3 Reasons You Need AI Voice Agents in Your Business - EP 014

00:00:02: Hello and welcome to this week's episode of the AI Boardroom.

00:00:09: And yeah, so Lana, it's been a minute, summer break.

00:00:13: Yeah, there's been a minute, but we're back into it.

00:00:17: So here to talk about AI voice agents.

00:00:20: So what are they?

00:00:21: What are some challenges organizations are experiencing by introducing them and some

00:00:26: different applications?

00:00:27: So we're going to go through

00:00:29: couple of practical business application, how different business verticals

00:00:34: industries are using them.

00:00:35: So I'm really excited about this one because yeah, I think it's been trending a

00:00:40: lot, especially with the voice agents and more and more organizations.

00:00:44: I think more in the US than here, like here people still trying to get chatbots

00:00:47: working.

00:00:50: And here I think he means Germany.

00:00:53: So we are kind of on two different sides of the world.

00:00:57: Yeah, I think the voice agents, I think, as the voice technology is becoming better

00:01:03: and better, and we're kind of seeing it from different angles, music is now using

00:01:09: it.

00:01:10: I think there's, you know, Chad GPT came out with, what is the name of...

00:01:16: Wanted to come out with like, so still...

00:01:18: Want to, I'm sorry.

00:01:20: Still trying to, yeah.

00:01:23: But I think that the voice technology is really...

00:01:26: kind of improving, which, you know, kind of makes it very lucrative for

00:01:31: organizations to be like, Hey, is this something that is ready now for us to

00:01:36: replace agents, kind of the call center agents and other places, which, yeah,

00:01:42: we'll talk about more, but do you want to just dive in and maybe let's, let's start

00:01:46: with defining what are voice agents?

00:01:48: Yeah.

00:01:49: So, yeah, basically, like the, the general definition is.

00:01:55: that we have some input that we usually get as text, which we now get as voice.

00:02:02: So we can have a call or a conversation, yeah, in a natural manner with our voice

00:02:09: instead of like typing stuff in.

00:02:11: So that's the basic idea.

00:02:13: And we basically have technically like two ways to approach it.

00:02:16: So we have the first way would be to actually take what we do in our chatbots,

00:02:23: take the text.

00:02:25: just synthesize it and yeah, give back the synthesized audio so that you, instead of

00:02:33: giving back the text, you speak.

00:02:35: This has several challenges, which we will go into in the next part, but that's

00:02:43: basically the one thing.

00:02:45: The new thing that's coming around now is that the LLMs getting multimodal, so they

00:02:53: don't...

00:02:54: take the input text, get text out and you synthesize it, but you give them directly

00:03:01: the audio feed and they spit out audio on the other side.

00:03:06: So that's basically the idea so that you don't need to go through all the synthesis

00:03:13: tasks, which all costs time, but you're able to get direct voice output.

00:03:20: That said.

00:03:22: It's still really early and we basically don't have a model yet, which makes this

00:03:27: really applicable.

00:03:29: There was one released this week.

00:03:31: I think it's called Mosh or something, which open source such a technology, but

00:03:38: the model itself is pretty, yeah, pretty early stage.

00:03:42: So, but yeah, it's getting through to be actually usable.

00:03:46: But I think through this episode, we might lean more on the classical way of

00:03:52: transforming the text.

00:03:55: But yeah.

00:03:56: Yeah.

00:03:56: And I think with another notable difference, because what's different

00:04:00: between the phone agents that you can have the robotic voice answering the 1 -800

00:04:05: number and this AI kind of voice agents we're speaking about is because there's

00:04:11: some intelligence behind it.

00:04:12: So the traditional 1 -800 number like press dial two for this option, dial three

00:04:19: for this option are very much rule -based.

00:04:22: types of systems.

00:04:23: So there's predefined script.

00:04:25: So someone like literally creates like a tree of decisions and tree of different

00:04:30: tasks for this robotic voice to, to voice out.

