Latest / The Adults In The AI Room / Earley AI Podcast - Episode 91: Real-Time Voice Intelligence, Fraud Detection, and AI Guardrails with Mike Pappas

Transcript

Earley AI Podcast - Episode 91: Real-Time Voice Intelligence, Fraud Detection, and AI Guardrails with Mike Pappas · The Adults In The AI Room

• 0:01
  SPEAKER_03: This is the Early AI Podcast hosted by Seth's detailed digital actions, expertise, and experiences of early AI technology to talk about what's possible with artificial intelligence.
• 0:14
  SPEAKER_03: The Early AI podcast is sponsored by Early Information Science, your digital transformation journey with design and deployment of innovative technology solutions as well as better.
• 0:26
  SPEAKER_03: Now enjoy the show.
• 0:28
  SPEAKER_02: Welcome to the Early AI Podcast.
• 0:31
  SPEAKER_02: I'm your host, Seth Early, and what we do in each episode is we examine how AI is changing business from the perspective of creating value, managing information, managing customer experiences, managing employee experiences, and changing how they do business.
• 0:46
  SPEAKER_02: And today we're going to be talking about a part of AI that's becoming increasingly critical, and that is voice, not just speech recognition, but real-time voice intelligence, where AI has to interpret intent, emotion, and risks as conversations unfold.
• 1:03
  SPEAKER_02: Joining me today is Mike Pappas.
• 1:05
  SPEAKER_02: He's CEO of Modulate, and his work began in gaming, which is one of the most difficult environments for voice, but his team's technology is now being applied to a much broader set of challenges, including customer abuse prevention, fraud detection, and AI guardrails for voice-driven systems.
• 1:20
  SPEAKER_02: Mike, welcome to the show.
• 1:22
  SPEAKER_00: Thanks so much for having me, Seth.
• 1:23
  SPEAKER_00: Excited to chat.
• 1:24
  SPEAKER_02: So, Mike, one of the things we like to start with is a misconception.
• 1:27
  SPEAKER_02: I didn't really think that uh there are a lot of corporate applications for voice analysis, right?
• 1:33
  SPEAKER_02: Certainly voice to text.
• 1:34
  SPEAKER_02: And I thought of that as a solved problem, but I really didn't see uh some of the uh implications that we to ended up talking about.
• 1:41
  SPEAKER_02: So, what are the biggest misconceptions that you see about AI when it comes to voice and real-time understanding?
• 1:48
  SPEAKER_00: Yeah, thanks, Seth.
• 1:49
  SPEAKER_00: Uh you you basically just named that misconception.
• 1:52
  SPEAKER_00: It's that taking voice and transcribing it is the same as understanding it.
• 1:56
  SPEAKER_00: And even if we're very good at understanding text, transcribe text is not the same as audio.
• 2:01
  SPEAKER_00: When you hear a voice, you hear the intonation, you hear the emotion, you hear pregnant pauses, you even hear things like the timbre of my voice that tell you things about my age and potential gender.
• 2:11
  SPEAKER_00: There's so much information that's being carried across in that audio that gets lost when you pull down to a transcript.
• 2:17
  SPEAKER_00: And whenever we talk to someone who professionally works in a contact center or deals with large amounts of audio content, they are always saying to us, we know that these transcripts are losing tons of good value.
• 2:29
  SPEAKER_00: We just assume that that's all we have.
• 2:32
  SPEAKER_00: That's the only way for computers to understand audio is to transcribe it.
• 2:36
  SPEAKER_00: And what Modulate has been building is kind of the revolution in a fundamentally new way to listen to audio and turn that into data computers can understand.
• 2:45
  SPEAKER_02: So, so as I say, a lot of people think of voice as solved, right?
• 2:49
  SPEAKER_02: Because they think of it in that one dimension, because they think of, you know, we have speech text, we have voice assistance.
• 2:55
  SPEAKER_02: But what do people misunderstand about what it it takes to make voice AI useful, safe, viable in those real environments, right?
