← Power Hour Optometry
Power Hour Optometry
Building AI for Optometry: A Live Walkthrough of Practical AI You Can Use in Your Practice
00:56:02
Loading summary
Transcript118 lines
- [00:00]
A
Welcome to the Power Hour, Optometry's biggest and longest running show. I'm your host, Eugene Shotsman, and I am so excited because today we're doing something we've never done before on the Power Hour. As you know, I'm a huge advocate of AI for just. For the simple fact that it can offer, in many cases, a better patient experience, greater practice efficiency, and be a driver for revenue. So today I got one of the top AI consultants to talk AI with me. But we're not just talking about artificial intelligence, we're about talking. We're actually going to be building an AI tool live on the show. So my guest is Obi Bharacharya, and he has been on the front lines of AI innovation for years. He's worked at athenahealth, he's helped develop Amazon Alexa, he's consulted with vsp, and he's joining me to prove that creating practical AI solutions for optometry does not have to take months and months and months or years and cost a fortune. And because together we're literally going to build a voice agent from the ground up that calls and confirms patient appointments live from scratch in about 30 minutes on the show. And the coolest part is you'll actually get to hear this voice agent call me at some point during the episode. So it's kind of like a neat little preview of something I'm super passionate and very excited about. A voice AI agent that can take calls that are coming into your practice and schedule patients at a really high close rate. Something that my team is working on and something I'm super, super proud of. But you get to get a look behind the scenes. And the point of this episode is not to, like, transform you into an AI developer, but just to show you how accessible some of these tools really are for practices. And if you ever thought AI sounds great, but, you know, where would I even start? This episode is for you because you'll see what's possible, how fast it can be built, and how there are tools out there that can build workflows that can allow you to take the busy work off of your team's plat. So my suggestion, and actually this is my literal ask during the show, is that I need your feedback. After the episode, I need you to send me a Note on Eugene Shotsman.com or on the Power Hour website, because Obi said he will be back and we will build something that you can use in your office during our next show together. My initial vision for the episode that we're going to build Two tools of different flavors, and he decided to tackle the harder one first. And we kind of ran out of time, so he, rather than me, tell him what next tool will be on the next episode. We're going to ask you, the audience, to write in, give me some feedback, tell me what you want us to build, and we'll literally build it live on the show, and I'll do my best to make it available to everybody. So before we dive in, Obi asked me to share that everything that we talk about on the show is purely for educational purposes. He has no interest in any of the companies we discussed. And make sure that if you're experimenting things, you know, the laws in your state and HIPAA rules and all that stuff. Again, I am really looking for your feedback. Any additional ideas for future episodes, you can reach out to me directly at Eugene Shotsman.com or at the Power Hour website. And now let's go build some AI live on the Power Hour. All right, Javi, welcome to the Power Hour. Excited to have you on the show, man.
- [03:23]
B
Thanks, Eugene. Really, really excited to be here. Yeah.
- [03:26]
A
And, you know, it's really. I think it's super helpful if we can give some context to the listeners as to where you're coming from, and you've been on the forefront of this AI innovation curve the. The entire time. So maybe give people. Before we do what we're about to do on this episode, let's give them a little bit of context so that people understand what your experience is like with AI and technology in general.
- [03:53]
B
Yeah, no, thanks. So, you know, AI is not new. AI has been around for literally decades. You know, if you go back to, I'm an engineer, so, you know, when I was learning how to code 25, 26 years ago, you know, we. We would start with things like if statements, and everybody kind of knows, like what an if statement is. You know, if something happens, do this.
- [04:14]
A
Right.
- [04:15]
B
That's an algorithm. It's a form of intelligence. Right. And it's artificial. What's new is generative AI. And in the last 10 years or so, as I've been working in health tech, I cut my teeth at Athena health, working on EMRs incorporating, you know, AI of what I call kind of the last century or the 2010s, right into our EMR. And from there, I went on to. During the pandemic, I was working on a trusted startup called Trusted Health. We had built a marketplace of nurses, connecting them with health health facilities on demand during the pandemic to kind of fill staffing buffers. We were using AI again there, you know, to figure out how we might predict staffing gaps on nursing units. We would look at census, we would look at people calling off, we'd look at leave and absences. We try to look at those demand signals and we'd try to determine like, all right, what do we, what do we need to fill these gaps? From there, I went to Amazon and I got kind of a front row seat to how LLMs are transforming these assistants that we use in every day in our everyday lives. I had the privilege of working on Alexa. We were trying to make Alexa smarter using LLMs. And that generative AI component is the new thing where AI now has the ability to essentially do pattern matching at scale.