00:04:34: And it's very constant with AI voice agents.

00:04:38: They are tied to some intelligent systems.

00:04:40: I think here you're referring to large language models that are able to pivot.

00:04:45: So there's not a predefined script.

00:04:47: So depending on the context and you can be very

00:04:50: You can be very interactive with this agent in a way, so you don't have to say

00:04:55: things in a particular way, enunciate them in a specific way without an accent.

00:05:00: And it can pivot, not based on the predefined script or tree of decisions.

00:05:05: It can actually go different paths.

00:05:09: And sometimes instead of taking you to another agent, like a live agent, it can

00:05:13: answer those questions live based on the knowledge base that it kind of has access

00:05:18: to.

00:05:19: Yeah, I think that's also a differentiation factor that I think is

00:05:24: important to highlight.

00:05:26: Yeah, definitely.

00:05:29: So, yeah, that being said, I think we should directly switch to the challenges

00:05:35: topic.

00:05:36: So, yeah, like I said before, so we have some technical, some also, yeah, like soft

00:05:47: skills stuff.

00:05:50: Talking about the technical stuff.

00:05:52: So one major thing that really breaks the immersion is always the latency.

00:05:59: So if I say something and I need to wait seconds without any response from the

00:06:05: system, it makes the waiting awkward.

00:06:11: And yeah, it kind of feels like you're going a step, waiting, then going the next

00:06:16: step, waiting.

00:06:17: So yeah, that's basically where you're at when you have too much of latency.

00:06:23: Which is not natural, I think.

00:06:25: And that's why I think just to kind of give an example, it's like, you and I are

00:06:29: kind of talking and then I'll ask you a question and you're like, silence

00:06:34: processing is just very unnatural because we're not really sure.

00:06:38: Is it because you don't really have that immediate feedback, right?

00:06:43: Like, did the understanding of my question and processing is a...

00:06:48: Am I going to get hung up on?

00:06:50: And so I think that inference is the reason why it's important.

00:06:53: That delay in getting a response is because it's just not natural for us to

00:06:59: wait that long.

00:07:01: And it's actually quite off -putting to customers to have to wait for a response.

00:07:07: And it's also even more awkward because they don't have a face.

00:07:10: Now we do, of course, a podcast, but we also have a video feed to that.

00:07:14: So if you're talking

00:07:17: And if you ask me something and then you wait for an answer, you see that I'm

00:07:22: thinking actually.

00:07:23: So we get some feedback.

00:07:26: If you have only a call, you might ask yourself, like, are we disconnected?

00:07:32: Is there something wrong?

00:07:33: Like, did you not understand?

00:07:35: Because you don't get any feedback because that's what you have to work with.

00:07:39: There are mitigation strategies.

00:07:43: So you could.

00:07:44: I spoke to a company, they actually, that's their main product is call services

00:07:49: automated through cloud systems.

00:07:55: And they, like you said, had basically the old call screen stuff where just like step

00:08:01: -by -step goes through something.

00:08:02: And now of course implement AI.

00:08:05: And what they did is they tried to first reduce the latency as far as they could,

00:08:09: but they still have several seconds delay.

00:08:12: And what they do is like,

00:08:13: on a random basis, they put in like arms and type, even typing, even like typing of

00:08:24: a keyboard or something to simulate that someone's searching up something in the

00:08:29: system.

00:08:30: So, and latency stays the same, but the perceived time of waiting is a lot more,

00:08:37: it's a lot more enjoyable to the customer because he then has some feedback.

00:08:42: It's the same if you're on the website, you click a button and nothing happens,

00:08:45: even if the actual action gets through.

00:08:50: It's just a bad experience.

00:08:52: And if you don't have a video feed, it's the same on audio.

00:08:56: If you wait without any feedback, it's just a bad experience.

00:09:01: I like that.

00:09:01: I actually, I can appreciate them thinking kind of going beyond, above and beyond.

00:09:06: And you think again, what we talked about previously about ChadGPT releasing their

00:09:10: kind of voice agent.