• 3:03
  SPEAKER_02: I was getting some value, and then I listened to the podcast and I was like, oh, okay, now I see what they're saying here.
• 3:08
  SPEAKER_02: Now there's emphasis here.
• 3:09
  SPEAKER_02: Ah, they were they were interspersing this with another conversation.
• 3:13
  SPEAKER_02: Okay, now I get much more of it.
• 3:14
  SPEAKER_02: So I could see right there where the transcript lost so much meaning.
• 3:19
  SPEAKER_02: So tell me more about how people misunderstand what it takes to get that intelligence from from voice.
• 3:25
  SPEAKER_00: Yeah, I mean, the the classic example for something like this is you get a relate to an event with your friend and you get a text from them saying something like, You okay?
• 3:33
  SPEAKER_00: Are they genuinely concerned?
• 3:35
  SPEAKER_00: Are they being passive aggressive?
• 3:36
  SPEAKER_00: Are they trying to figure out if you're even coming to this thing?
• 3:40
  SPEAKER_00: You can't tell when you're looking at that text, right?
• 3:42
  SPEAKER_00: You you need the audio to understand it.
• 3:44
  SPEAKER_00: So that that is the first piece again.
• 3:47
  SPEAKER_00: There's something fundamental to the audio.
• 3:49
  SPEAKER_00: And then in terms of talking about audio being solved, there are a lot of really impressive demos of technology that seems to understand this stuff.
• 3:58
  SPEAKER_00: But the problem is that they're all designed for these really pristine environments where your recording quality is perfect and everyone is speaking to the camera and trying to make sure that they're understood by the AI.
• 4:10
  SPEAKER_00: And when you jump into the real world, you have a messy cell phone connection with, you know, background noise and all kinds of emotion and jargon that people are chat throwing back and forth at each other.
• 4:20
  SPEAKER_00: And being able to build a system that holds up under those circumstances and actually still understands, you know, Seth and Mike are trading insults, but it's because they're old friends and they're just having a good time.
• 4:32
  SPEAKER_02: Like that kind of distinction that's versus they're a fight, they're having a fight here, and this is really aggressive, right?
• 4:37
  SPEAKER_02: Exactly, right?
• 4:38
  SPEAKER_00: And that that's why gaming was such a good starting point for us, is that gaming is just filled with this kind of very contextual, very complicated conversations where you need to deal with emotion and noise and all that to be able to tell the difference between something positive and something negative.
• 4:57
  SPEAKER_02: So tell me more about those use cases.
• 4:59
  SPEAKER_02: You mentioned harassment, you mentioned abuse, mentioned a bunch of different things.
• 5:03
  SPEAKER_02: So talk a little bit about what those early uh use cases were in the gaming world.
• 5:07
  SPEAKER_00: So yeah, in in these online games, something like Call of Duty or Grand Theft Auto, who we both work with, they are trying to encourage players to use voice chat to coordinate with each other, to ally, to get to know each other and just have a more immersive experience.
• 5:23
  SPEAKER_00: Um, but the danger is that when you have that sense of anonymity on the internet, as we know, some people are hostile.
• 5:29
  SPEAKER_00: And even when you're not being hostile, there can be confusion, there can be sort of cultural misunderstandings that can lead to a perception of hostility.
• 5:37
  SPEAKER_00: So these games were asking us, first of all, to look for things like bullying and hate speech and harassment so that we could identify those sort of intentional ways of injuring another player and escalate that to the studio so that they could prevent it.
• 5:51
  SPEAKER_00: They also asked us to look for positive behaviors and be able to say, hey, this guy Seth is actually really good at coaching new players and helping them feel more welcome in this community.
• 6:01
  SPEAKER_00: Maybe when we have a new player looking for a partner, we should pair them up with Seth instead of that jerk Mike who just keeps, you know, trashing on them and taking advantage of the fact that they're not that skilled.
• 6:11
  SPEAKER_02: You're not looking for keywords, you're looking for intent and meaning.
• 6:15
  SPEAKER_02: So, how do you approach that problem when you're when you're you know when you're doing this?
