WEBVTT
00:00:01.760 --> 00:00:10.960
OpenAI announced a new offering yesterday in OpenAI Presence, a service that I thought would make a pretty big splash.
00:00:11.119 --> 00:00:18.079
Yet in the first 24-ish hours since its announcement, it's been surprisingly silent.
00:00:18.320 --> 00:00:22.960
I mean, kind of below average chatter online, below average media coverage.
00:00:23.199 --> 00:00:27.679
Nothing really went viral or took off when maybe it should have.
00:00:28.079 --> 00:00:32.079
But this emerging category needs your attention.
00:00:32.320 --> 00:00:34.320
So what is OpenAI Presence?
00:00:34.479 --> 00:00:44.880
Well, it's an enterprise AI platform that lets companies build, deploy, and manage real-time voice in chat agents for customer support, sales, HR, and IT.
00:00:45.039 --> 00:00:49.439
OpenAI says each agent improves through real-world experience.
00:00:49.600 --> 00:00:57.200
And I've thought for years that eventually one of the first big disruptions in AI would be customer service.
00:00:57.439 --> 00:01:02.960
But the real-time models weren't narrow enough, or fast enough, or nearly smart enough.
00:01:03.200 --> 00:01:04.319
Now they are.
00:01:04.560 --> 00:01:08.640
And the path for disruption is almost impossible to ignore.
00:01:08.799 --> 00:01:19.120
I mean, customer service has been historically poor quality, slow, not helpful, and seemingly running on pre-internet technology, let alone pre-AI technology.
00:01:19.359 --> 00:01:30.000
The other main thing worth pointing out: while Anthropic has recently been grabbing headlines by pulling models from subscription, OpenAI is seemingly focusing beyond the model now.
00:01:30.239 --> 00:01:34.560
So will your company be employing customer service agents anytime soon?
00:01:34.799 --> 00:01:38.640
And might that be a good thing or a bad thing?
00:01:38.879 --> 00:01:40.400
Well, let's find out.
00:01:40.480 --> 00:01:42.640
And welcome to Everyday AI.
00:01:42.799 --> 00:01:44.879
Here's the big picture on what's new.
00:01:44.959 --> 00:01:50.719
So OpenAI has launched their new presence to automate customer service.
00:01:50.879 --> 00:01:56.640
So presence lets companies deploy real-time voice and chat agents for support, sales, and internal workflows.
00:01:56.719 --> 00:02:04.239
And OpenAI says that these things just get smarter the more that your company uses them in the real world interfacing with customers.
00:02:04.400 --> 00:02:09.120
And right now, customer service, does anyone actually like it?
00:02:09.520 --> 00:02:11.919
For me, it's almost the bane of my existence.
00:02:12.000 --> 00:02:18.000
Uh, I know I'm gonna have to, you know, get transferred six times if I have to actually call the company.
00:02:18.159 --> 00:02:24.560
Those companies that have tried to roll out voice agents were maybe ahead of the curve and the technology really wasn't what ready.
00:02:24.719 --> 00:02:32.319
So I always thought that AI would be uh, you know, kind of fitting this uh huge gap, but it was too early.
00:02:32.479 --> 00:02:42.319
But now we're actually there because if you haven't used OpenAI's new GPT live or GPT real-time 2 models, you would not know any better.
00:02:42.479 --> 00:02:48.479
You would say, Oh, any AI agent, yeah, I've I've heard these talking to customer service, they're slow and very robotic.
00:02:48.639 --> 00:02:50.000
Well, that's not the case anymore.
00:02:50.159 --> 00:02:58.960
So, on today's show, here's what you're gonna learn you're gonna know what open AI presence actually is and why it's not the new voice model everyone assumes.
00:02:59.199 --> 00:03:04.960
You're gonna know everything that goes into building one of these agents and how open AI keeps it in check and how your company can too.
00:03:05.199 --> 00:03:16.960
You're gonna know the one number OpenAI is using to prove presence works and everything that number leaves out, and what this could mean for your company, your team, and every customer who maybe calls in if you are using it.
00:03:17.120 --> 00:03:19.120
All right, welcome to everyday AI.
00:03:19.199 --> 00:03:22.400
If you're new here, my name is Jordan Wilson and we do this every day.
00:03:22.560 --> 00:03:31.520
This is your unedited, unscripted daily live stream, podcast, and free daily newsletter, helping business leaders like you and me keep up with everything that's happening in the world of AI.
00:03:31.680 --> 00:03:33.280
I tell you what matters, what doesn't.
00:03:33.439 --> 00:03:35.439
You use that information to grow your company and career.
00:03:35.520 --> 00:03:36.479
It's a cheat code.
00:03:36.639 --> 00:03:38.000
So it starts here.
00:03:38.240 --> 00:03:41.360
But make sure you go to our website at your everydayai.com.
00:03:41.439 --> 00:03:47.759
We're going to be recapping the main points of today's show as well as all of the other AI developments that you need to keep in mind.
00:03:48.000 --> 00:03:50.240
All right, so open AI presence.
00:03:50.319 --> 00:03:51.439
Let's get into it.
00:03:51.680 --> 00:03:57.599
Because I'm wondering if the customer service takeover is finally now arriving.
00:03:57.759 --> 00:04:03.520
Uh, because we saw some of these earlier products launch, you know, like two or so years ago.
00:04:03.680 --> 00:04:09.520
It was kind of like the 2023, maybe 2024 normal AI agents launch, right?
00:04:09.599 --> 00:04:19.759
There's all this hype behind it, but the technology, I don't think for actual on-demand agents that can help you with your work, I don't think that really arrived until like the past six months.
00:04:20.000 --> 00:04:22.720
So the same thing we saw with voice agents, right?
00:04:22.800 --> 00:04:34.879
As soon as voice models became available, uh, you know, commercially and from the consumer side in like late 2023, early 2024, you know, there was already this first warning sign, right?
00:04:34.959 --> 00:04:41.199
The the the boy who cried wolf came and said, Hey, customer service, you know, all these jobs are gonna go away.
00:04:41.279 --> 00:04:43.839
But it wasn't the case, the models weren't good.
00:04:43.920 --> 00:04:48.160
Uh, but now this may be different now.
00:04:48.800 --> 00:05:04.160
But first, I want to zoom out and not just talk about open AI presence, and I want to talk about the space in general because oftentimes it is the competition uh that ultimately drives the end product or the services that those product that those products get rolled into.
00:05:04.319 --> 00:05:13.199
And reportedly, right, and we'll see and we'll be covering this tomorrow uh on our uh you know Friday feature show where we go over features that are available for everyone to use.
00:05:13.360 --> 00:05:24.399
But uh reportedly we'll be getting a codex real-time voice mode soon, uh, right, that acts more like a personal assistant, like a smart AI-powered series that actually works.
00:05:24.639 --> 00:05:25.199
All right.
00:05:25.600 --> 00:05:36.639
So that would be amazing, but they're not the only company that's obviously investing heavily in voice as more and more people seemingly become more comfortable talking to or with AI, right?
00:05:36.720 --> 00:05:38.639
Maybe it's because of the dictation.
00:05:38.800 --> 00:05:46.480
Uh, maybe it's because, well, now the models are actually smart enough and responsive enough that we can actually use these voice models and people see how powerful that is, right?
00:05:46.639 --> 00:05:55.279
Everyone's always wanted to have their own, you know, Jarvis that you could just say something to and it uh you know understands and it thinks about it, and it doesn't just give you a robotic response back.
00:05:55.439 --> 00:06:08.399
So not only maybe we be getting uh a new real-time voice mode from OpenAI via codex, but also we've seen some recent leaks that show Anthropic is also uh gonna be rolling out some new voice modes as well.
00:06:08.720 --> 00:06:12.560
And then, I mean, you can't forget Google and their Gemini Live.
00:06:12.720 --> 00:06:17.199
So when it was announced, it was actually kind of the leader of the pack.
00:06:17.279 --> 00:06:26.800
Uh, but we know that Gemini is at least temporarily behind on getting new models uh to market, although we did just get the uh Gemini 3.6 Flash.
00:06:26.879 --> 00:06:31.759
But in terms of their pro series and Gemini Live updates, we haven't seen anything in a long time.
00:06:31.839 --> 00:06:41.839
But I would assume that whether it is part of the Gemini 4 that is now under pre-training or the Gemini 3.5 Pro that you know has been delayed now for a couple of months.
00:06:42.000 --> 00:06:45.600
My assumption is we will get a new Gemini Live mode.
00:06:45.680 --> 00:06:59.600
So the three big players here are seemingly starting to either invest or re-invest in the real-time uh voice space because, like I said, now the models are smart enough.
00:06:59.680 --> 00:07:08.399
You can talk to a model, and kind of how OpenAI runs this is you talk to a model and it responds, and then it sends another model in the background to call tools and to search the web.