- [05:36]
A
Right.
- [05:37]
B
You can give it some context. Alexa, stop. Alexa wants to join this call. So I don't know if you can hear Alexa. That's hilarious. So good to hear it. Okay. Oh, fantastic. So anyway, so I've had kind of a front row seat to this from where it started.
- [05:58]
A
Right.
- [05:59]
B
Kind of traditional algorithms all the way to generative AI. And what's been really, really impactful for me in vision care specifically is how much of an opportunity there is to not just use generative AI, but any kind of intelligence in some of these workflows. I think I'm amazed at how much practices still do manually. So where you don't even need generative AI, a little automation could help.
- [06:27]
A
So, yeah, and I think this is why we wanted to do this wild and crazy thing on the episode today where I want. So some of you know, and I think some of our listeners know that I'm on this tour where we talk about practical applications of artificial intelligence in eye care. That, you know, there's all these emerging technologies that kind of solve a problem, sort of give somebody an opportunity. But I think the industry is still really not truly fully adopting the full potential of AI. And I think the. The feedback that I hear, oftentimes it's like, it's hard. It takes a really long time to do. And so we wanted to try to disprove that a little bit on the show. Today is going to be crazy. First time we've done anything like this ever on the Power Hour. But we're literally just going to build a high tools for the next 30, 40 minutes. And for those of you who are watching on YouTube, you'll get a chance to do it or you get a chance to watch us do it. For those of you who are listening, I'm going to go ahead and Narrate for you so you kind of know what's going on on the screen. But you may also enjoy jumping on the YouTube video and watching it happen, if you're interested. But Abhi, I'm going to turn it over to you and I'm going to just give you two challenges. We're going to solve two different problems today. Problem number one, let's do some sort of voice agent workflow where we're basically following up or maybe confirming an appointment. Does that sound reasonable?
- [07:59]
B
Sounds like a pretty, pretty common one. So. Yep, I like that.
- [08:03]
A
Straightforward. So we'll test that out. And then the second thing is, let's solve another problem, which is I've got an employee who I want to make better at doing patient counseling for dry eye cases. Right. Because what happens usually is I've got doctor talks about a treatment plan and it's $2,000 and the patient walks out to go talk to the employee. And the employee is like, yeah, I know, it's kind of expensive. And then the patient walks away without buying anything. And that's where I would like to see an AI solution. Be like a master sales trainer for that particular dry eye patient, counselor, employee. Now, full disclosure, you're hearing that idea for the first time, right?
- [08:52]
B
That's right. That's right. Yeah. No prep work. Yeah, that's right.
- [08:57]
A
So you decide which one you want to handle first. And we're going to do one before the break and one after the break. But we're literally just going to try to build it live here on the Power Hour just to show people how you can get off the ground and running without really having to spend years and years on developing something.
- [09:16]
B
Yeah. And I think that's the biggest thing that I've been seeing is missing is that folks are talking about AI. They're talking about it a lot, you know, in shows and on YouTube, all over the place. Right. People are talking about it a lot. There's not a lot of practical, like, you know, show me, right? Like, show me exactly how I can apply this in my practice. So what I've been doing is kind of getting past that fear because I think a lot of times people get stopped or blocked and, you know, like you just said, this is too hard, it's going to take too long. And instead of being blocked by that, accept that this is going to be a testing and learning process. Even in this scenario today, like, you know, we're going to walk through these couple scenarios. You going to see errors, you're going to see bugs. You're going to see issues, and that's okay. Like, it's all right. Don't let perfection be the enemy of done right. And I think that's the big thing I want the audience to take away is you don't have to have a perfect plan before you just start playing with this stuff. So. Excellent. So let's start with the appointment reminder, because I think that one is a fun one and I'm just going to share my screen. Um, but before I, before I do that, I, I sometimes like to kind of do these sessions by starting off with kind of the end product because it's a little bit like a cooking show, you know, like you end up with like the, the end product and then they, they, they kind of like, all right, let me show you how to make it right. And they kind of walk through it. But then, you know, like the, the, the montage always kind of like fast forwards to, here's the end product. Let me show you the end product first. All right, so, so I'm just going to have the agent because it's. Well, this one I did have a little bit of time to prep and I've done it before for, for other folks. So I'm going to have the agent actually just call my phone. It's my phone right here. And it's just going to call me and it's going to do an appointment reminder. Sound good?