• 6:19
  SPEAKER_02: And how do you account for that, the ways that humans actually speak, context, tone, and so on?
• 6:24
  SPEAKER_00: The simple answer is that there just is no shortcut.
• 6:28
  SPEAKER_00: You have to look at every single aspect of the audio in order to see all the things that a human being can see.
• 6:35
  SPEAKER_00: And so the way Modulate approaches this is we actually have what's known as an ensemble model that consists of over a hundred different component machine learning models.
• 6:45
  SPEAKER_00: Some of those models are looking at emotion, some are looking at the cadence and prosody of the speech, some are looking at the timbre of the voices, some are looking at, you know, particular higher-level characteristics like interruptiveness or whether the conversation feels scripted.
• 6:59
  SPEAKER_00: And of course, plenty are looking at the words that are being said as well.
• 7:03
  SPEAKER_00: But the the sort of first innovation from modulate standpoint is just having all of those models that are able to look at every possible sort of lens of the audio and make sure that we're not losing something valuable that was in that original speech.
• 7:17
  SPEAKER_00: The second big innovation is the ability to weave those all together.
• 7:21
  SPEAKER_00: So five years ago, you could have bought a sentiment analysis tool that could tell you, hey, this comment is sarcastic.
• 7:29
  SPEAKER_00: And you certainly could have bought a transcription tool that could tell you the comment was nice job, genius.
• 7:34
  SPEAKER_00: But if you tried to then pass that into a summary system, it would say, Mike complimented Seth's great idea.
• 7:41
  SPEAKER_00: Even though all the data was there, it would really struggle to connect those dots and understand that it really changes the meaning of the conversation.
• 7:49
  SPEAKER_00: It's not just sort of a fun fact that that happens to be sarcastic.
• 7:53
  SPEAKER_00: So that's the other work that we put in is combining these different data points together, we can recognize that that sarcasm poisons the meaning of the rest of the conversation in a certain way and reinterpret everything else that's being said through that lens.
• 8:09
  SPEAKER_02: That's great.
• 8:10
  SPEAKER_02: So talk about some of the other use cases.
• 8:13
  SPEAKER_02: So many times these uh platforms are relying on user reporting or after-the-fact types of review, right?
• 8:21
  SPEAKER_02: Even when you're gathering that data.
• 8:23
  SPEAKER_02: So, what changes when you can do this in intervene in real time?
• 8:27
  SPEAKER_02: Why is that shift so important for trust and safety and user retention?
• 8:33
  SPEAKER_00: Yeah, I mean, in in social and gaming spaces, it's important for two reasons.
• 8:37
  SPEAKER_00: One is that if you know, I'm actively harassing you right now, Seth, and the gamer platform is able to come in and put a stop to it, they are protecting your experience.
• 8:47
  SPEAKER_00: They're making sure that you don't have as negative an experience as you otherwise would have.
• 8:51
  SPEAKER_00: Also, for me as the bully in this scenario, getting that feedback fast drastically changes how quickly I learn.
• 8:58
  SPEAKER_00: If I, you know, am saying all this awful stuff to you, and then at the end of the match, I get an email 30 minutes later saying, Hey, by the way, we noticed that you did something wrong and we're gonna suspend your account.
• 9:09
  SPEAKER_00: That just infuriates me.
• 9:10
  SPEAKER_00: It doesn't actually lead to a change in behavior.
• 9:12
  SPEAKER_00: But when live in the conversation, I'm getting a notification saying, Hey, that's not okay.
• 9:17
  SPEAKER_00: This is our code of conduct, you need to behave this way, it actually just systematically changes behavior.
• 9:23
  SPEAKER_00: In Call of Duty, found that this kind of intervention led to about 8% drop-off of repeat user rate month over month.
• 9:30
  SPEAKER_00: So as we continue to run, we would just see the number of repeat users dwindle over time because they were frankly just learning more constructive ways to engage, or learning that being purely destructive was not welcome on that platform.
• 9:43
  SPEAKER_02: Just talk a little bit about shifting from the gaming world to the corporate world and what some of those applications and use cases are.