00:07:08.480 --> 00:07:10.480
And that was one of the big things that was missing.
00:07:10.639 --> 00:07:14.399
Well, number one, initially, the models were not real time and neural enough.
00:07:14.480 --> 00:07:19.040
They sounded robotic and you know, it might have paused for a half second or a second.
00:07:19.199 --> 00:07:23.839
And when it comes to customer service and customer interactions, that's all you need to completely ruin it, right?
00:07:23.920 --> 00:07:31.279
And that's why you know, back in 2024, so many companies did not jump on this bandwagon, rightfully so, because the technology wasn't there.
00:07:31.439 --> 00:07:43.680
But now these models, right, they respond to you in real time, especially in the new uh GPT Live 1 and the uh GPT 2 real time, which we'll be talking about those models and the differences between them, right?
00:07:43.759 --> 00:07:53.680
So, not only can they respond to you right away and sound fairly human, uh, but it can send another model in the background to call tools because before they couldn't really do that.
00:07:53.839 --> 00:08:03.439
So sometimes these voice models hallucinated or they couldn't really work with your data or understand you know what it was that you were talking about because it couldn't really do that in the background.
00:08:03.519 --> 00:08:04.399
Now it can.
00:08:04.720 --> 00:08:17.040
All right, so with that in mind, let's talk a little bit more kind of about this uh real-time voice race, because I always think it's important to understand that before we jump into any you know product or service.
00:08:17.279 --> 00:08:22.720
So open AI, though, has been quietly dominating the real-time AI voice space, right?
00:08:22.800 --> 00:08:26.480
And and again, I think this is one of the future modalities.
00:08:26.639 --> 00:08:40.559
I think multimodal AI uh is so AI is interacting with your desktop, at least right now, and probably after that, um, an AI that you can just talk to and it can complete work on your desktop or your phone, right?
00:08:40.639 --> 00:08:45.919
I think that is kind of the next iteration or the next wave that's coming with helpful AI.
00:08:46.480 --> 00:08:52.399
And it starts with well, a model that you can talk to is smart and can kind of do things in the background for you.
00:08:52.559 --> 00:08:56.399
So if we look at the artificial analysis speech-to-speech index, right?
00:08:56.480 --> 00:09:03.759
So that's probably the best uh single index, right, that takes into account multiple um, you know, different benchmarks.
00:09:03.840 --> 00:09:13.200
And open AI has been winning this space, like I said, kind of quietly, because everyone, for the most part, is talking about just the basic, you know, text or agent benchmarks.
00:09:13.360 --> 00:09:23.519
Very few people are focusing on speech to speech, which is interesting to me because if more and more people, at least that are playing on the edge of AI, are dictating everything, like me, right?
00:09:23.679 --> 00:09:28.320
I think that's an early indicator of where the space as a whole is headed, right?
00:09:28.480 --> 00:09:39.840
I think we'll and maybe we're all just waiting on Apple uh to finally get the smart Siri right, but that I think will start to shift how people work and interface with technology.
00:09:40.000 --> 00:09:45.679
So uh Open AI, when it comes to this benchmark, they've been winning and it hasn't really been close.
00:09:45.840 --> 00:09:56.000
They actually have four of the top six models, including the best model in the artificial analysis speech to speech index in their new GPT, uh GPT Real Time 2.
00:09:56.320 --> 00:10:09.360
But what's interesting to me is even their GPT Real Time 1.5, the old version, uh, is is better than what Google has in their uh Gemini 3.1 live.
00:10:09.840 --> 00:10:23.519
All right, so let's take a quick look and I'm gonna show our uh live stream audience here a little bit on the kind of marketing side for what open AI is putting out there for open AI presence.
00:10:23.600 --> 00:10:30.159
Uh so what they said is put trusted AI agents to work across customer channels and internal workflows.
00:10:30.320 --> 00:10:33.840
So right now you do have to reach out to their sales team if you want to access.
00:10:33.919 --> 00:10:42.480
So it's not like you're just gonna log into ChatGPT or log into your uh, you know, your company's uh, you know, open AI playground and start deploying these things.
00:10:42.639 --> 00:10:46.480
So it is uh a very limited uh rollout right now.
00:10:46.639 --> 00:10:56.799
But what they say, this is trusted AI agents, and they say each agent improves through real-world experience with evaluations, guardrails, and human approval governing every change.
00:10:57.519 --> 00:11:05.440
So, kind of the uh the three ways that they break this out to explain it is they say that this gives you one presence across every channel.
00:11:05.600 --> 00:11:10.639
They say you can show up in consistent in a consistent way across voice and chat.
00:11:10.720 --> 00:11:11.840
So ultimately, right?
00:11:11.919 --> 00:11:15.440
So if your company is looking to deploy voice agents, right?
00:11:15.519 --> 00:11:23.200
So a phone number, maybe you don't have all the humans you need to handle the volumes of calls, that's one thing, but then also on your website, uh right.
00:11:23.279 --> 00:11:29.679
So it's it's kind of a uh one system that can work and stay in sync with both.
00:11:29.840 --> 00:11:33.200
So they say show up in a consistent way across voice and chat.
00:11:33.279 --> 00:11:41.519
You decide what remains consistent, such as policies, evaluations, and escalation rules and what should change for each workflow or channel.
00:11:41.679 --> 00:11:55.919
Then they say trust built into every uh deployment, connect agents to company systems, define policies and permissions, and validate performance through simulations, evaluations, guardrails, and escalation paths.
00:11:56.240 --> 00:12:00.480
And then last but not least, they say improve quality continuously.
00:12:00.559 --> 00:12:12.799
Uh, they say presence continuously gets better with use, it uses production conversations and quality signals to recommend improvements that make agents more capable.
00:12:13.120 --> 00:12:26.320
And then OpenAI has shared some use cases with some big companies that they've been testing this out with, such as BBVA, Softbank, um, IHE, and then obviously their own internal use cases as well.
00:12:26.399 --> 00:12:34.559
So things that they're saying this can be used for is customer support, demand generation, claims, procurement, it helped desk, uh, human resource, etc.
00:12:34.799 --> 00:12:53.440
But essentially, anywhere right now where maybe your company is experiencing human bottleneck and you can't really scale um at the point that you need to, whether it's for customers or maybe if you're a large organization, sometimes it's an internal, uh, you know, almost ITHR uh type shortcoming that you're running into.
00:12:53.519 --> 00:13:02.960
Um, and and you know, I kind of want to go back to this concept of what AI unlocks, and I think that initial voice agents maybe got this all wrong.
00:13:03.120 --> 00:13:21.440
Because what I think initial voice agents that were too early, it was more of just uh scaling this technology out to as many people as possible, but in a similar way, which I think is not the point of generative AI and artificial intelligence, right?
00:13:21.600 --> 00:13:25.279
I think a lot of people I talk about this a lot, but they think, oh, this is an easy button.
00:13:25.360 --> 00:13:29.759
Let's get a blanket approach that works for everyone, which I think is the absolute wrong approach, right?
00:13:29.919 --> 00:13:39.440
The way that you should be doing this, and I think kind of tying in uh the you know website chat side with the voice side is very helpful, right?
00:13:39.519 --> 00:13:46.559
I think this is you know, we'll see what kind of B2B applications work well for this, but I think you should think of B2C, right?
00:13:46.639 --> 00:13:51.360
So if you're uh you know working at a company or maybe a consumer, right?
00:13:51.440 --> 00:13:54.559
Maybe you are constantly like Amazon, right?
00:13:54.639 --> 00:13:59.120
That's not to pick on one certain company, but you know, bless up my wife is always the one.
00:13:59.200 --> 00:14:07.679
If something happens with Amazon, she's the one that braves, you know, getting on the website and you know, talking on the phone to three or four or five or six different voice agents.
00:14:07.759 --> 00:14:09.919
I think Amazon is actually okay at this, right?
00:14:10.080 --> 00:14:16.240
But think of those companies, if it's not your company that could use something like this, think of those companies that you interface with.
00:14:16.320 --> 00:14:33.279
So I do think this is on the B uh B2C side, but having that data follow you around from whether it is you are online chatting or on the phone and not having to re-explain those things or you know what went wrong with an order, because presumably it's all going to be tracked, right?
00:14:33.360 --> 00:14:39.120
That's the thing that open AI says here, you know, talking about how it connects to your own company systems.
00:14:39.279 --> 00:14:58.159
Uh, so in the same way that you know, most companies have been rushing to get their data in order, you know, over the past, you know, post-Chat GPT phase when they're like, oh my gosh, all of a sudden, we really need to have our data in order uh for AI systems to be able to read and make use of this, right?
00:14:58.320 --> 00:15:05.679
This is this holds true uh for voice and customer success or customer experience uh agents as well.
00:15:06.159 --> 00:15:12.559
So that's a kind of brief overview of what OpenAI is uh saying that this is.