- [11:12]
A
Let's do it.
- [11:14]
B
I'm just going to have it execute this workflow. I'm getting a phone call over here. Hey. Hi, may I speak with Abby? Yeah, this is him. Hi, Abby. This is Vision Care Optometry calling. I'm just reminding you about your eye exam appointment tomorrow at 2pm can you confirm you'll be able to make it? Oh, man. Yeah, I totally forgot about that. But yeah, I think I can make it tomorrow. Yeah, sure. Perfect. We'll see you tomorrow at 2pm if anything changes, please call us at 555010. Have a great day. All right, thank you.
- [11:55]
A
Cool.
- [11:56]
B
So that's Proof of Life. That was the agent giving me a call. So let's see how, let's break it down. Like, you know, it was pretty simple, right? Pretty simple phone call. But how did we actually get to that point? Right, so I'm going to share my.
- [12:12]
A
Screen and for all of you listening as Avi shares a screen, I will narrate for you so that you know what we're looking at and kind of what, what we're dealing with. So as that's Coming up. Tell us what tools you're using.
- [12:28]
B
Absolutely. So what I like to do is always use stuff that's off the shelf as much as possible, right? Because you can write code and kind of get really deep into this stuff, but what can you do quickly? So when you look at the voice part of it, like the thing that actually called me, there's a tool called VAPI AI that allows folks to create voice agents really, really quickly. So that actually handled the voice part of it that actually made the phone call. But just like a human, the voice and making the phone call is only part of this, right? You need something that actually knew what appointments were coming up that day, right? To, like, you know, decide, all right, I need to call these people. You needed something to be able to, you know, actually extract kind of the information from that. That list, right? Make the phone call and then go back and update, you know, the. The list and say, yep, I made that phone call. So in real life, if somebody was doing this at a practice, it'd probably be somebody inside of an EMR pulling a list of the appointments that are coming up that day, picking up the phone, making those dials, and then logging that, right? So for that part of that brain that's kind of doing all that work, we use a product called N8N and. And N8N is a workflow automation tool. We're going to use N8N today to essentially be that task manager, right? Kind of how do I get data from the spreadsheet? How do I send it to VAPI to make the phone call? And then how do I ultimately kind of get the results of that back into a spreadsheet? So that's kind of the tools. Vapi and N8N, those are kind of tools. And then the last piece is we're using Google Sheets, like just with a list of appointments in here. So that's kind of the input, right? So before we get started, one of the things I want to say is you might be thinking right now, well, gosh, how is this going to work with my emr? How am I going to get it to sync with EMR and all that? What I tell people is don't try to focus on, like, the hardest parts of the problem first. You know, integrating something into the EMR is time consuming. So figure out ways to get the data out of there quickly. Like if you just did an export, right, into an. Into a Google Sheet or an Excel spreadsheet the day before, whatever, it's an easy way for you to get started with a voice agent. Because the thing that you're really trying to automate here is that time consuming making phone calls.
- [14:54]
A
Right.
- [14:55]
B
You know, you probably have somebody that's told you that, hey, if you make a phone call and you make an appointment reminder call, you're more likely to have the person show up for their appointment.
- [15:03]
A
Right.
- [15:04]
B
You want to try to automate that work. You're not necessarily trying to automate the entire thing. Right. So focus on something small that you can just get out. That's going to make an impact. Don't get caught up in, oh, gosh, how does this integrate with my emr? There are ways to do it, but you don't need to focus on that right now. The big thing is how do I get these calls off my plate?