• 9:50
  SPEAKER_00: Yeah, absolutely.
• 9:51
  SPEAKER_00: So in the corporate world, the philosophy is still the same.
• 9:54
  SPEAKER_00: If something bad is happening in this conversation, I need to know about it immediately so that I can prevent it.
• 10:01
  SPEAKER_00: But now that something bad looks different.
• 10:04
  SPEAKER_00: It could be fraud.
• 10:05
  SPEAKER_00: You know, if you're actually talking to deepfake Mike right now and he's in the process of convincing you to give him access to his bank account, then you'd better get notified in real time that that's actually a deep fake.
• 10:17
  SPEAKER_00: So one of those hundred models in our in our ensemble is actually our own custom-trained deepfake model, which is number one on the worldwide leaderboards.
• 10:26
  SPEAKER_00: We're really good at this stuff.
• 10:27
  SPEAKER_00: Another use case is AI agent guardrails.
• 10:29
  SPEAKER_00: So if you're deploying real-time voice agents, there is a risk that those AIs go off the rails.
• 10:35
  SPEAKER_00: We spoke to some folks in a recruiting context who were trying to use AI interviewers, and they had prompted the AI interviewer to check if the candidate was flexible.
• 10:44
  SPEAKER_00: Um, and so what they found out was that the AI interviewer would sometimes ask the candidate to perform a yoga pose during the interview.
• 10:50
  SPEAKER_00: You genuinely can't make this stuff up, right?
• 10:52
  SPEAKER_00: It's but you need a system that is noticing in real time what's happening so it can escalate to you and you can take action about it because otherwise you just lost that customer.
• 11:03
  SPEAKER_00: You lost that candidate, and in that kind of case, maybe you're even in for a lawsuit, right?
• 11:07
  SPEAKER_00: You you need to be able to catch this stuff before it escalates so that you can actually resolve the problem.
• 11:13
  SPEAKER_00: So, one of our first ever customers in the corporate space came to us after seeing some of our work in Call of Duty, and they were a Fortune 500 company that works in food delivery.
• 11:23
  SPEAKER_00: And they started out saying, Hey, we just want to know if someone is harassing the pizza delivery driver.
• 11:28
  SPEAKER_00: You know, you're five minutes late, what the hell?
• 11:30
  SPEAKER_00: I'm gonna shoot you.
• 11:31
  SPEAKER_00: Unfortunately, that's a real kind of thing that happens, and that driver understandably says, I no longer want to go to this person's front door.
• 11:38
  SPEAKER_00: But when we deployed with them, we very quickly found, hey, we were actually detecting a lot of attempted scams against these drivers as well.
• 11:45
  SPEAKER_00: And those scams usually look like one of two things.
• 11:49
  SPEAKER_00: The first is what I would call sort of return fraud um or order fraud, where they're basically reaching out to the driver as they're on their way to the restaurant and they're saying, hey, the app charged me 200 bucks, but it only placed a hundred bucks worth of the order.
• 12:04
  SPEAKER_00: So you need to get the restaurant to give you the extra money or the extra food that I paid for.
• 12:09
  SPEAKER_00: And if you don't, I'm gonna force you to pay me a hundred bucks out of your own pocket when you show up here.
• 12:15
  SPEAKER_00: Um, and they kind of make that enough of a threat that the driver feels like this is legitimate and they do their best to resolve the issue.
• 12:21
  SPEAKER_00: The other is you actually save the driver's information and then you call them a few days later um and say, hi, I'm from the IT department of whatever this delivery company is, and I need to check a few details.
• 12:36
  SPEAKER_00: And now you're stealing social security numbers and doing all this horrific stuff.
• 12:39
  SPEAKER_00: It's a you know, it's always terrible, but especially to target these drivers who are, you know, just just trying to make a living.
• 12:46
  SPEAKER_00: So us being able to recognize and say, hey, the platform now with modulate has the tools to monitor these calls, not in an invasive way where we're, you know, telling you everything this driver says and does, but just so that if someone is actively trying to manipulate them, the platform has the information that they need to be able to say, oh, let's jump in and actually protect them in this case.