00:15:13.039 --> 00:15:15.039
So it's not a new model, all right?
00:15:15.200 --> 00:15:19.039
It's just kind of the infrastructure that surrounds their voice models.
00:15:19.279 --> 00:15:26.799
So every presence agent, according to OpenAI, starts with one narrow job and only the system access that the job needs.
00:15:26.879 --> 00:15:34.080
So companies set the allowed actions, required approvals, and exact moments that presence should escalate to a human, right?
00:15:34.159 --> 00:15:35.759
So you can set that level.
00:15:35.840 --> 00:15:48.720
Maybe you set it very low, and I would probably recommend that when this does roll out to the masses, is you know, don't set that escalation bar high, especially when you're testing out a new technology in production, right?
00:15:48.799 --> 00:15:49.759
You have to get it right.
00:15:49.919 --> 00:15:57.200
And again, I think people always look at maximum autonomy when it comes to trying out new AI, which is the exact wrong way to look at it.
00:15:57.279 --> 00:16:11.679
You look at your uh low stakes, high quantity uh uh instances, and then you have to be able to develop uh a system internally in the same way that I've always encouraged companies to create your own benchmarks, right?
00:16:11.759 --> 00:16:13.519
But those are text-based benchmarks.
00:16:13.679 --> 00:16:21.759
Don't just look at artificial analysis, don't just look at, I don't know, whatever you're you're looking at, uh, arena or you know, humanities last exam, right?
00:16:21.840 --> 00:16:25.679
You can't look at a single benchmark uh for enterprise AI adoption.
00:16:25.840 --> 00:16:27.759
Those are great starting points, right?
00:16:27.840 --> 00:16:46.159
But you need to be uh for text-based or agent uh kind of agent-based um systems, you have to have your own internal text-based benchmarks that say this is what uh qualifies as a win, this is what qualifies as a loss, these are our guardrails, this is how we deploy this throughout our organization.
00:16:46.320 --> 00:16:54.320
You have to do the exact same thing for voice agents, which is a little more difficult because there's things like nuance and sarcasm in your voice.
00:16:54.480 --> 00:17:15.839
So that's one thing, and maybe one reason um why these voice agents, I think initially when we saw this first puff of smoke in 2024 with oh my gosh, voice AI agents are gonna take over customer service, and they well, they didn't, is because the models weren't even smart enough to get over that big uh gray area of hey, people are ambiguous, right?
00:17:16.000 --> 00:17:19.119
If someone responds yes, you can say yes, right?
00:17:19.200 --> 00:17:20.559
That someone's excited about it.
00:17:20.720 --> 00:17:26.240
If they say yeah, right, that means they're maybe being sarcastic and not actually excited about something.
00:17:26.400 --> 00:17:39.920
So, models for the most part, I think couldn't even get over one of those first big initial humps because the reason why most companies have right customer support with real humans is to understand that human nuance.
00:17:40.000 --> 00:17:54.240
Uh, so if if if all the model was doing in 2023, 2024 was essentially translating uh your speech to text and then having a model in the background read that text, that's the reason why it just didn't work initially.
00:17:55.200 --> 00:18:01.759
So, uh what open AI actually bundles together is what presence actually is.
00:18:01.920 --> 00:18:10.160
So they're giving you the policies, uh the guardrails, approved actions, simulations, and evaluation tools inside of one product.
00:18:10.319 --> 00:18:15.519
So this is a codex-powered improvement process that's built directly into the platform, it's not sold separately.
00:18:15.680 --> 00:18:24.480
And open AI's own uh FDEs or uh forward deployed engineers, uh connect each company's systems and bring the agent live.
00:18:24.720 --> 00:18:28.160
So, yeah, we've been talking about this this FDE, right?
00:18:28.319 --> 00:18:30.880
Uh, you'll see it in our newsletter today, right?
00:18:30.960 --> 00:18:46.079
But Amazon just, you know, shut down kind of their AGI department or you know, laid off a bunch of people in their AGI department, and they're focusing more and more resources on FDEs, or you know, actually deploying putting their humans inside of these other companies.
00:18:46.240 --> 00:18:51.359
So it looks like that's what OpenAI's um kind of angle is here, right?
00:18:51.440 --> 00:18:58.160
They already have their FDEs, you know, big companies already have their dedicated open AI uh people working inside there.
00:18:58.319 --> 00:19:06.720
So this is just one other system that those FDEs or for deployed engineers will be able to implement uh for open AI customers.
00:19:07.279 --> 00:19:10.799
So here's the thing you have to talk about.
00:19:10.960 --> 00:19:34.400
Um is getting from beta, getting from demo to deployment, because that is what is ultimately going to decide whether OpenAI presence is kind of the next big wave of not just customer service, but maybe where all the other AI players will play, um, versus just, oh, it's a new feature that may or may not pick up steam.
00:19:34.559 --> 00:19:43.519
So, right now, the way that open AI frames it is that teams are able to simulate those common requests, edge cases, and higher risk scenarios before any customer actually sees the agents.
00:19:43.680 --> 00:19:48.559
So you can set those guardrails, humans can monitor and step in mid-conversation, uh, etc.
00:19:48.799 --> 00:19:56.480
And then after launch, codecs can investigate real production signals and propose updates that teams can actually test before rollout.
00:19:56.880 --> 00:20:00.880
So now let's try to understand a little bit of what's happening under the hood.
00:20:01.039 --> 00:20:10.079
And I think this part, especially if you use OpenAI's new GPT live, and it is bonkers, y'all.
00:20:10.160 --> 00:20:10.960
It is good.
00:20:11.119 --> 00:20:12.720
I wanted to do a show on this, right?
00:20:12.799 --> 00:20:17.519
But in our newsletter for our Wednesday demos, I usually have you all vote.
00:20:17.839 --> 00:20:27.920
Um, I really want to do a show on this because I think GPT Live, that's their new um essentially the replacement for advanced voice mode inside of Chat GPT.
00:20:28.079 --> 00:20:31.039
If you haven't used it, my gosh, it is good.
00:20:31.200 --> 00:20:42.960
Um, not just being able to understand your data, but you you know, I think you can finally experience talking to an agent and then knowing that it's actually calling tools and looking things up in the background.
00:20:43.119 --> 00:20:54.319
Whereas before, I think live voice agents, whether it was Chat GPT's advocate uh advanced voice mode, uh, you know, Gemini Live, those are probably the two uh, you know, big consumer versions.
00:20:54.480 --> 00:21:01.200
It almost seemed like all these voice agents would give you an intentionally uh ambiguous and vague answer.
00:21:01.279 --> 00:21:06.240
And you're like, okay, it seems to maybe be looking up something in the background, but maybe not.
00:21:06.319 --> 00:21:08.079
And it's almost like you couldn't tell.
00:21:08.240 --> 00:21:23.039
I remember one time I was literally just trying to test like you know, like the hallucination or ambiguous rate uh with uh you know, head to head between Gemini Live and uh the advanced voice mode with ChatGPT, but regardless, the technology just wasn't good.
00:21:23.119 --> 00:21:27.920
But I think this last iteration from OpenAI makes it good and it's multiple models.
00:21:28.079 --> 00:21:31.920
So, first, let's talk about GPT Real Time 2.
00:21:32.079 --> 00:21:35.839
Well, technically it's now GPT Real Time 2.1 because they just came out with an update.
00:21:35.920 --> 00:21:41.839
But that's essentially the developer tool for building custom voice agents through OpenAI's API.
00:21:42.000 --> 00:21:44.640
So they essentially have this technology, right?
00:21:44.720 --> 00:21:47.519
Let's just call it the super smart AI voice, right?
00:21:47.599 --> 00:21:54.079
So if you're a developer and if you want to build with this new super smart AI voice, you would use GBT Real Time 2.1.
00:21:54.240 --> 00:21:59.599
If you're a consumer and you want to experience it just using your Chat GPT account, that's called GPT.
00:22:00.240 --> 00:22:01.039
Live one.
00:22:01.119 --> 00:22:04.000
So both of these are new in the last two weeks.
00:22:04.079 --> 00:22:21.039
So that powers the new chat GPT voice where you can hold a live conversation while the AI quietly handles harder work off to another model, whether it's calling tools, looking at your data, uh searching something on the internet, uh, thinking, reasoning in the background about something that would normally take an AI model a little bit longer.
00:22:21.119 --> 00:22:21.680
Uh, right.
00:22:21.759 --> 00:22:31.039
The last technical details we we saw was that they kicked this over to GPT 5.5 and it would think and reason about things while the real-time model kept the conversation going.
00:22:31.200 --> 00:22:34.160
Uh, so presence though sits above both.
00:22:34.400 --> 00:22:36.720
So it's not a new model, right?
00:22:36.799 --> 00:22:43.039
It sits above both, and open AI has not yet revealed uh which model actually powers uh presence.