00:09:13: They did, I think, include the ums and ifs.

00:09:16: And so when even if there was some influence...

00:09:20: But that was actually the model.

00:09:24: That was not to mitigate time delay.

00:09:27: They actually did all the...

00:09:29: okay.

00:09:29: It was part of their personality.

00:09:31: Yeah, actually.

00:09:34: And that's what they did.

00:09:37: They removed all the processing between by...

00:09:40: the model the ability to produce audio, to like get audio and to produce audio

00:09:45: without any intermediate steps.

00:09:47: Same for images, by the way.

00:09:48: But we still don't have access to the full model.

00:09:51: So I cannot give too much detail about that.

00:09:54: But that's basically what they do.

00:09:56: They created a multimodal model, which means basically get different modalities

00:10:02: in and get different modalities out.

00:10:04: You could also...

00:10:05: in the future, for example, say you get audio in and get an image out.

00:10:09: So that's the beauty of multimodality.

00:10:12: So yeah, and that's actually why they got this natural conversation to that point

00:10:18: because they reduced latency to milliseconds.

00:10:23: And well, which you have a natural conversation and that also, and that's the

00:10:30: next challenge, you cannot interrupt.

00:10:34: So

00:10:35: Think about you take text, you put it in a machine and wait for the audio file and

00:10:40: then you can play the audio file.

00:10:41: In all of the process, if someone's interrupting you, it's not easy to react

00:10:48: to this because you play the whole audio file.

00:10:52: You might interrupt the audio file, but it will be even more natural than the

00:10:57: conversation was anyway because you have a lot of disturbances.

00:11:01: And if you do the new GPT -4 .0 way, or I think Google Gemini also wants to

00:11:07: introduce this, you're actually able to interrupt the model and it reacts

00:11:12: naturally to the interruption like a human would.

00:11:16: And that's kind of a big win on their side too.

00:11:21: Yeah, and I think we talked about just to drive home that kind of how it works

00:11:26: behind the scenes.

00:11:28: So it's almost like a radio.

00:11:29: I don't know if you've ever...

00:11:31: when you had a child, when you were a child had these radio things that you

00:11:35: walked, you know, across different rooms and you're like, Hey, can you listen to

00:11:38: me?

00:11:38: You have to press the button down.

00:11:40: Like a walkie talkie type of thing.

00:11:42: Yeah.

00:11:42: And then you press down and then the other person can listen when you talk, but they

00:11:47: can't press that, that thing and you hear it then vice versa.

00:11:52: And so that's typical.

00:11:52: That's basically the state of voice agents currently that are again, apart from some

00:11:57: of these newer technologies that aren't yet.

00:12:01: to the public, it is very like take turns and wait until kind of you're done with

00:12:09: your part and it stops.

00:12:11: And then it takes their part.

00:12:13: Like basically AI takes time to actually process and return a response for you to

00:12:19: take your turns.

00:12:20: But what I think you're referring to is with open AI, you can actually interrupt

00:12:26: mid sentence and

00:12:29: have the AI system pivot basically in a different direction without completing

00:12:34: that specific task, which I think could be game changing actually.

00:12:38: Definitely.

00:12:39: Yeah.

00:12:40: I don't have any info about how long can a conversation be before it reaches context

00:12:46: limits and stuff.

00:12:48: But I think they will stick to the 128 ,000 tokens they have right now, which to

00:12:55: give you some point of reference.

00:12:59: The first book of Harry Potter from the text tokens would be 117 ,000 or

00:13:08: something.

00:13:09: So you're saying it's a book of Harry Potter, basically?

00:13:12: The context window.

00:13:14: So it might be different for audio tokens.

00:13:18: I don't know.

00:13:20: But yeah.

00:13:20: That would be awesome.

00:13:22: Because I would imagine that companies have...

00:13:26: either voice scripts or policies that are hopefully less than a Harry Potter book.