• 13:08
  SPEAKER_00: As you said, you know, voice voice is it's how we express emotion, it's how we understand emotion, and therefore, like emotion is a powerful way to move other human beings.
• 13:17
  SPEAKER_00: So when you're trying to manipulate someone, you rely on creating an emotional context that makes them sort of fit fail to meet their better judgment because they are trying to respond to those emotions.
• 13:29
  SPEAKER_00: So, examples that that we've seen sort of as I suppose new age fraud attempts to some degree, obviously deep fakes are in there.
• 13:37
  SPEAKER_00: But there's other things as well, such as hearing in the background a baby crying, and the fraudster is saying, Oh, I'm so sorry, my baby's really angry, I just need to get this done and trying to create that sense of urgency.
• 13:49
  SPEAKER_00: And it's a recording of a baby, and you can tell.
• 13:53
  SPEAKER_00: If you're just looking at transcripts, right, you're never gonna catch that that's a fake baby sound.
• 13:57
  SPEAKER_00: But if you know how to listen to the audio, you can actually say things like the echo of this voice is not matching the echo of the speaker at all.
• 14:04
  SPEAKER_00: They're in completely different territories, and therefore I know that this was faked.
• 14:09
  SPEAKER_00: Um, we're also seeing, you know, fraudsters because they can scale so much up with AI right now, they are scaling up their operations accordingly.
• 14:17
  SPEAKER_00: And so there's a greater need to be looking for repeat scripts.
• 14:21
  SPEAKER_00: You might see a call center.
• 14:23
  SPEAKER_00: Maybe you have, you know, five different call centers scattered across the country, and you've got a bunch of different agents in each of those call centers, and suddenly a single fraudster is able to place a bunch of automated calls that hit 20 of your agents all at the same time, and it's running the same script in the same playbook.
• 14:42
  SPEAKER_00: And their strategy is I bet I can find the one week link faster than you can shut me down.
• 14:47
  SPEAKER_00: And so recognizing things like the cadence of this conversation is off in a way that feels scripted becomes extremely valuable, even if the words seem legitimate, or being able to simply in real time cross-reference and say, Hey, we just noticed that in the last 30 seconds we have nine different people all claiming to be Seth.
• 15:07
  SPEAKER_00: That's a novel.
• 15:08
  SPEAKER_00: Something weird is happening here.
• 15:09
  SPEAKER_00: Not a lot of Seth's in the world, not that many.
• 15:12
  SPEAKER_00: So all of these different pieces sort of come together, and the the challenge is that fraudsters are good at their job, right?
• 15:18
  SPEAKER_00: Their whole job is to convince you not to follow your policies.
• 15:22
  SPEAKER_00: They will be very, very good at convincing you that they have a good reason you should not follow your policy in this case.
• 15:29
  SPEAKER_00: So the policy alone cannot protect you.
• 15:33
  SPEAKER_00: What you need is the ability to recognize that that reason they're claiming is not legitimate.
• 15:38
  SPEAKER_00: And catching these inconsistencies in the audio or these trends of sort of false behaviors are a really critical way to do that.
• 15:45
  SPEAKER_02: How should organizations think about guardrails for AI-driven voice systems, especially when they operate at speed and scale?
• 15:54
  SPEAKER_00: If you're deploying an AI voice system out there, then there's two critical things you need to be able to know about what that AI system is doing.
• 16:02
  SPEAKER_00: The first is when it is actually in the middle of a call and that call is going very poorly.
• 16:08
  SPEAKER_00: You need to be able to know about it in real time so you can jump in and do something about bring in a human or or redirect the AI system.
• 16:16
  SPEAKER_00: And that going very badly, each company will set their guidelines differently.
• 16:21
  SPEAKER_00: For some customers, some companies, a customer just saying, Hey, I'm I'm likely to churn, or I'm flying heavily that they're likely to churn, is a crisis in and of itself.