- [15:22]
A
Well, yeah, and I think for the purpose of today's discussion is it's not necessarily to have you buy an off the shelf solution. And in fact, you know, full disclosure is, you know, be my organization is working on an appointment scheduling AI tool that will be released in a few weeks. And the reality is that it's exactly the same process we just went, we're going to go through. That is similar to the process we went through. And of course, eventually you're right. The hardest part, you have to connect it to the ehr, which, you know, and we ended up connecting it to like 10ehrs or 9ehrs or something. But reality is that you're absolutely right. The hardest part first is let's, let's figure out where we can save the time. Right. Like or sorry. The most impactful part is let's figure out where we can save the time. How do we make this, how do, how do we get this to do something that ultimately has an impact on the practice?
- [16:20]
B
Totally. So let's talk about how we would actually build this agent. Right. What I want to show you is how you can use Claude to actually help you guide exactly, step by step on building this workflow yourself. I always like to do this so that you're not alone in building this. A lot of times I watch like a YouTube video or watch a podcast and like somebody's doing something really cool, you know, you might lose that video, whatever. Remember, you've always got cloud. So I'm going to start with a prompt and I'm going to start with a role. I'm going to give Claude a role to follow. Because think about in real life, if you're going to build this voice agent, who would you want to talk to? You probably want to Talk to like an expert AIML developer who has experience building voice agents.
- [17:09]
A
Maybe somebody who has the knowledge of the entire Internet in their back pocket too.
- [17:15]
B
Exactly. Which Claude happens to have the entire knowledge of the Internet, which is nice. And that gives it that role so that Claude actually could go in and bring that relevant knowledge to this task. So now I'm going to say, do.
- [17:28]
A
You always give your prompt like when you're talking to chat, GPT or Claude, do you always start with a role?
- [17:33]
B
Almost always. And I find that it gives me much better results than when I don't start with a role. And I think it's because if you don't give it a role, it's, it is literally got the whole context of the Internet, right? So it's, it's trying to like guess exactly at what you're trying to do.
- [17:51]
A
Right?
- [17:52]
B
And that's good. Sometimes when you're exploring, you're kind of maybe let's say doing like market research or something like that. Like, hey, show me the competitive optometry practices in Napa, California. Right? Like, you know, you know, it's fine to kind of ask it questions like that in that way. But let's say now you're like, hey, I want to actually build a website that's better than all of the competitors in my area. You want to give it a role at that point, right? And say, hey, I want you to act as an expert web designer, expert social media manager, whatever, right? To kind of give it that role so that you can get the artifact that you want just like real life. You know, I always say think about who you would want to talk to in real life when you're trying to get an outcome from AI and then give the AI that, that role so that it can emulate that. There's some really fun ones I've done. Like, hey, pretend that you're Steve Jobs, right? And act as your Steve Jobs. Blow up my startup idea for me, right? Like totally tear it down, right? Acting as Steve Jobs. And it's, it's quite entertaining. It will actually kind of like emulate what what Steve would say or somebody else would say.
- [19:05]
A
That's cool. That's super cool, smart.