00:13:34: So I can give you some promise that it would in the future, sometime in the

00:13:40: future when some of this technology, especially from OpenAI is released.

00:13:43: But I think that there's some of this already available and again, but we'll go

00:13:46: into use cases, but it's already being used.

00:13:49: It's also not far off.

00:13:51: So we actually will get access to these technologies pretty soon.

00:13:55: So yeah.

00:13:56: We will also, if any updates might be delivered on what we talk about today,

00:14:03: then we record another episode.

00:14:05: Promise you that.

00:14:07: One challenge I want to emphasize on that note, which also differs the direct

00:14:13: system, which was upcoming and the actual system that you could implement today, is

00:14:20: emotions.

00:14:21: So if you transform audio to text,

00:14:26: You most you use like one channel of a conversation, which is like the the

00:14:31: subtext, the emotions, the tension in the voice, stuff like that.

00:14:35: You completely lose that.

00:14:39: And you could go ahead use like there's a human eye, I think they have they focus

00:14:46: only on emotion detection in like basically every modality.

00:14:50: But that would add latency to your response.

00:14:54: So

00:14:56: You really have to weigh how much value it delivers to you before you actually use

00:15:02: stuff like that.

00:15:05: Could be also interesting in something like a chatbot to actually evaluate the

00:15:11: emotions to improve your answer.

00:15:17: But yeah, that's something that if you process the audio directly, which the new

00:15:21: GPT -4 .0 chatbot does, for example, it reacts and it also changes its voice based

00:15:29: on what your emotions are.

00:15:31: While today, it's really a manual process, which is kind of tedious to implement.

00:15:38: And even if you got it to a point where it's kind of working, you might run into

00:15:43: some issues.

00:15:44: So...

00:15:45: If you have something where a lot of emotions would be needed to be tracked,

00:15:50: hotline for first aid for people with any problem, then the where emotions have a

00:16:02: big part in understanding the people, the human on the other side.

00:16:07: You might wait for newer technologies because or add latency to the pipeline.

00:16:14: by knowing you could deliver a better response, still, yeah, you have to

00:16:19: evaluate if there is a need of processing the emotions in the message.

00:16:27: Yeah.

00:16:28: So these agents can get then quite complex, because I think if you're

00:16:32: accommodating by filling in the gaps or the ums and the thinking or...

00:16:42: whatever, and then considering emotions and things like that.

00:16:46: So there's multiple factors.

00:16:48: I don't know that they're challenges, but I think they're really important

00:16:50: considerations for how do you actually deliver the best customer experience for

00:16:55: your target use case?

00:16:57: And I think that's where we're probably headed in next is like, how are these

00:17:01: organizations and like, what are they using these agents for?

00:17:04: And it's, again, best AI is aligned to the business objectives and cater to your

00:17:10: customers.

00:17:10: And I like that you've

00:17:11: given that example about kind of emotional state may be important for like hotline

00:17:18: for like mental health or I don't know, some other kind of response system where

00:17:26: emotions are really part of your business objectives like to detect them.

00:17:30: And sometimes that's the reason why you have voice agents 24 seven to basically

00:17:35: the one that better detects them than a human.

00:17:38: But there are, again, like emerging approaches to actually help with some of

00:17:42: that.

00:17:43: But just talk of to kind of dive into some of these other use cases that we're seeing

00:17:48: a lot of and you probably have heard of.

00:17:50: And I think this one's an exciting one that I've kind of heard just to give you

00:17:55: an example of how else you can be thinking about agents as sales.

00:18:00: So I saw a recent use case and I think some demo, I want to say in the last

00:18:04: couple of weeks where

00:18:06: the voice agents are actually used to do a follow -up post bookings that are made on

00:18:12: the website.

00:18:12: So think real estate, building new communities, you fill in the form, you

00:18:17: wanna talk to an agent, if it's over the weekend, then chances are the agents are

00:18:22: not working, they're taking time off.