• 16:30
  SPEAKER_00: In other cases, it might only be when you have significant sort of regulatory liability if you mishandle this particular caller, if there's a large chance of fraud.
• 16:38
  SPEAKER_00: But you need some way in real time to know whether your AI system is doing something wrong.
• 16:45
  SPEAKER_00: And the challenge with that is that most AI systems cannot introspect.
• 16:50
  SPEAKER_00: They cannot check and ask themselves, is something going wrong here?
• 16:55
  SPEAKER_00: They're just doing their best at each moment.
• 16:58
  SPEAKER_00: And so you need something that is supervising them, that's sitting on top, that's able to say, hey, this is not what your actual goals in the conversation were.
• 17:06
  SPEAKER_00: The other thing that you need is if you're using AI agents, it's probably because you need them to be at large scale.
• 17:12
  SPEAKER_00: You want those AI agents to be handling a lot of conversations, which means you can't QA them all.
• 17:17
  SPEAKER_00: You can't review them even after the fact, and you can't learn things like key trends that are happening.
• 17:22
  SPEAKER_00: Um, so you also need a system that can digest all of those conversations, whether in real time or after the fact, and be able to say, hey, the number of callers who are frustrated about billing issues went up 300% in the last day.
• 17:36
  SPEAKER_00: Something is wrong.
• 17:37
  SPEAKER_00: We need to investigate that on the product side.
• 17:39
  SPEAKER_00: Those kinds of insights are critical as well.
• 17:42
  SPEAKER_02: You know, it's interesting when you start and when you look at how we interact these days when you have voice assistants and they're terrible, right?
• 17:49
  SPEAKER_02: They're terrible.
• 17:50
  SPEAKER_02: Their speech interpretation is incorrect.
• 17:52
  SPEAKER_02: Make mistake.
• 17:53
  SPEAKER_02: They don't understand.
• 17:54
  SPEAKER_02: I mean, it it blows my mind how bad they are.
• 17:57
  SPEAKER_02: But we know we're on a path for them to get better.
• 17:59
  SPEAKER_02: But you know, you get in endless loops and you end up with uh, you know, getting handed off from one voice assistant to another, right?
• 18:07
  SPEAKER_02: Uh, and and and one department to another, and finally you get to a human and they ask you to go through everything all over again, right?
• 18:14
  SPEAKER_02: So that's a really interesting uh thing.
• 18:16
  SPEAKER_02: So we're hoping that this is going to get better.
• 18:18
  SPEAKER_02: So when you think about this idea of being able to understand the safety, have those interventions, but not violate trust.
• 18:26
  SPEAKER_02: There's a fine line between safety systems and surveillance systems, right?
• 18:31
  SPEAKER_02: So, how do you think about designing voice AI that improves safety and trust, but doesn't cross that boundary that makes users and employees uncomfortable?
• 18:41
  SPEAKER_00: Yeah.
• 18:42
  SPEAKER_00: Um, privacy is uh is a really important and and ultimately very tricky balance here because, as you say, in order to protect what's happening in the call, you do need some way to listen to it.
• 18:54
  SPEAKER_00: In general, you should be collecting and certainly storing as little of that information as possible.
• 19:00
  SPEAKER_00: There is not usually a need to compile everything that is said across every single call.
• 19:05
  SPEAKER_00: At least that module, the way that we partner with folks around a problem like this, is we basically encourage lay out in advance what things you care about.
• 19:14
  SPEAKER_00: Whether it's toxicity in a gaming context or fraud in a banking context or AI misbehavior in an interviewing context, let's lay out in advance what the kinds of things are that you care about.
• 19:25
  SPEAKER_00: We can actually program that into the system from the get-go.
• 19:28
  SPEAKER_00: And then we are escalating to you the conversations where that kind of bad thing is happening so that you can take a step in and resolve it, rather than saying, here is a digest of every single call and where we're storing all of it.
• 19:44
  SPEAKER_00: And that that sort of methodology means that we end up collecting much, much less data, and we're only surfacing the stuff that these organizations were trying to get surfaced, anyways.