00:22:43.119 --> 00:22:47.839
Although I'm guessing it's probably something on the GPT real-time to side.
00:22:47.920 --> 00:22:48.960
That would be my guess.
00:22:49.599 --> 00:22:50.160
All right.
00:22:51.200 --> 00:22:54.319
So why do you have to pay attention to this?
00:22:55.200 --> 00:23:04.319
And I think this is something where I think trust either increases or erodes depending on your experience, right?
00:23:04.480 --> 00:23:06.400
A very simple example.
00:23:07.680 --> 00:23:10.160
I don't I forgot what the company is, right?
00:23:10.240 --> 00:23:14.079
But it's it's either a heating or air conditioning company that I use here in Chicago.
00:23:14.240 --> 00:23:18.079
They had their voice agents and they're actually tricky at first.
00:23:18.240 --> 00:23:21.599
The first time I called, I don't know, heater, air conditioner wasn't working very well.
00:23:21.759 --> 00:23:24.960
And I'm like, wait, this kind of sounds real.
00:23:25.359 --> 00:23:32.319
And then instantly, obviously, being around AI, I knew it was AI and I was having fun trying to you know jailbreak it on the phone, that kind of stuff.
00:23:32.400 --> 00:23:38.480
Um, but it couldn't get human handoff, and that was a bad thing.
00:23:38.559 --> 00:23:50.799
Uh, because ultimately what matters here, it's not about the technology, it's about are consumers' uh queries being resolved or not, or are people just more frustrated?
00:23:51.039 --> 00:23:54.000
But the the reality is, I think consumers want this.
00:23:54.079 --> 00:24:00.960
There was actually a study from 5.9 that said that 80% of consumers are willing to use AI customer service, right?
00:24:01.200 --> 00:24:16.000
Yes, I know there's some companies out there that do customer service well, but I would say this is an argue that is literally just uh sorry, this is an area that is literally begging to be disrupted because it is so bad, the quality is so slow, and ultimately, I don't know.
00:24:16.240 --> 00:24:21.119
Anytime I have to make a call to customer service, I feel there is no resolution.
00:24:21.279 --> 00:24:26.079
So maybe this is something I'm personally rooting for for the industry, right?
00:24:26.160 --> 00:24:37.039
Um I don't let that impact, you know, if if you're a decision maker at your company, don't let that impact, you know, your decision on whether you should be using these types of voice in real-time agents or not.
00:24:37.200 --> 00:24:41.759
But regardless, the consumer demand is almost overwhelming, right?
00:24:41.839 --> 00:24:50.400
You've you've seen this, whether it's uh it has any merit or not, you've seen this recent kind of uprising against AI.
00:24:50.480 --> 00:24:53.119
And I'm not just talking about AI slop, right?
00:24:53.359 --> 00:24:57.680
Um, companies I think are deploying sometimes AI in an irresponsible way.
00:24:57.759 --> 00:25:01.680
Uh, when it comes to job creation, job growth, uh job growth, things like that.
00:25:01.759 --> 00:25:12.799
So sometimes there's this kickback against AI, uh, but at least when it comes to the appetite for uh deploying real good AI agents, it seems like customer service, right?
00:25:12.960 --> 00:25:14.480
Consumers want this.
00:25:14.640 --> 00:25:23.279
So if you are a business leader, maybe you haven't even uh, you know, maybe you're in the in the space where it's like, yeah, we don't even really offer customer service.
00:25:23.440 --> 00:25:25.039
Well, maybe you probably can.
00:25:25.440 --> 00:25:28.000
Uh so that's another thing to uh think about.
00:25:29.200 --> 00:25:35.279
So, like I started the show off with as we wrap here, this was a quiet launch, which was interesting to me.
00:25:35.359 --> 00:25:40.640
And presence, I think, drew so little attention because essentially it's not sexy, right?
00:25:40.799 --> 00:25:45.359
Enterprise infrastructure doesn't look exciting when it's first launched.
00:25:45.440 --> 00:25:46.960
But that's what powers this thing.
00:25:47.039 --> 00:25:52.480
If you don't have the infrastructure set up, this isn't going to be a reality for you.
00:25:52.720 --> 00:25:58.079
So I think that obviously customer service organizations they do have the infrastructure, but what about everyone else?
00:25:58.240 --> 00:26:05.680
What if customer you know, service or phone service is, you know, not even a top 10 priority for you, but you've always been uh offering it.
00:26:05.839 --> 00:26:12.079
Uh, I I think customer service though may become AI's kind of first mass consumer disruption, right?
00:26:12.240 --> 00:26:14.240
Aside from text, that's that's come and gone.
00:26:14.319 --> 00:26:16.559
So maybe it's the next big iteration.
00:26:16.720 --> 00:26:25.200
Um, yes, we've all seen AI photos and AI videos um in advertisements, but I think that's going to continue to blur the line between what's real and what's not.
00:26:25.279 --> 00:26:51.519
But I think what is real is that this is a space that is probably likely going to explode because, like I started with uh both OpenAI, uh Google, um, Anthropic, and even technically Grok, all the players are starting to in are starting to invest heavily in this space because the technology is finally good enough where this can actually be a customer service tool that is worthwhile and that uh the users want.
00:26:52.079 --> 00:26:57.920
But whether that helps customers depends entirely on what companies optimize this first.
00:26:58.079 --> 00:27:08.240
And if you are just looking for that big red easy button, or if you have the infrastructure in human power to actually deploy this in a responsible way that helps meet that demand.
00:27:08.400 --> 00:27:12.880
All right, so that's a wrap on what's new with open AI presence.
00:27:13.039 --> 00:27:24.559
And hey, my kind of hot take uh takeaway on this is like I said, I'm rooting for this, not just for open AI presence, um, I'm rooting for this segment.
00:27:25.200 --> 00:27:31.680
Because personally, as a consumer, I have hated customer service for the longest time.
00:27:31.839 --> 00:27:37.759
I've very rarely had a good experience, and at least according to studies, it seems like that is normal across the board.
00:27:37.920 --> 00:27:43.279
So, in my uh in my experience, this is something that I think enterprise companies should be looking at.
00:27:43.440 --> 00:27:51.200
Obviously, if you're already in the B2C space, if you already have dedicated customer support, there's probably a good chance that you have the infrastructure to take advantage of this.
00:27:51.279 --> 00:27:52.480
So you should be looking at it.
00:27:52.640 --> 00:27:56.559
But for everyone else, I think this also begs the question: should we be doing this?
00:27:56.720 --> 00:28:10.799
Whether it's for or maybe you're in the B2B space, maybe you are in the B2C uh space, but you just haven't had kind of a dedicated customer phone or customer chat service, and maybe you should, or even for larger organizations looking to deploy something like this internally, right?
00:28:10.960 --> 00:28:25.200
If you have tens of thousands of employees, being able to give that level of personalization, like I was talking about easier, uh or like I was talking about earlier, where I think companies, when they find generative AI, they're like, okay, we could just put out more content that appeals to everyone.
00:28:25.279 --> 00:28:36.400
When I think it's the reverse, you have to uh ingest those signals, ingest those data points, and you should just be putting out more personalized uh messaging and more helpful messaging to more people.
00:28:36.480 --> 00:28:38.799
So maybe that's internally, maybe that's externally.
00:28:38.880 --> 00:28:42.319
But regardless, I think the technology is finally there and that's exciting.
00:28:42.480 --> 00:28:46.720
And well, we'll see if open AI presence gets more eyeballs on it.
00:28:46.960 --> 00:28:49.119
So uh that's a wrap for today's show.
00:28:49.200 --> 00:28:49.920
Thank you for tuning in.
00:28:50.000 --> 00:28:55.759
If you haven't already, please go to your everydayai.com, sign up for the free daily newsletter, and make sure to tune in tomorrow.
00:28:55.839 --> 00:28:57.519
We're gonna be going over our Friday features.
00:28:57.680 --> 00:29:02.319
There's a lot of new stuff that dropped both uh yesterday and that I know is coming today.
00:29:02.400 --> 00:29:03.920
So you're not gonna want to miss that one.
00:29:04.079 --> 00:29:04.880
Thanks for tuning in.
00:29:04.960 --> 00:29:07.440
We'll see you back tomorrow and every day for more everyday AI.
00:29:07.519 --> 00:29:08.480
Thanks, y'all.
00:29:09.920 --> 00:29:13.200
And that's a wrap for today's edition of Everyday AI.
00:29:13.359 --> 00:29:14.480
Thanks for joining us.
00:29:14.640 --> 00:29:18.400
If you enjoyed this episode, please subscribe and leave us a rating.
00:29:18.559 --> 00:29:20.160
It helps keep us going.
00:29:20.400 --> 00:29:27.440
For a little more AI magic, visit your everydayai.com and sign up to our daily newsletter so you don't get left behind.
00:29:27.680 --> 00:29:30.319
Go break some barriers, and we'll see you next time.