00:18:25: And there's research, and don't quote me on this, but there's research that states

00:18:30: that if you don't return a call, the better conversions are...

00:18:35: best within a certain time period.

00:18:37: So you get the best conversion if you have a return call within, let's say, the first

00:18:42: hour, and then it declines after that.

00:18:45: So if you do wait two days before reaching back to that client, the chance of that

00:18:50: client converting to a paying customer is less and less.

00:18:54: And so voice agents are now used to do follow -ups.

00:18:57: So when someone completes the forms online and within minutes, sometimes you can

00:19:03: even...

00:19:04: put a timer on it, within five minutes, they'll return a call asking you questions

00:19:10: that clarify your responses.

00:19:12: So typically, like these forms on the website, so typically like legions, so

00:19:16: they're very light, you're probably asking them for a name, phone number and email,

00:19:20: something light.

00:19:21: So this agent is used to then collect additional information so that when a live

00:19:27: agent does call to do this relationship building, they have more data to then look

00:19:32: into.

00:19:33: maybe that customer's background, look into options and things like that.

00:19:37: But it also gives you a touch point.

00:19:40: So instead of just having this kind of informal conversation with the website,

00:19:46: you now have talked to an agent.

00:19:48: So that's another touch point for the company.

00:19:50: And then it gives you information to build a better relationship the first time an

00:19:54: agent, like a live agent, talks to that customer.

00:19:57: So again, you're kind of building up that relationship through these multiple

00:20:01: touch points, but the business objective there, especially being kind of in sales,

00:20:06: real estate, again, you want to get more people converting.

00:20:11: And so for their perspective time, the response time for that follow -up is

00:20:16: really important.

00:20:17: And that's where kind of these voice agents come in and really deliver on that

00:20:22: value for them.

00:20:23: So.

00:20:24: Yeah.

00:20:26: I think you mentioned something important.

00:20:28: So you have background information of the customer.

00:20:30: And the good thing is if you have an agent that might do the call for you, you can

00:20:35: collect a lot of context upfront.

00:20:37: So you don't have to do it during the call.

00:20:41: So on every new lead that comes in, you try to collect as much information as you

00:20:46: can.

00:20:46: Before I had to take seconds with an AI system built for that.

00:20:50: And then you already have something you could combine with a call script.

00:20:55: By the way, it's one of the first things I did when I was doing my first blog article

00:21:03: on AI.

00:21:04: I took HubSpot call scripts and made a pipeline that if a call center agent is

00:21:14: talking, that the system would receive the message and based on the call script give

00:21:18: you the correct answer, or like the best answer by the synthesis of

00:21:24: what's incoming, what we found in the vector database.

00:21:31: And I was hugely impressed because I got the latency of that down to like one and a

00:21:36: half seconds.

00:21:36: And it was just the first try.

00:21:39: I didn't even optimize anything.

00:21:42: So yeah, that's basically something you could apply to all your AI approaches.

00:21:48: Like what can I do beforehand to then make

00:21:53: make the actual answer quicker and better at the same time.

00:21:58: Yeah.

00:21:59: And I think it's, again, I think the beauty of these AI agents and what you're

00:22:02: kind of talking about is having that context, but also having more of a human

00:22:07: interaction because of this intelligence that is built into it.

00:22:11: And so this make it feel, again, more research that says on this, I think people

00:22:16: acknowledge that it's not a realistic conversation, but...

00:22:21: it is better than the voice prompting and they still find it useful to give

00:22:27: information so that they're not wasting time later.

00:22:29: And they're still kind of in that state of providing more information.

00:22:33: They're curious, they've just been on your website, so it's top of mind for them.

00:22:37: So they are kind of in that prime state to give information rather than, again, if

00:22:41: you call them two days later to follow up with the live agent, they're like, I

00:22:45: forgot why I was searching your website.

00:22:48: Nevermind, right?

00:22:49: kind of capture them in the state and have that conversation, even if it's not the

00:22:54: most ideal.

00:22:55: But people, again, are willing to interact.