- [19:08]
B
So act as an AML developer who has experience building voice agents. Give me a step by step guide in building a voice agent for doing appointment reminders for my optometry practice using N8N and Vapi AI, which were the two tools that we, you know, talked about that we're going to use for this, this exercise. So I want to just send that through. And what Claude gives me back now is, gives me an exactly. Kind of a step by step workflow, right? Exactly. What I have to do in Vapi, what I would have to do in N8N. It kind of just gives me the whole thing, which is really nice. So I'm not going to go through obviously this particular step by step for this particular example, because we've built it so many times. I can show it to you from memory. But this is, this is an excellent way to kind of think about doing something you've never done before, right? Use your. Use your friend to kind of help you kind of get that first, first version out. All right, so let's jump in. So I think just for this exercise, what we're going to want to do is take some appointments in and we're going to assume, just for the sake of like making this simple, going to assume that you have the appointments in like a Google sheet, right? You can see like in here you got patient name, phone number. I kind of just made this before, but these are just kind of patient name, phone number, appointment date, appointment time. And then, you know, here you got reminder sent or no, no reminder sent, like, what is the status of that appointment? Like, has a reminder actually been done or not? You could export this from your emr, right? And you might be thinking, well, gosh, I'd love for it to just integrate with my emr. And this again, I'll remind you, like, start with the thing that you're actually trying to automate first, right? You're trying to automate the time consumption of all making those phone calls. That's really the time consuming thing here. Focus on that, the EMR integration. That's something you could do later, right? Just focus on the thing that's most impactful, which is making those calls. So imagine you've got a whole bunch of these patients in here, right? And then we're going to go to vapi. And what VAPI is, is again, it's that voice skill, right? It gives you the voices, it gives you the models. It's got. The way I kind of think about it is if AI was a human, right? The VAPI is like the mouthpiece, right? It's the ability to kind of talk, right? So I'm going to open this up and we're going to build the voice part of this first. And arguably it's kind of the most fun part because building the phone, the. The voice part of it is, you know, you get to kind of pick the personality and all that kind of stuff is really cool. So. So I'm gonna go to assistants over here, and you're gonna go and create assistant. And there's a bunch of, like, templates in here that you can choose from. You can pick, like, you know, like, you know, even has, like, an appointment scheduler. It even has some, like, healthcare things like care coordination or whatever. I'm just going to start with a blank one, and I'm going to call this assistant, Eileen, and we're going to create this. And what you see over here, you get, like, kind of the cost over here you get, like, models. So again, going back to that human analogy, right, The AI that you're going to be using here to kind of make these phone calls, it needs a brain. And the model is kind of the brain of any AI assistant. So here you can choose kind of what you want to use. I don't touch this. I'm just going to keep the defaults. Already has OpenAI in here. That's great. And then there's a couple things in here that you're going to want to deep dive a little bit on. So there's this concept of the first message. This is what the AI agent is going to say when it picks up the phone and when it calls the person. Right. What is the first thing that you're going to want it to say?
- [23:12]
A
Right.
- [23:13]
B
And then there's this thing called a system prompt. The system prompt is essentially kind of all of the directives that the AI needs to follow in order to complete its job. All right, and these are blank right now, right? These are completely blank. So what are you going to do here? These are blank. How do I kind of fill this out? Let's go back to our friend. Right, so this is a concept called tool chaining, right. You're going to go back to Claude and have it generate that stuff for you. So now you're going to say, okay, I need a system prompt for my voice assistant. Her name is Eileen. So now it's going to generate a system prompt.
- [23:57]
A
Right.
- [23:58]
B
So it kind of just gave me a whole bunch of, you know, information here.
- [24:02]
A
Right.
- [24:03]
B
So I can take this and I can simply come in and cut and paste it. Now you can obviously read it. You get more time. You can get read it and kind of be like, oh, you know, a.
- [24:13]
A
Little bit more of the. For. For those not watching but listening. Yeah. Give them, like, at least some of the categories that are part of that system prompt, just so that people understand what a system prompt actually is. Yeah.
- [24:23]
B
Let's go a little deeper. Good call. So it's very much similar to kind of how I was describing how you prompt Claude to begin with. So you'll see the first thing it actually starts with is role. So you want to give Eileen here a job. You know, your job is to remind patients about their upcoming eye care appointments and confirm that they're going to show up, right? You represent the practice with professionalism, personable and helpful. I call this the job description.
- [24:49]
A
Right?
- [24:50]
B
You are now giving the AI a job description. It's important to kind of think about this as like hiring an employee. You know, like when you're doing this, you know, you're, you're, you're, you're hiring an employee. So what would you tell an employee?
- [25:02]
A
Right?