00:00:01.760 --> 00:00:10.960
OpenAI announced a new offering yesterday in OpenAI Presence, a service that I thought would make a pretty big splash.
00:00:11.119 --> 00:00:18.079
Yet in the first 24-ish hours since its announcement, it's been surprisingly silent.
00:00:18.320 --> 00:00:22.960
I mean, kind of below average chatter online, below average media coverage.
00:00:23.199 --> 00:00:27.679
Nothing really went viral or took off when maybe it should have.
00:00:28.079 --> 00:00:32.079
But this emerging category needs your attention.
00:00:32.320 --> 00:00:34.320
So what is OpenAI Presence?
00:00:34.479 --> 00:00:44.880
Well, it's an enterprise AI platform that lets companies build, deploy, and manage real-time voice in chat agents for customer support, sales, HR, and IT.
00:00:45.039 --> 00:00:49.439
OpenAI says each agent improves through real-world experience.
00:00:49.600 --> 00:00:57.200
And I've thought for years that eventually one of the first big disruptions in AI would be customer service.
00:00:57.439 --> 00:01:02.960
But the real-time models weren't narrow enough, or fast enough, or nearly smart enough.
00:01:03.200 --> 00:01:04.319
Now they are.
00:01:04.560 --> 00:01:08.640
And the path for disruption is almost impossible to ignore.
00:01:08.799 --> 00:01:19.120
I mean, customer service has been historically poor quality, slow, not helpful, and seemingly running on pre-internet technology, let alone pre-AI technology.
00:01:19.359 --> 00:01:30.000
The other main thing worth pointing out: while Anthropic has recently been grabbing headlines by pulling models from subscription, OpenAI is seemingly focusing beyond the model now.
00:01:30.239 --> 00:01:34.560
So will your company be employing customer service agents anytime soon?
00:01:34.799 --> 00:01:38.640
And might that be a good thing or a bad thing?
00:01:38.879 --> 00:01:40.400
Well, let's find out.
00:01:40.480 --> 00:01:42.640
And welcome to Everyday AI.
00:01:42.799 --> 00:01:44.879
Here's the big picture on what's new.
00:01:44.959 --> 00:01:50.719
So OpenAI has launched their new presence to automate customer service.
00:01:50.879 --> 00:01:56.640
So presence lets companies deploy real-time voice and chat agents for support, sales, and internal workflows.
00:01:56.719 --> 00:02:04.239
And OpenAI says that these things just get smarter the more that your company uses them in the real world interfacing with customers.
00:02:04.400 --> 00:02:09.120
And right now, customer service, does anyone actually like it?
00:02:09.520 --> 00:02:11.919
For me, it's almost the bane of my existence.
00:02:12.000 --> 00:02:18.000
Uh, I know I'm gonna have to, you know, get transferred six times if I have to actually call the company.
00:02:18.159 --> 00:02:24.560
Those companies that have tried to roll out voice agents were maybe ahead of the curve and the technology really wasn't what ready.
00:02:24.719 --> 00:02:32.319
So I always thought that AI would be uh, you know, kind of fitting this uh huge gap, but it was too early.
00:02:32.479 --> 00:02:42.319
But now we're actually there because if you haven't used OpenAI's new GPT live or GPT real-time 2 models, you would not know any better.
00:02:42.479 --> 00:02:48.479
You would say, Oh, any AI agent, yeah, I've I've heard these talking to customer service, they're slow and very robotic.
00:02:48.639 --> 00:02:50.000
Well, that's not the case anymore.
00:02:50.159 --> 00:02:58.960
So, on today's show, here's what you're gonna learn you're gonna know what open AI presence actually is and why it's not the new voice model everyone assumes.
00:02:59.199 --> 00:03:04.960
You're gonna know everything that goes into building one of these agents and how open AI keeps it in check and how your company can too.
00:03:05.199 --> 00:03:16.960
You're gonna know the one number OpenAI is using to prove presence works and everything that number leaves out, and what this could mean for your company, your team, and every customer who maybe calls in if you are using it.
00:03:17.120 --> 00:03:19.120
All right, welcome to everyday AI.
00:03:19.199 --> 00:03:22.400
If you're new here, my name is Jordan Wilson and we do this every day.
00:03:22.560 --> 00:03:31.520
This is your unedited, unscripted daily live stream, podcast, and free daily newsletter, helping business leaders like you and me keep up with everything that's happening in the world of AI.
00:03:31.680 --> 00:03:33.280
I tell you what matters, what doesn't.
00:03:33.439 --> 00:03:35.439
You use that information to grow your company and career.
00:03:35.520 --> 00:03:36.479
It's a cheat code.
00:03:36.639 --> 00:03:38.000
So it starts here.
00:03:38.240 --> 00:03:41.360
But make sure you go to our website at your everydayai.com.
00:03:41.439 --> 00:03:47.759
We're going to be recapping the main points of today's show as well as all of the other AI developments that you need to keep in mind.
00:03:48.000 --> 00:03:50.240
All right, so open AI presence.
00:03:50.319 --> 00:03:51.439
Let's get into it.
00:03:51.680 --> 00:03:57.599
Because I'm wondering if the customer service takeover is finally now arriving.
00:03:57.759 --> 00:04:03.520
Uh, because we saw some of these earlier products launch, you know, like two or so years ago.
00:04:03.680 --> 00:04:09.520
It was kind of like the 2023, maybe 2024 normal AI agents launch, right?
00:04:09.599 --> 00:04:19.759
There's all this hype behind it, but the technology, I don't think for actual on-demand agents that can help you with your work, I don't think that really arrived until like the past six months.
00:04:20.000 --> 00:04:22.720
So the same thing we saw with voice agents, right?
00:04:22.800 --> 00:04:34.879
As soon as voice models became available, uh, you know, commercially and from the consumer side in like late 2023, early 2024, you know, there was already this first warning sign, right?
00:04:34.959 --> 00:04:41.199
The the the boy who cried wolf came and said, Hey, customer service, you know, all these jobs are gonna go away.
00:04:41.279 --> 00:04:43.839
But it wasn't the case, the models weren't good.
00:04:43.920 --> 00:04:48.160
Uh, but now this may be different now.
00:04:48.800 --> 00:05:04.160
But first, I want to zoom out and not just talk about open AI presence, and I want to talk about the space in general because oftentimes it is the competition uh that ultimately drives the end product or the services that those product that those products get rolled into.
00:05:04.319 --> 00:05:13.199
And reportedly, right, and we'll see and we'll be covering this tomorrow uh on our uh you know Friday feature show where we go over features that are available for everyone to use.
00:05:13.360 --> 00:05:24.399
But uh reportedly we'll be getting a codex real-time voice mode soon, uh, right, that acts more like a personal assistant, like a smart AI-powered series that actually works.
00:05:24.639 --> 00:05:25.199
All right.
00:05:25.600 --> 00:05:36.639
So that would be amazing, but they're not the only company that's obviously investing heavily in voice as more and more people seemingly become more comfortable talking to or with AI, right?
00:05:36.720 --> 00:05:38.639
Maybe it's because of the dictation.
00:05:38.800 --> 00:05:46.480
Uh, maybe it's because, well, now the models are actually smart enough and responsive enough that we can actually use these voice models and people see how powerful that is, right?
00:05:46.639 --> 00:05:55.279
Everyone's always wanted to have their own, you know, Jarvis that you could just say something to and it uh you know understands and it thinks about it, and it doesn't just give you a robotic response back.
00:05:55.439 --> 00:06:08.399
So not only maybe we be getting uh a new real-time voice mode from OpenAI via codex, but also we've seen some recent leaks that show Anthropic is also uh gonna be rolling out some new voice modes as well.
00:06:08.720 --> 00:06:12.560
And then, I mean, you can't forget Google and their Gemini Live.
00:06:12.720 --> 00:06:17.199
So when it was announced, it was actually kind of the leader of the pack.
00:06:17.279 --> 00:06:26.800
Uh, but we know that Gemini is at least temporarily behind on getting new models uh to market, although we did just get the uh Gemini 3.6 Flash.
00:06:26.879 --> 00:06:31.759
But in terms of their pro series and Gemini Live updates, we haven't seen anything in a long time.
00:06:31.839 --> 00:06:41.839
But I would assume that whether it is part of the Gemini 4 that is now under pre-training or the Gemini 3.5 Pro that you know has been delayed now for a couple of months.
00:06:42.000 --> 00:06:45.600
My assumption is we will get a new Gemini Live mode.
00:06:45.680 --> 00:06:59.600
So the three big players here are seemingly starting to either invest or re-invest in the real-time uh voice space because, like I said, now the models are smart enough.
00:06:59.680 --> 00:07:08.399
You can talk to a model, and kind of how OpenAI runs this is you talk to a model and it responds, and then it sends another model in the background to call tools and to search the web.