• 19:57
  SPEAKER_00: Um, in a gaming context, right, you were going to encourage users to submit those reports.
• 20:01
  SPEAKER_00: And now we're accelerating that process, but it's still only the stuff where something is actually going wrong.
• 20:07
  SPEAKER_00: In some place like a bank or a contact center, they might be required by law or at least sort of by their standard practice to collect all of that stuff anyway.
• 20:16
  SPEAKER_00: Everyone's heard the this call is being recorded for quality assurance.
• 20:19
  SPEAKER_02: And what we're doing training purposes, yes.
• 20:21
  SPEAKER_00: Exactly.
• 20:22
  SPEAKER_00: And what we're doing is sort of surfacing insights more accurately from that.
• 20:25
  SPEAKER_00: I don't expect any privacy diehards to be overly swayed by this next argument, but I will say if you're already collecting all of that data and your options are either we're going to have to pay a bunch of people to listen to it, or we can have the AI run through it, extract only the information that we care about, and focus our attention there so that we're not, you know, taking a look at any of that other stuff.
• 20:48
  SPEAKER_00: I would still argue that's actually a more privacy-safe approach.
• 20:52
  SPEAKER_00: There's sort of less, less exposure risk of other kinds of things that were not what the platform was looking for, suddenly getting surfaced or misshared.
• 21:02
  SPEAKER_02: One of the things I've been trying to train my some of my business development people on is uh is their voice when they're leaving voicemail messages, right?
• 21:10
  SPEAKER_02: And or when they're talking to people like not to use an upward inflection of your tone, like this is Seth early, right?
• 21:17
  SPEAKER_02: So something that conveys confidence.
• 21:19
  SPEAKER_02: I would imagine that sales training has got to be a really important uh uh function or capability around that emotional uh analysis and that voice and tone and pacing and cadence and all of those types of things.
• 21:32
  SPEAKER_00: Yeah, we we certainly talk a lot with organizations that are thinking about standard operating procedures and coaching, whether for their support agents or for their salespeople and how to navigate a conversation, take advantage of the fact that voice is again full of more than just text.
• 21:49
  SPEAKER_00: I think one of the challenges people run into is you can write in your standard operating procedures, like be cheery and pleasant.
• 21:55
  SPEAKER_02: And that's very hard.
• 21:57
  SPEAKER_02: Not too cheery and pleasant.
• 21:58
  SPEAKER_00: Well, first of all, yes, specify.
• 22:00
  SPEAKER_00: It is really hard.
• 22:01
  SPEAKER_00: And then also, like, not everyone has the gift of reading those words and saying, okay, I'm cheerful now.
• 22:07
  SPEAKER_00: Right.
• 22:07
  SPEAKER_00: Like it's hard to fake that kind of emotion.
• 22:09
  SPEAKER_00: One of the things you can do with like a real-time emotion-aware AI is actually provide that real-time feedback.
• 22:15
  SPEAKER_00: Um, and too much of that gets very heavy-handed and very difficult to field.
• 22:20
  SPEAKER_00: Being able, especially when you're sort of navigating different cultures, to be able to say something like, Hey, you don't realize that this person is actually offended right now.
• 22:29
  SPEAKER_00: The thing that you said that you thought was a joke actually upset them.
• 22:32
  SPEAKER_00: You should, you know, express a bit of contrition around that.
• 22:35
  SPEAKER_00: Or, hey, this person is expecting you to be presenting in this way.
• 22:39
  SPEAKER_00: And if you're not presenting that way, that's going to be an issue.
• 22:42
  SPEAKER_00: And actually, in some cases, even provide real-time calibration.
• 22:46
  SPEAKER_00: So you could say something like, Hey, how happy are we?
• 22:49
  SPEAKER_00: Or on the flip side, how angry are we?
• 22:51
  SPEAKER_00: I've spoken with some contact centers that are looking at actually rating the sort of aggression level of the caller so that they can say, Hey, agent, you just had to deal with a brutal call.