00:22:58: And again, some of the early research with these voice agents saying that people are

00:23:04: seeing value in those conversations, even if they are not, again, probably the

00:23:11: OpenAI level yet.

00:23:14: Yeah.

00:23:15: But still, I think that there's

00:23:16: some really exciting use cases and sales, real estate that you can explore with

00:23:23: voice agents.

00:23:25: Yeah, I think customer service is another one that we're seeing.

00:23:29: I don't know if you kind of have any use cases that you are kind of specifically

00:23:35: thinking of.

00:23:36: So I have a use case I would have wished I could have.

00:23:40: We were flying with kids and you can

00:23:46: you're allowed to take one seat per kid, like a car seat and one stroller.

00:24:00: I was like, are there any restrictions?

00:24:02: How does it work?

00:24:04: I wasn't getting any proper information.

00:24:09: But just like, yeah, you can.

00:24:13: You can do stuff.

00:24:16: Because usually when you fly you have like a billion restrictions like for everything

00:24:21: and everything's like and I just would have wanted to have an AI which has a

00:24:25: proper database Talking about data Which has proper database of explanations maybe

00:24:33: an fhq whatever Who could could just answer my question in a quick way?

00:24:40: Yeah, and I would have preferred the voice agent

00:24:45: over a chat at that point, because I wanted a quick answer.

00:24:50: And all this chatting takes more time than me just speaking it out and getting an

00:24:55: answer within seconds.

00:24:56: It's just faster than typing.

00:24:58: So it's generally, I think for customer service, it's even the better use case,

00:25:02: because it's not even close to being as slow as the old calling systems are.

00:25:12: And it saves time compared to

00:25:15: compared to stuff.

00:25:16: And you don't have to get as much information.

00:25:18: Like you just have to get like standard information from a vector database, which

00:25:21: is pretty quick, handed over to the AI, synthesize the stuff and go out.

00:25:26: So I think customer service is more often than not pretty straightforward because

00:25:33: like most of the things that are asked are pretty simple and mundane.

00:25:40: So yeah.

00:25:41: I think I am looking forward to add it to our own customer service solution too.

00:25:46: So yeah.

00:25:47: And I think one thing about these, just kind of to drive home on the customer

00:25:52: service, I do not, I hate talking to chat bots on websites.

00:25:59: Like hate is a strong word, but I do because of like, it is so trial and error

00:26:03: and that is very scripted, right?

00:26:05: Even like those agents are the equivalents of like these.

00:26:08: prompt systems.

00:26:09: And so the chances of you even going through the agent that is not AI enabled

00:26:14: to get a response.

00:26:15: It's like the success rate is so low that I don't even try.

00:26:20: And so I usually like my default is like you, and especially in circumstances where

00:26:24: you have your child or children, you're traveling, like you're not going to like

00:26:28: have a stroller and like a kid and then start a type questions and interact with

00:26:33: the chatbot.

00:26:34: Like for customer service, I think it is important

00:26:38: for you to deliver that quick response systems.

00:26:41: And so usually for me, I'm like, okay, what's my success rate if I call this 1

00:26:45: -800 number and press 0 -0 -0?

00:26:49: Until I get to the right one.

00:26:51: Yeah, even if you think about it, even if you can just have an AI system, understand

00:26:55: the request and route the customer to the right endpoint.

00:27:01: So even that would already be a huge improvement on what we have today in most

00:27:05: of the systems.

00:27:08: So yeah, to make me happy and press less of the zero buttons.

00:27:13: Because I do, like literally, I just need to speak to an agent that I just need a

00:27:16: quick response instead of navigating all of these kind of prompts.

00:27:22: If it accepts voice input, I always like directly try to human agent, human agent,

00:27:29: human agent until it just gives up and gives it, puts me through.

00:27:33: Exactly.

00:27:33: Yeah.