00:07:08.480 --> 00:07:10.480
And that was one of the big things that was missing.
00:07:10.639 --> 00:07:14.399
Well, number one, initially, the models were not real time and neural enough.
00:07:14.480 --> 00:07:19.040
They sounded robotic and you know, it might have paused for a half second or a second.
00:07:19.199 --> 00:07:23.839
And when it comes to customer service and customer interactions, that's all you need to completely ruin it, right?
00:07:23.920 --> 00:07:31.279
And that's why you know, back in 2024, so many companies did not jump on this bandwagon, rightfully so, because the technology wasn't there.
00:07:31.439 --> 00:07:43.680
But now these models, right, they respond to you in real time, especially in the new uh GPT Live 1 and the uh GPT 2 real time, which we'll be talking about those models and the differences between them, right?
00:07:43.759 --> 00:07:53.680
So, not only can they respond to you right away and sound fairly human, uh, but it can send another model in the background to call tools because before they couldn't really do that.
00:07:53.839 --> 00:08:03.439
So sometimes these voice models hallucinated or they couldn't really work with your data or understand you know what it was that you were talking about because it couldn't really do that in the background.
00:08:03.519 --> 00:08:04.399
Now it can.
00:08:04.720 --> 00:08:17.040
All right, so with that in mind, let's talk a little bit more kind of about this uh real-time voice race, because I always think it's important to understand that before we jump into any you know product or service.
00:08:17.279 --> 00:08:22.720
So open AI, though, has been quietly dominating the real-time AI voice space, right?
00:08:22.800 --> 00:08:26.480
And and again, I think this is one of the future modalities.
00:08:26.639 --> 00:08:40.559
I think multimodal AI uh is so AI is interacting with your desktop, at least right now, and probably after that, um, an AI that you can just talk to and it can complete work on your desktop or your phone, right?
00:08:40.639 --> 00:08:45.919
I think that is kind of the next iteration or the next wave that's coming with helpful AI.
00:08:46.480 --> 00:08:52.399
And it starts with well, a model that you can talk to is smart and can kind of do things in the background for you.
00:08:52.559 --> 00:08:56.399
So if we look at the artificial analysis speech-to-speech index, right?
00:08:56.480 --> 00:09:03.759
So that's probably the best uh single index, right, that takes into account multiple um, you know, different benchmarks.
00:09:03.840 --> 00:09:13.200
And open AI has been winning this space, like I said, kind of quietly, because everyone, for the most part, is talking about just the basic, you know, text or agent benchmarks.
00:09:13.360 --> 00:09:23.519
Very few people are focusing on speech to speech, which is interesting to me because if more and more people, at least that are playing on the edge of AI, are dictating everything, like me, right?
00:09:23.679 --> 00:09:28.320
I think that's an early indicator of where the space as a whole is headed, right?
00:09:28.480 --> 00:09:39.840
I think we'll and maybe we're all just waiting on Apple uh to finally get the smart Siri right, but that I think will start to shift how people work and interface with technology.
00:09:40.000 --> 00:09:45.679
So uh Open AI, when it comes to this benchmark, they've been winning and it hasn't really been close.
00:09:45.840 --> 00:09:56.000
They actually have four of the top six models, including the best model in the artificial analysis speech to speech index in their new GPT, uh GPT Real Time 2.
00:09:56.320 --> 00:10:09.360
But what's interesting to me is even their GPT Real Time 1.5, the old version, uh, is is better than what Google has in their uh Gemini 3.1 live.
00:10:09.840 --> 00:10:23.519
All right, so let's take a quick look and I'm gonna show our uh live stream audience here a little bit on the kind of marketing side for what open AI is putting out there for open AI presence.
00:10:23.600 --> 00:10:30.159
Uh so what they said is put trusted AI agents to work across customer channels and internal workflows.
00:10:30.320 --> 00:10:33.840
So right now you do have to reach out to their sales team if you want to access.
00:10:33.919 --> 00:10:42.480
So it's not like you're just gonna log into ChatGPT or log into your uh, you know, your company's uh, you know, open AI playground and start deploying these things.
00:10:42.639 --> 00:10:46.480
So it is uh a very limited uh rollout right now.
00:10:46.639 --> 00:10:56.799
But what they say, this is trusted AI agents, and they say each agent improves through real-world experience with evaluations, guardrails, and human approval governing every change.
00:10:57.519 --> 00:11:05.440
So, kind of the uh the three ways that they break this out to explain it is they say that this gives you one presence across every channel.
00:11:05.600 --> 00:11:10.639
They say you can show up in consistent in a consistent way across voice and chat.
00:11:10.720 --> 00:11:11.840
So ultimately, right?
00:11:11.919 --> 00:11:15.440
So if your company is looking to deploy voice agents, right?
00:11:15.519 --> 00:11:23.200
So a phone number, maybe you don't have all the humans you need to handle the volumes of calls, that's one thing, but then also on your website, uh right.
00:11:23.279 --> 00:11:29.679
So it's it's kind of a uh one system that can work and stay in sync with both.
00:11:29.840 --> 00:11:33.200
So they say show up in a consistent way across voice and chat.
00:11:33.279 --> 00:11:41.519
You decide what remains consistent, such as policies, evaluations, and escalation rules and what should change for each workflow or channel.
00:11:41.679 --> 00:11:55.919
Then they say trust built into every uh deployment, connect agents to company systems, define policies and permissions, and validate performance through simulations, evaluations, guardrails, and escalation paths.
00:11:56.240 --> 00:12:00.480
And then last but not least, they say improve quality continuously.
00:12:00.559 --> 00:12:12.799
Uh, they say presence continuously gets better with use, it uses production conversations and quality signals to recommend improvements that make agents more capable.
00:12:13.120 --> 00:12:26.320
And then OpenAI has shared some use cases with some big companies that they've been testing this out with, such as BBVA, Softbank, um, IHE, and then obviously their own internal use cases as well.
00:12:26.399 --> 00:12:34.559
So things that they're saying this can be used for is customer support, demand generation, claims, procurement, it helped desk, uh, human resource, etc.
00:12:34.799 --> 00:12:53.440
But essentially, anywhere right now where maybe your company is experiencing human bottleneck and you can't really scale um at the point that you need to, whether it's for customers or maybe if you're a large organization, sometimes it's an internal, uh, you know, almost ITHR uh type shortcoming that you're running into.
00:12:53.519 --> 00:13:02.960
Um, and and you know, I kind of want to go back to this concept of what AI unlocks, and I think that initial voice agents maybe got this all wrong.
00:13:03.120 --> 00:13:21.440
Because what I think initial voice agents that were too early, it was more of just uh scaling this technology out to as many people as possible, but in a similar way, which I think is not the point of generative AI and artificial intelligence, right?
00:13:21.600 --> 00:13:25.279
I think a lot of people I talk about this a lot, but they think, oh, this is an easy button.
00:13:25.360 --> 00:13:29.759
Let's get a blanket approach that works for everyone, which I think is the absolute wrong approach, right?
00:13:29.919 --> 00:13:39.440
The way that you should be doing this, and I think kind of tying in uh the you know website chat side with the voice side is very helpful, right?
00:13:39.519 --> 00:13:46.559
I think this is you know, we'll see what kind of B2B applications work well for this, but I think you should think of B2C, right?
00:13:46.639 --> 00:13:51.360
So if you're uh you know working at a company or maybe a consumer, right?
00:13:51.440 --> 00:13:54.559
Maybe you are constantly like Amazon, right?
00:13:54.639 --> 00:13:59.120
That's not to pick on one certain company, but you know, bless up my wife is always the one.
00:13:59.200 --> 00:14:07.679
If something happens with Amazon, she's the one that braves, you know, getting on the website and you know, talking on the phone to three or four or five or six different voice agents.
00:14:07.759 --> 00:14:09.919
I think Amazon is actually okay at this, right?
00:14:10.080 --> 00:14:16.240
But think of those companies, if it's not your company that could use something like this, think of those companies that you interface with.
00:14:16.320 --> 00:14:33.279
So I do think this is on the B uh B2C side, but having that data follow you around from whether it is you are online chatting or on the phone and not having to re-explain those things or you know what went wrong with an order, because presumably it's all going to be tracked, right?
00:14:33.360 --> 00:14:39.120
That's the thing that open AI says here, you know, talking about how it connects to your own company systems.
00:14:39.279 --> 00:14:58.159
Uh, so in the same way that you know, most companies have been rushing to get their data in order, you know, over the past, you know, post-Chat GPT phase when they're like, oh my gosh, all of a sudden, we really need to have our data in order uh for AI systems to be able to read and make use of this, right?
00:14:58.320 --> 00:15:05.679
This is this holds true uh for voice and customer success or customer experience uh agents as well.
00:15:06.159 --> 00:15:12.559
So that's a kind of brief overview of what OpenAI is uh saying that this is.