• 23:02
  SPEAKER_00: Maybe take a break, maybe unwind for a second, because that was sort of unfairly bad for you.
• 23:07
  SPEAKER_02: Or route that to your black belt agent, right?
• 23:10
  SPEAKER_02: Who knows how to deal with angry uh people.
• 23:12
  SPEAKER_00: Yeah, I would certainly start by being really clear about the difference between voice like voice analysis AI and voice agents.
• 23:21
  SPEAKER_00: And I think right now there's a lot of confusion where people are just saying voice AI for both, and both have their place.
• 23:29
  SPEAKER_00: Voice agents can be an enormously helpful way to support your customers at scale to provide different kinds of value to them.
• 23:35
  SPEAKER_00: But as I mentioned earlier, they cannot introspect.
• 23:38
  SPEAKER_00: They will be suckered in by all kinds of fraudulent and manipulative activities.
• 23:44
  SPEAKER_00: They will not notice when a call is going wildly off the rails.
• 23:48
  SPEAKER_00: They will not be able to prevent themselves from having an issue.
• 23:53
  SPEAKER_00: That's not their purpose.
• 23:55
  SPEAKER_00: Their purpose is to respond to the caller in the way that they understand is best.
• 23:59
  SPEAKER_00: On the flip side, analysis AI cannot talk to you.
• 24:05
  SPEAKER_00: We are not building an agent.
• 24:06
  SPEAKER_00: There's nothing for you to talk to.
• 24:08
  SPEAKER_00: It is just a supervisor.
• 24:11
  SPEAKER_00: But that supervisor is enormously accurate and allows you to deploy voice agents with much greater confidence.
• 24:18
  SPEAKER_00: So the first thing to anyone building this is make sure that you're really clear on what thing you're trying to do and what your actual KPIs are.
• 24:26
  SPEAKER_00: If you have, for instance, a contact center, you're probably measuring CSAT customer satisfaction.
• 24:31
  SPEAKER_00: Deploying any kind of voice AI should not be driving CSAT down.
• 24:36
  SPEAKER_00: You should be building out your analytics suite first to make sure that you are able to tell as you're deploying that AI system, is it actually decreasing CSAT, which is common.
• 24:47
  SPEAKER_00: A lot of people have AI fatigue these days.
• 24:49
  SPEAKER_00: Um this is all common sense, right?
• 24:52
  SPEAKER_00: I'm I'm not saying anything revolutionary, but it is really important.
• 24:55
  SPEAKER_00: If you're deploying a system, make sure you know why and make sure that you can actually validate that it's solving the problem you think it is.
• 25:02
  SPEAKER_02: Well, Mike, this has been a fascinating conversation.
• 25:06
  SPEAKER_02: You've really uh made it clear that voice is not just another input channel, uh, and it is a very uh high, highly it's a very complex and high impact environment to apply AI.
• 25:17
  SPEAKER_02: And especially when you're talking about trust and safety and and uh fraud prevention, and those all are things that are on the line.
• 25:23
  SPEAKER_02: So, you know, I really want to thank you for your time and and appreciate your insights.
• 25:27
  SPEAKER_00: Well, thank you very much, Seth.
• 25:29
  SPEAKER_00: This was a great conversation, and I'm glad to have had the chance to talk through it all.
• 25:32
  SPEAKER_02: And and thank you to our listeners for tuning in.
• 25:35
  SPEAKER_02: Stay with us for more conversations about how AI is trans transforming business technology and the ways organization operate in the real world.
• 25:42
  SPEAKER_02: And we will see you next time in the next episode of the Early AI Podcast.
• 25:48
  SPEAKER_03: Thank you for joining us on another deep dive into AI innovation.
• 25:52
  SPEAKER_03: Tune in next time when we introduce another industry expert and discuss how to maximize AI in your world.
• 25:58
  SPEAKER_03: The Early AI Podcast is sponsored by Early Information Science and Vector.
• 26:04
  SPEAKER_03: To learn more, visit early.com at E-A-R-L-B-Y.com.
• 26:09
  SPEAKER_03: Thank you for tuning in.