00:27:34: And I think a lot of it actually is like when you do these prompt systems, I don't

00:27:38: know if people, I guess those who are listening who operate, you know, customer

00:27:42: service or call center agents, a lot of these promptings are often used for

00:27:47: analytics.

00:27:48: They're not actually helping you to divert to multiple different centers.

00:27:54: Sometimes they do.

00:27:55: So if you have like a really, really huge multi -product service kind of

00:27:59: organization, you have no...

00:28:02: for example, that way.

00:28:04: How many times I've called like maybe smaller offices and you have like 500

00:28:08: different things, different buttons and like things, decisions to navigate and you

00:28:13: still land on the same person at the end.

00:28:16: And it's like, so I do think that in some cases, it is mostly used for analytics to

00:28:23: understand like staffing needs.

00:28:24: So if there's more types of questions that are used in like one part of the, you

00:28:29: know,

00:28:31: people are pressing more of these functions.

00:28:33: They're like, maybe we need to staff more people or educate our people on these

00:28:37: types of questions and then have very nuanced staff.

00:28:41: But sometimes if you do have a call center that is across different products, so for

00:28:46: example, you need specialized staff in those.

00:28:49: But yeah, it's just not really helpful.

00:28:51: And sometimes I would rather just go to the general agent, who is the router for

00:28:57: them all, who can answer my question.

00:28:59: then go through these prompts.

00:29:01: And sometimes again, like this is where I think an AI agent, like why can't the

00:29:05: person when they call the 1 -800 number hear a voice speaking and ready to answer

00:29:12: a question?

00:29:13: If it's at least somewhat conversational and not like a predefined script, it

00:29:20: already improves the customer experience by a lot.

00:29:24: And with that customer satisfaction and with that maybe also

00:29:30: reduced churn, whatever.

00:29:32: So, yeah, one thing I also want to add is you can be multilingual from the get -go.

00:29:39: So, at least like in the main languages, I'm not sure how the whole voice synthesis

00:29:44: works with like smaller or like not as much used languages, but if you have

00:29:50: French, English, German, it's pretty much solved.

00:29:54: So, yeah, if also think about like

00:29:57: Do you have an international customer base or even for like internal stuff?

00:30:01: Do you have different subsidiaries of your company which use different languages?

00:30:06: Maybe a centralized knowledge management could be done by a voice agent, which is

00:30:13: then also like taking the English text and just translates it properly because that's

00:30:18: something that like large language models do basically perfectly.

00:30:22: So, yeah.

00:30:23: Yeah.

00:30:24: And I think just to demonstrate like,

00:30:26: Maybe another use case that I think I'd love to kind of dive in.

00:30:29: So right now we've kind of talked about cutting the response rate and then

00:30:33: responding to customers immediately to close out the rate.

00:30:37: So that was like one objective, business objective that you want to do and use

00:30:41: voice agents for.

00:30:42: Now we talked about improving customer experience.

00:30:45: How do you actually deliver the best experience to your customer to increase

00:30:49: satisfaction, your kind of rankings for the organization.

00:30:55: But and then the other extreme I would say like you don't have enough staff and I

00:30:59: think a good example that hits home is within healthcare You do have you know

00:31:04: specialized clinics?

00:31:05: Let's say and you have higher demand than you have the supply of agents and in this

00:31:11: case You know where the people who are behind the phone often especially in these

00:31:14: specialized clinics or in taking appointments or who are in taking

00:31:19: information from Patients they there's not enough of them

00:31:23: So in my experience, again, if you do have specialized and very nuanced clinics, you

00:31:29: just, they're often operated by nurses.

00:31:32: Right now there's a nurse shortage kind of behind it.

00:31:36: It's just not enough people to answer the phone because they have to be highly

00:31:42: skilled, trained, and you know, you can't, like the ramp up time for those types of

00:31:47: workers is not that quick.

00:31:49: So how can you augment the,

00:31:52: you know, these nurses, what part of their workflow or data collection can you take

00:31:57: in order to leave just kind of, again, the nuanced work that you do need the nurse

00:32:02: for in order to validate?