00:15:13.039 --> 00:15:15.039
So it's not a new model, all right?
00:15:15.200 --> 00:15:19.039
It's just kind of the infrastructure that surrounds their voice models.
00:15:19.279 --> 00:15:26.799
So every presence agent, according to OpenAI, starts with one narrow job and only the system access that the job needs.
00:15:26.879 --> 00:15:34.080
So companies set the allowed actions, required approvals, and exact moments that presence should escalate to a human, right?
00:15:34.159 --> 00:15:35.759
So you can set that level.
00:15:35.840 --> 00:15:48.720
Maybe you set it very low, and I would probably recommend that when this does roll out to the masses, is you know, don't set that escalation bar high, especially when you're testing out a new technology in production, right?
00:15:48.799 --> 00:15:49.759
You have to get it right.
00:15:49.919 --> 00:15:57.200
And again, I think people always look at maximum autonomy when it comes to trying out new AI, which is the exact wrong way to look at it.
00:15:57.279 --> 00:16:11.679
You look at your uh low stakes, high quantity uh uh instances, and then you have to be able to develop uh a system internally in the same way that I've always encouraged companies to create your own benchmarks, right?
00:16:11.759 --> 00:16:13.519
But those are text-based benchmarks.
00:16:13.679 --> 00:16:21.759
Don't just look at artificial analysis, don't just look at, I don't know, whatever you're you're looking at, uh, arena or you know, humanities last exam, right?
00:16:21.840 --> 00:16:25.679
You can't look at a single benchmark uh for enterprise AI adoption.
00:16:25.840 --> 00:16:27.759
Those are great starting points, right?
00:16:27.840 --> 00:16:46.159
But you need to be uh for text-based or agent uh kind of agent-based um systems, you have to have your own internal text-based benchmarks that say this is what uh qualifies as a win, this is what qualifies as a loss, these are our guardrails, this is how we deploy this throughout our organization.
00:16:46.320 --> 00:16:54.320
You have to do the exact same thing for voice agents, which is a little more difficult because there's things like nuance and sarcasm in your voice.
00:16:54.480 --> 00:17:15.839
So that's one thing, and maybe one reason um why these voice agents, I think initially when we saw this first puff of smoke in 2024 with oh my gosh, voice AI agents are gonna take over customer service, and they well, they didn't, is because the models weren't even smart enough to get over that big uh gray area of hey, people are ambiguous, right?
00:17:16.000 --> 00:17:19.119
If someone responds yes, you can say yes, right?
00:17:19.200 --> 00:17:20.559
That someone's excited about it.
00:17:20.720 --> 00:17:26.240
If they say yeah, right, that means they're maybe being sarcastic and not actually excited about something.
00:17:26.400 --> 00:17:39.920
So, models for the most part, I think couldn't even get over one of those first big initial humps because the reason why most companies have right customer support with real humans is to understand that human nuance.
00:17:40.000 --> 00:17:54.240
Uh, so if if if all the model was doing in 2023, 2024 was essentially translating uh your speech to text and then having a model in the background read that text, that's the reason why it just didn't work initially.
00:17:55.200 --> 00:18:01.759
So, uh what open AI actually bundles together is what presence actually is.
00:18:01.920 --> 00:18:10.160
So they're giving you the policies, uh the guardrails, approved actions, simulations, and evaluation tools inside of one product.
00:18:10.319 --> 00:18:15.519
So this is a codex-powered improvement process that's built directly into the platform, it's not sold separately.
00:18:15.680 --> 00:18:24.480
And open AI's own uh FDEs or uh forward deployed engineers, uh connect each company's systems and bring the agent live.
00:18:24.720 --> 00:18:28.160
So, yeah, we've been talking about this this FDE, right?
00:18:28.319 --> 00:18:30.880
Uh, you'll see it in our newsletter today, right?
00:18:30.960 --> 00:18:46.079
But Amazon just, you know, shut down kind of their AGI department or you know, laid off a bunch of people in their AGI department, and they're focusing more and more resources on FDEs, or you know, actually deploying putting their humans inside of these other companies.
00:18:46.240 --> 00:18:51.359
So it looks like that's what OpenAI's um kind of angle is here, right?
00:18:51.440 --> 00:18:58.160
They already have their FDEs, you know, big companies already have their dedicated open AI uh people working inside there.
00:18:58.319 --> 00:19:06.720
So this is just one other system that those FDEs or for deployed engineers will be able to implement uh for open AI customers.
00:19:07.279 --> 00:19:10.799
So here's the thing you have to talk about.
00:19:10.960 --> 00:19:34.400
Um is getting from beta, getting from demo to deployment, because that is what is ultimately going to decide whether OpenAI presence is kind of the next big wave of not just customer service, but maybe where all the other AI players will play, um, versus just, oh, it's a new feature that may or may not pick up steam.
00:19:34.559 --> 00:19:43.519
So, right now, the way that open AI frames it is that teams are able to simulate those common requests, edge cases, and higher risk scenarios before any customer actually sees the agents.
00:19:43.680 --> 00:19:48.559
So you can set those guardrails, humans can monitor and step in mid-conversation, uh, etc.
00:19:48.799 --> 00:19:56.480
And then after launch, codecs can investigate real production signals and propose updates that teams can actually test before rollout.
00:19:56.880 --> 00:20:00.880
So now let's try to understand a little bit of what's happening under the hood.
00:20:01.039 --> 00:20:10.079
And I think this part, especially if you use OpenAI's new GPT live, and it is bonkers, y'all.
00:20:10.160 --> 00:20:10.960
It is good.
00:20:11.119 --> 00:20:12.720
I wanted to do a show on this, right?
00:20:12.799 --> 00:20:17.519
But in our newsletter for our Wednesday demos, I usually have you all vote.
00:20:17.839 --> 00:20:27.920
Um, I really want to do a show on this because I think GPT Live, that's their new um essentially the replacement for advanced voice mode inside of Chat GPT.
00:20:28.079 --> 00:20:31.039
If you haven't used it, my gosh, it is good.
00:20:31.200 --> 00:20:42.960
Um, not just being able to understand your data, but you you know, I think you can finally experience talking to an agent and then knowing that it's actually calling tools and looking things up in the background.
00:20:43.119 --> 00:20:54.319
Whereas before, I think live voice agents, whether it was Chat GPT's advocate uh advanced voice mode, uh, you know, Gemini Live, those are probably the two uh, you know, big consumer versions.
00:20:54.480 --> 00:21:01.200
It almost seemed like all these voice agents would give you an intentionally uh ambiguous and vague answer.
00:21:01.279 --> 00:21:06.240
And you're like, okay, it seems to maybe be looking up something in the background, but maybe not.
00:21:06.319 --> 00:21:08.079
And it's almost like you couldn't tell.
00:21:08.240 --> 00:21:23.039
I remember one time I was literally just trying to test like you know, like the hallucination or ambiguous rate uh with uh you know, head to head between Gemini Live and uh the advanced voice mode with ChatGPT, but regardless, the technology just wasn't good.
00:21:23.119 --> 00:21:27.920
But I think this last iteration from OpenAI makes it good and it's multiple models.
00:21:28.079 --> 00:21:31.920
So, first, let's talk about GPT Real Time 2.
00:21:32.079 --> 00:21:35.839
Well, technically it's now GPT Real Time 2.1 because they just came out with an update.
00:21:35.920 --> 00:21:41.839
But that's essentially the developer tool for building custom voice agents through OpenAI's API.
00:21:42.000 --> 00:21:44.640
So they essentially have this technology, right?
00:21:44.720 --> 00:21:47.519
Let's just call it the super smart AI voice, right?
00:21:47.599 --> 00:21:54.079
So if you're a developer and if you want to build with this new super smart AI voice, you would use GBT Real Time 2.1.
00:21:54.240 --> 00:21:59.599
If you're a consumer and you want to experience it just using your Chat GPT account, that's called GPT.
00:22:00.240 --> 00:22:01.039
Live one.
00:22:01.119 --> 00:22:04.000
So both of these are new in the last two weeks.
00:22:04.079 --> 00:22:21.039
So that powers the new chat GPT voice where you can hold a live conversation while the AI quietly handles harder work off to another model, whether it's calling tools, looking at your data, uh searching something on the internet, uh, thinking, reasoning in the background about something that would normally take an AI model a little bit longer.
00:22:21.119 --> 00:22:21.680
Uh, right.
00:22:21.759 --> 00:22:31.039
The last technical details we we saw was that they kicked this over to GPT 5.5 and it would think and reason about things while the real-time model kept the conversation going.
00:22:31.200 --> 00:22:34.160
Uh, so presence though sits above both.
00:22:34.400 --> 00:22:36.720
So it's not a new model, right?
00:22:36.799 --> 00:22:43.039
It sits above both, and open AI has not yet revealed uh which model actually powers uh presence.