- [25:02]
B
A new employee that you were hiring? You'd probably give them a job description to start, right? You'd probably tell them how you expect them to interact with your customers. You'd probably give them the responsibilities that, you know, you expect for, for that employee to complete. So in this case, greeting the patient warmly by name, right? You probably have, I'm sure your staff and your practices probably have like little cheat sheets of exactly what they have to do for these calls. Think about some of that as context, right? Clearly stating the details of the appointment, date, time, who's it with, confirm whether they can attend, handle basic questions, collect that information if they need to reschedule or cancel, and then it goes into the guidelines of the conversation. How do you open? What's the confirmation flow look like? What are some questions that the patient might have that you might have to answer, like location and address of the office, things to bring to the appointment. How long would it typically take, right? Whether they would need to have their eyes dilated. This is an opportunity to maybe even pick up some signals, right? In that call for procedures or things like that that, you know, you might want to do in the office. You can also give it guardrails. You know, the, the data you're going to feed this thing. Remember, it's not a lot. You're only giving them patient name, phone number, appointment date, appointment time, and that's it. So they can't really answer medical questions, you know, so like if the patient asks, what was my last prescription? You know, you can't answer that stuff, right? As a, as an employee. And, you know, it's probably not worth.
- [26:32]
A
Go off script at all. I've seen. And there's people who, once they realize they're talking to an AI, they think it's really fun to try to break the AI, myself included, by the way, but I do it for science. The reality, you know, your office, that the law doesn't exist about this yet. Right. So your office technically could be liable if somebody gets into an engaging conversation with an AI and that AI really wants to be helpful and then it starts offering advice and then the person takes that advice and then gets hurt somehow. You know, all of that, those conversations are recorded. And so that's where guardrails are so, so important for just about any AI that you're using, especially one that's patient facing.
- [27:18]
B
Absolutely. And you know, the nice thing about Claude is you can even, you know, say, like I want to make sure that I've got the best guardrails in place.
- [27:27]
A
Right.
- [27:27]
B
That it stays on target, stays on topic, doesn't veer into those other, other areas. And I think a broader kind of takeaway here for implementing AI in your practice. One of the things I've been encouraging folks to think about is non clinical use cases because I think that's where a lot of the lift can actually be provided.
- [27:48]
A
Right.
- [27:48]
B
Is things like appointment scheduling or billing and this kind of, you know, these marketing and sales. Right. Like those types of functions are really ripe for AI versus like trying to get into clinical stuff which you know you're going to want them to talk to a, an OD about that stuff anyway. Right. So those guardrails are really good. Tone and style is kind of the next part of the system prompt. So here you can kind of tweak it. You know, patient and kind, especially with elderly patients, like brief responses. Don't over explain or ramble. This is actually a problem that AI kind of has in general is sometimes just gets very verbose and you know, again talking about, hey, if, if the patient seems confused based on like the responses, maybe speak more slowly and kind of gives it some of those triggers to watch out for. You know what, I'm going too fast. Maybe I should slow down. You might get some difficult situations as an, as a member of the, of the front office staff here, right. So handling difficult situations is another section in here. Wrong numbers. Somebody died, right? Calling a deceased patient, patient, an angry patient, somebody you know, just, you just copped them on the wrong, wrong day and the wrong side of the bed. A language barrier, which, what's really nice about VAPI is actually it's multilingual. You can actually have it speak in different languages, which is quite entertaining. So, so there's about 60 languages that it supports. But, but you can you know, you can kind of handle that here as well. Important rules, again, is more about guardrails. Don't argue with a patient if they insist that they don't have an appointment. Keep the calls under two to three minutes, end it professionally, and then there's some sample responses here. So if you all already do these calls and you already have like, let's say transcripts or you have scripts, this is a great place to put some of those scripts to kind of give the AI a little bit of knowledge of how these calls typically go. Right. And what are some of the typical responses? And that just helps the AI be as realistic as possible.
- [29:53]
A
And if we were doing this live, and if we were EHR connected, meaning, like if we were building like a full go to market tool, and if were EHR connected, you would imagine a world where you would also include in this, in this master prompt, you would include some information about, well, if the patient says, no, I can't make it, then you would also try to connect to the EHR and offer some, offer some opportunities to reschedule. Right. And you'd say, okay, well, I'm sorry, you can't make it. What's a good time for you? And then you would try to reschedule, which is again, one of the efficiency as you can gain by going with an AI tool that can actually work to reach the patient rather than just leave a voice message. And when it reaches the patient, what can it do? Well, one of the things it can do is if somebody says, I can't, then it can possibly do this in a, in a more effective and pleasant way than a, than a person saying, oh, well, then I got to put you on hold and see what we have available and that kind of thing.
