify the keywords that we need to extract out of that carrier just happened to be, in this case, Turkish-German.

17:26SlatorPod Host:Okay, that's fascinating. I mean, that's so niche. Like, okay, airport, specific location, specific type of linguistic community almost for that specific task, right? And then you guys are going in, identifying that and making the model work. Wow, okay. How replicatable would this be? I mean, okay, you fix it for the German airport, but then would this kind of relatively easily scale across all of the airports and all of the workforces? Two separate things here. One, again, is to identify the use cases where we make the biggest impact on the shortest period of time with minimum obstacle possible.

18:00Like you give an example. Baggage is great, but, for example, safety for airlines and airports, it's paramount. It's super critical. And actually, we're turning, it's a culture thing, we're turning every single employee within that company or airline or specific airport to safety ambassadors. So incident reporting and time for resolution shrunk significantly. So, you know, this is a use case. forget about the language as an accent for a second. The key here, and it's a big challenge, is all the big LLM, all the big players you mentioned, 11 Labs, DeepGram, Assembly, all those guys, Whisper, OpenAI, they'll give you a plain vanilla.

18:40Here's an API, good luck, Florian. We're saying that's going to take you that far. It's going to hit the ceiling when it comes to jargon, to keywords, all these things we talked about a minute ago, our ability, you know, I don't know if you like truffle, tartufo, okay? An omelet is just an omelet until you put truffle on top. We are the tartufo, okay? Good one. Okay? Especially when I like it. Some people feel it too strong, okay? But for me, that's the point here. Now, it's not sustainable and it's not scalable unless you're able to do that automatically. And the ability to tailor the nuance for a specific form of workflow or inspection is really the art that we generated that's directly linked to return of investment.

19:29We're not necessarily selling it to CIOs. Our biggest friends are the CFOs and CEOs and CEOs because it allows them to do that, you know, to do more with less. The second part, going back to those exotic type of processes that's far away from Palo Alto, in a way, I feel like we fast track and skip generations. Because everybody now, you know, when we started the project with the grocery chain, it's in Canada. And French-Canadian is a very interesting language to start with. When we succeeded, I got a call from the CEO. And he said, I'm here for three years, and it's a$25 billion company. For three years, every board meeting started with a question.

20:16Say, Mr. CEO, what are you doing in AI? And he said, mostly I'm flailing, creating some out here with my arms and talking about chatbots and all kinds of things. And for the first time, I was sitting tall. It's a real use case. They have to inspect the temperature of the meat or the consumable goods inside a refrigerator, not a refrigerator, and reported in real time, say, on a single process like this, with first 600 stores, they saved 110 ,000 hours, they went directly to EBITDA. The ROI is 5x. And in addition now, it's not waiting for the goods to go bad and then try to investigate what happened.

20:54Actually, if the temperature is above the threshold, it's automatically allowed them to be proactive. And, you know, this is a 2 % margin industry. and Frank Sinatra saying about New York, if I can make it here, I can make it anywhere. And I realized if I can help a supermarket chain, which is not very advanced technology-wise, that never invests in AI, to actually leverage the own generated data to make better decisions, and I have such an ROI, I can do it everywhere else.

21:20SlatorPod Host:So this is a very kind of, you mentioned CFO, CEO, so this is a very challenging, maybe even a little bit long enterprise sales cycle, I guess you're in, and you have the ROI, as like a pretty big arrow in your sales quiver, right? So how do you go about kind of go-to-market enterprise sales? Do they come to you? Do you go to them? It's a great question. And just to give you a hint about my background, this is really the areas where I specialize. One of my previous companies, I took from 16 direct enterprise clients to 16 ,000 channel clients. And I've done that, I've built that many times, I've had many mistakes, but I kind of perfected it along the way.

22:00Several things here. The biggest challenge was when we started is like, why speech? Okay, and then when you start, it's not like, oh, I need to have a speech solution and I budget it and I'm looking for the best solution. That's not the case. The first, why speech? And I do it normally to get them excited, to fuel the imagination and demonstrate the art of the possible. That normally leads to what we call a whiteboard session where they start to throw ideas. and I try to navigate into fine areas where we can make the biggest impact at the shortest period of time with minimum obstacle possible.

22:32Now, once I've done it and I've validated it, I realize the market is huge, but there's no time to spend. So I believe that this is, and this is connected to what we just announced about our UST collaboration. It ties in very nicely in a minute. We realize that once we prove it, we don't have the domain expertise, but there's a lot of professional services or digital transformation companies that their job is to do exactly right. And they realized, by the way, I went to the largest of them. It's a 900 ,000-person gorilla. And I tried to talk about what we're doing here, and I got a cold reaction initially.

23:12And I asked them, may I go to the whiteboard and write something? I said, sure. And I wrote W-B-Y-H-W-B-Y-T. Say that's in Hebrew? I said, no. It's what brought you here will not bring you there. And their playground is changing. Their sandbox is changing. So what we are doing right now, we partner with USD. USD is an amazing organization, 30 ,000 people strong, very agile. And among other things, by the way, they have their own Salesforce practice, the big system integrated to Salesforce, to take just what we have done right now for this Fortune 50 company, go get them and replicate it at scale.