00:22:43.119 --> 00:22:47.839
Although I'm guessing it's probably something on the GPT real-time to side.
00:22:47.920 --> 00:22:48.960
That would be my guess.
00:22:49.599 --> 00:22:50.160
All right.
00:22:51.200 --> 00:22:54.319
So why do you have to pay attention to this?
00:22:55.200 --> 00:23:04.319
And I think this is something where I think trust either increases or erodes depending on your experience, right?
00:23:04.480 --> 00:23:06.400
A very simple example.
00:23:07.680 --> 00:23:10.160
I don't I forgot what the company is, right?
00:23:10.240 --> 00:23:14.079
But it's it's either a heating or air conditioning company that I use here in Chicago.
00:23:14.240 --> 00:23:18.079
They had their voice agents and they're actually tricky at first.
00:23:18.240 --> 00:23:21.599
The first time I called, I don't know, heater, air conditioner wasn't working very well.
00:23:21.759 --> 00:23:24.960
And I'm like, wait, this kind of sounds real.
00:23:25.359 --> 00:23:32.319
And then instantly, obviously, being around AI, I knew it was AI and I was having fun trying to you know jailbreak it on the phone, that kind of stuff.
00:23:32.400 --> 00:23:38.480
Um, but it couldn't get human handoff, and that was a bad thing.
00:23:38.559 --> 00:23:50.799
Uh, because ultimately what matters here, it's not about the technology, it's about are consumers' uh queries being resolved or not, or are people just more frustrated?
00:23:51.039 --> 00:23:54.000
But the the reality is, I think consumers want this.
00:23:54.079 --> 00:24:00.960
There was actually a study from 5.9 that said that 80% of consumers are willing to use AI customer service, right?
00:24:01.200 --> 00:24:16.000
Yes, I know there's some companies out there that do customer service well, but I would say this is an argue that is literally just uh sorry, this is an area that is literally begging to be disrupted because it is so bad, the quality is so slow, and ultimately, I don't know.
00:24:16.240 --> 00:24:21.119
Anytime I have to make a call to customer service, I feel there is no resolution.
00:24:21.279 --> 00:24:26.079
So maybe this is something I'm personally rooting for for the industry, right?
00:24:26.160 --> 00:24:37.039
Um I don't let that impact, you know, if if you're a decision maker at your company, don't let that impact, you know, your decision on whether you should be using these types of voice in real-time agents or not.
00:24:37.200 --> 00:24:41.759
But regardless, the consumer demand is almost overwhelming, right?
00:24:41.839 --> 00:24:50.400
You've you've seen this, whether it's uh it has any merit or not, you've seen this recent kind of uprising against AI.
00:24:50.480 --> 00:24:53.119
And I'm not just talking about AI slop, right?
00:24:53.359 --> 00:24:57.680
Um, companies I think are deploying sometimes AI in an irresponsible way.
00:24:57.759 --> 00:25:01.680
Uh, when it comes to job creation, job growth, uh job growth, things like that.
00:25:01.759 --> 00:25:12.799
So sometimes there's this kickback against AI, uh, but at least when it comes to the appetite for uh deploying real good AI agents, it seems like customer service, right?
00:25:12.960 --> 00:25:14.480
Consumers want this.
00:25:14.640 --> 00:25:23.279
So if you are a business leader, maybe you haven't even uh, you know, maybe you're in the in the space where it's like, yeah, we don't even really offer customer service.
00:25:23.440 --> 00:25:25.039
Well, maybe you probably can.
00:25:25.440 --> 00:25:28.000
Uh so that's another thing to uh think about.
00:25:29.200 --> 00:25:35.279
So, like I started the show off with as we wrap here, this was a quiet launch, which was interesting to me.
00:25:35.359 --> 00:25:40.640
And presence, I think, drew so little attention because essentially it's not sexy, right?
00:25:40.799 --> 00:25:45.359
Enterprise infrastructure doesn't look exciting when it's first launched.
00:25:45.440 --> 00:25:46.960
But that's what powers this thing.
00:25:47.039 --> 00:25:52.480
If you don't have the infrastructure set up, this isn't going to be a reality for you.
00:25:52.720 --> 00:25:58.079
So I think that obviously customer service organizations they do have the infrastructure, but what about everyone else?
00:25:58.240 --> 00:26:05.680
What if customer you know, service or phone service is, you know, not even a top 10 priority for you, but you've always been uh offering it.
00:26:05.839 --> 00:26:12.079
Uh, I I think customer service though may become AI's kind of first mass consumer disruption, right?
00:26:12.240 --> 00:26:14.240
Aside from text, that's that's come and gone.
00:26:14.319 --> 00:26:16.559
So maybe it's the next big iteration.
00:26:16.720 --> 00:26:25.200
Um, yes, we've all seen AI photos and AI videos um in advertisements, but I think that's going to continue to blur the line between what's real and what's not.
00:26:25.279 --> 00:26:51.519
But I think what is real is that this is a space that is probably likely going to explode because, like I started with uh both OpenAI, uh Google, um, Anthropic, and even technically Grok, all the players are starting to in are starting to invest heavily in this space because the technology is finally good enough where this can actually be a customer service tool that is worthwhile and that uh the users want.
00:26:52.079 --> 00:26:57.920
But whether that helps customers depends entirely on what companies optimize this first.
00:26:58.079 --> 00:27:08.240
And if you are just looking for that big red easy button, or if you have the infrastructure in human power to actually deploy this in a responsible way that helps meet that demand.
00:27:08.400 --> 00:27:12.880
All right, so that's a wrap on what's new with open AI presence.
00:27:13.039 --> 00:27:24.559
And hey, my kind of hot take uh takeaway on this is like I said, I'm rooting for this, not just for open AI presence, um, I'm rooting for this segment.
00:27:25.200 --> 00:27:31.680
Because personally, as a consumer, I have hated customer service for the longest time.
00:27:31.839 --> 00:27:37.759
I've very rarely had a good experience, and at least according to studies, it seems like that is normal across the board.
00:27:37.920 --> 00:27:43.279
So, in my uh in my experience, this is something that I think enterprise companies should be looking at.
00:27:43.440 --> 00:27:51.200
Obviously, if you're already in the B2C space, if you already have dedicated customer support, there's probably a good chance that you have the infrastructure to take advantage of this.
00:27:51.279 --> 00:27:52.480
So you should be looking at it.
00:27:52.640 --> 00:27:56.559
But for everyone else, I think this also begs the question: should we be doing this?
00:27:56.720 --> 00:28:10.799
Whether it's for or maybe you're in the B2B space, maybe you are in the B2C uh space, but you just haven't had kind of a dedicated customer phone or customer chat service, and maybe you should, or even for larger organizations looking to deploy something like this internally, right?
00:28:10.960 --> 00:28:25.200
If you have tens of thousands of employees, being able to give that level of personalization, like I was talking about easier, uh or like I was talking about earlier, where I think companies, when they find generative AI, they're like, okay, we could just put out more content that appeals to everyone.
00:28:25.279 --> 00:28:36.400
When I think it's the reverse, you have to uh ingest those signals, ingest those data points, and you should just be putting out more personalized uh messaging and more helpful messaging to more people.
00:28:36.480 --> 00:28:38.799
So maybe that's internally, maybe that's externally.
00:28:38.880 --> 00:28:42.319
But regardless, I think the technology is finally there and that's exciting.
00:28:42.480 --> 00:28:46.720
And well, we'll see if open AI presence gets more eyeballs on it.
00:28:46.960 --> 00:28:49.119
So uh that's a wrap for today's show.
00:28:49.200 --> 00:28:49.920
Thank you for tuning in.
00:28:50.000 --> 00:28:55.759
If you haven't already, please go to your everydayai.com, sign up for the free daily newsletter, and make sure to tune in tomorrow.
00:28:55.839 --> 00:28:57.519
We're gonna be going over our Friday features.
00:28:57.680 --> 00:29:02.319
There's a lot of new stuff that dropped both uh yesterday and that I know is coming today.
00:29:02.400 --> 00:29:03.920
So you're not gonna want to miss that one.
00:29:04.079 --> 00:29:04.880
Thanks for tuning in.
00:29:04.960 --> 00:29:07.440
We'll see you back tomorrow and every day for more everyday AI.
00:29:07.519 --> 00:29:08.480
Thanks, y'all.
00:29:09.920 --> 00:29:13.200
And that's a wrap for today's edition of Everyday AI.
00:29:13.359 --> 00:29:14.480
Thanks for joining us.
00:29:14.640 --> 00:29:18.400
If you enjoyed this episode, please subscribe and leave us a rating.
00:29:18.559 --> 00:29:20.160
It helps keep us going.
00:29:20.400 --> 00:29:27.440
For a little more AI magic, visit your everydayai.com and sign up to our daily newsletter so you don't get left behind.
00:29:27.680 --> 00:29:30.319
Go break some barriers, and we'll see you next time.