00:32:05: So like there are use cases where, you know, you can even do patient validation.

00:32:09: So for example, like you can hook up these voice agents to APIs behind the scenes to

00:32:15: validate the customer, you know, or validate certain conditions or validate to

00:32:21: data that's in electronic health records, for example, validate to the appointment

00:32:27: type that they're seeking, collect some information, determine the appointment

00:32:31: type, and then maybe can go as far as like actually scheduling.

00:32:35: So all of these tasks, you can't necessarily delegate because there's some

00:32:40: nuance and like some training that is involved, like for you to determine like

00:32:43: the types of visit that patient is looking for.

00:32:47: So for example, if it's like a cancer clinic,

00:32:49: and they need to schedule a biopsy and then they need to schedule like lab

00:32:53: results.

00:32:54: There's some logic to the way that these nurses are organizing.

00:32:59: And then there's timing between the last labs and the new lab that you can't

00:33:04: necessarily just apply a rule -based kind of algorithm.

00:33:07: There's just like validation that happens.

00:33:10: And that's why, again, these highly specialized nurses are operating these

00:33:13: phones.

00:33:14: What happens when you don't have enough of them?

00:33:16: Do you turn patients down?

00:33:17: Yeah.

00:33:18: And do you really want your nurses to do phone work?

00:33:23: Like that's, I think that's the most important stuff.

00:33:26: So like they have to do healthcare, like why should they do the phone job?

00:33:31: So, of course you can't like to be, to be completely real here, you cannot have a

00:33:39: hundred percent solution, but would it benefit you to free up 80 % of your stuff?

00:33:44: Yeah.

00:33:45: I guess so.

00:33:47: Exactly.

00:33:48: Especially if it's...

00:33:49: And it's also maybe fulfilling a shortage, right?

00:33:53: So you could be taking on or scaling and taking in more clients or helping more

00:33:59: customers if you had basically more of that supply, which you can again, fulfill

00:34:05: with voice agents.

00:34:07: So it really goes back and I think we talk about this a lot, but it comes back to

00:34:10: what problem are you trying to solve?

00:34:13: I think voice agents are a great way to scale your customer service operations,

00:34:19: like kind of behind the scenes work, whether it's to follow up, to provide

00:34:23: better experience or to augment your existing people to like free up their

00:34:28: time.

00:34:29: I think they're really great use cases.

00:34:31: Definitely.

00:34:33: Cool.

00:34:33: I think we had a lot of use cases now.

00:34:37: And I think general application of voice agents got through pretty

00:34:43: pretty well.

00:34:44: So yeah, I hope you out there are the same.

00:34:50: And yeah, enjoy this episode as much as we did.

00:34:54: Because, Lana, we already got to the end.

00:34:57: Yeah, I'm flying by it's like ridiculous.

00:35:00: Like it's always the same on the podcast starts the first 10 minutes you think,

00:35:04: what should we talk about like in 10 minutes, and then the next 10 minutes or

00:35:08: the next

00:35:09: thing that felt like 10 minutes was actually 25 minutes.

00:35:14: That's great.

00:35:15: I mean, I think we're both really deeply passionate about the AI and this topic and

00:35:21: really sharing our experience.

00:35:22: So yeah, definitely.

00:35:23: I agree.

00:35:24: Good to be back.

00:35:25: Time flies.

00:35:26: Yeah.

00:35:27: So thank you for joining us.

00:35:28: I'd love to invite you to subscribe to us if you enjoy these types of sessions that

00:35:35: we have.

00:35:36: I'll leave a comment.

00:35:37: let us know how we could improve and what topics you want us to cover next.

00:35:42: Thank you for joining.

00:35:43: And think about liking and subscribing the video if you liked it.

00:35:49: And subscribing to the channel, of course.

00:35:51: So yeah, really appreciate you.

00:35:54: And thanks for listening.

00:35:55: See you next time.

00:35:57: Bye.