- [30:53]
B
Absolutely. And what a, what a great experience for the patient.
- [30:57]
A
Right.
- [30:58]
B
Being able to seamlessly, right there, you know, get another appointment and get it rescheduled and go, go off and do something else. Right. And that's, I think that's some of the things that you could do. Like next, right after you kind of launch it, and you kind of like, all right, well, cool, this is working really well. How do I integrate it now with my EMR and do some of those more advanced workflows? So, so this is pretty cool. It's pretty good system prompt. I, I'm probably not inclined to change too much in here. It looks decent, so I've put it in here. I still have kind of this concept of the first message. I could type something in here. But since our friend Claude's on a roll. I'm just going to ask it give me a first message that is short and friendly.
- [31:46]
A
Let's ask it to give us three so that you and I can pick from them.
- [31:49]
B
Oh, you know what? I like that. Give me a few options for the first message for the voice agent that are short and friendly for Eileen to.
- [32:01]
A
Open calls with Claude's thinking, thinking, thinking. And here it is.
- [32:08]
B
Here we go. So we got Ultra short options, 10 to 12 seconds. Option one, let's say hi. Hi Eugene, this is Eileen from practice aim. I'm calling about your appointment tomorrow at whatever time it is with Dr. Whoever. Can you still make it seems very much like business. Just let's get down to business. Warm and efficient. Hi Eugene, it's Eileen calling from Bright Optometry. Just confirming your eye appointment tomorrow at 2pm Are we still seeing you? That also seems kind of like straight to the point, you know?
- [32:47]
A
Are we still.
- [32:49]
B
Yeah, I know. It's like, I mean, man, like butter me up a little bit. Casual, natural. Hey. Hey, Eugene, this is Eileen from Bright Optometry. Quick reminder, you've got an appointment with Dr. Obi tomorrow at 2pm you're good for that time. That one's a little better. Out of the three, I. I like that one the best thus far. So let's see, medium length. This is a little bit longer. He's like professional with personality. Hi Eugene, this is Eileen calling from Bright Optometry. I wanted to give you a quick heads up about your appointment tomorrow at 2pm with Dr. Obi. Will you be able to make it? Oh, this one's a good one. This next one. Good. Good morning. Like good time of day, right? Like good morning, Eugene. Eileen here.
- [33:37]
A
Right.
- [33:37]
B
I'm calling to remind you about your eye exam. Does that still work for your schedule? That's a good one.
- [33:42]
A
Yep.
- [33:44]
B
What do you think? You got a winner.
- [33:45]
A
Yeah, let's just go with that last one. It was good. Made the point that you can. And of course we can make some changes if we want to, but I.
- [33:54]
B
Like that you get.
- [33:55]
A
You get to control the process. Claude is your buddy, but is not necessarily your. Your end all be all.
- [34:03]
B
Yep, totally. So you got. You got first message in there and then one of my favorite things to do in here is picking the voice. So there's different providers of voices that you can actually choose from and you can pick even folks that have like accents. There's. There's all kinds of accents. There's a New York man who's actually quite entertaining in in this. But it doesn't seem like, like, like vapi or. Sorry, it doesn't seem Eileen friendly. But you can pick, like, different demographics, genders, accents. So there's a whole library of kind of different voices available. And it's. It's probably one of the more fun things to do is playing around with these voices. So actually, I'll show you real quick in here. Go to a different tab. It's got. There's a thing called voice library, and you can really kind of knock yourself out. I've spent countless hours just wasting time in here. But you can come in.
- [35:05]
A
It's actually really, really fun.
- [35:08]
B
Yeah, you can play. You can play with these voices, which is. Which is pretty cool. So I think just for this one, we're going to pick. I'm just going to pick a, A, a female voice. We'll call Paige, and then you can actually talk to the assistant live right here now. Won't do that entire workflow right now because remember, we haven't hooked it up to the brain, which is N8N. Like that's doing the task flow.
- [35:35]
A
Right.
- [35:36]
B
But if you just wanted to test, like, you know, hey, how does it sound? All that, you could do that right here. So you could hit talk to assistant and just say, yep, it's okay to use all these things. So it's actually now talking. Yes, it does. So anyway, I don't know if you could actually hear it because it's coming in through faint.
- [36:02]
A
But I. But I. I think I got the idea.
- [36:04]
B
Yeah. And when we get it, when we get on the phone call, you'll be able to kind of hear it right away. So. So what we're going to do is we're going to hook this up. We need to give the voice agent a phone number now so it can actually dial and stuff like that. So we're going to go to phone numbers, and I'm actually going to create a new phone number here. Vapi actually comes with 10 free phone numbers like that you can just use, which is nice. It actually has some free usage, too. So, like everything I'm showing you that I'm doing here, you don't actually have to start paying right away. It gives you some free credits to kind of play around with, which is nice. So I'm going to go in and I'm going to assign that assistant, which is, you know, we just called it Eileen. So we're going to just assign it to this phone number. I'm going to hit save. And what that does is just gives Eileen access to a phone, Right. You'll actually be able to call this phone number and talk to Eileen if you want, which I might have you do, Eugene, here in a minute just to kind of test it, because you could actually call it on your side if you like. So it gives you about a minute to activate. But while that's working, let's go to the workflow. So I'm going to actually create this from scratch. So N8N. So let me explain a little about N8 N8N. Like I said, it's kind of like this task manager. It helps connect the dots between. Hey, I've got this data sitting in a spreadsheet, right? How do I get this to Eileen so that she can make some phone calls? And then how do I get Eileen to come back and update the spreadsheet, right. And kind of put her notes in here that, yep, you know, she's showing the person showing up and the reminder has been done or whatever.
- [37:44]
A
Right.
- [37:45]
B
So N8N. Think about N8N as a way to just stitch those things together and, you know, make it so that these things are talking to each other. Right. So I'm going to come in, I'm going to create a new workflow. And remember, as I'm going into this, at any point, if you're feeling like, gosh, this is a bit much like, you lost me now, right? It was fun looking at the voices and all that, but this is just way too much work. Obi, you're killing me. Remember, you got your friend right here, right? Your friend is right here. At any time you can be like, yep, Claude is right here. Gonna be like, you know what? Give me a beginner step by step guide to configuring N8N. Now with this voice agent I configured in Vapi AI, right? And it'll, it'll walk you through as slowly as you want, right? And give you like little step, step by step instructions, including like, what website to go to. How do you kind of, you know, get in and start building this thing? So anytime you get stuck, you're like, man, this is just too much. You can always go back, right? So I'm actually going to start this workflow with a trigger. And a trigger is basically just how does this workflow start, right? And sometimes you can have like other things that are happening, like your website or whatever could trigger a workflow. You know, something happens in the emr. So for example, if this was like fully integrated emr, a trigger might be that the appointment is coming up, right? And it's coming up within the next 24 hours. That could be the trigger that starts this. But just for the simplicity sake, here we're just going to say somebody's going to. A human's going to come in and just hit this button and say, hey, I want to execute this workflow, right? That just makes it easy for us to kind of demo. So that's the very first step. So now what do we got to do? All right, then we're putting ourselves in Eileen's shoes. Now. Her job is to do these appointment reminder calls. What's the first thing she's going to need to look at? She's going to need to look at this spreadsheet and figure out who to call.
- [39:45]
A
Yeah.
- [39:45]
B
Who is she going to call? So let's make it so that she can do that. So I'm going to come in and we're going to get. We're going to find stuff from a Google sheet, right? So I'm actually going to just type in Google Sheets, which is nice. And I'm going to say I need to get a row from a Google Sheet, right? So I've got Google Sheets with all the patients in there, and I can come into here and just connect with Google Sheets right here. You can actually authenticate just using your Google account. Like, you just type in your credentials and it'll authenticate. I've already done it in this. And then you're going to find the sheet that you made. So in this case, this one's called Appointment reminders, right? And you're going to just pick the sheet, right? So this is sheet, sheet one, right? You're just going to pick the sheet that that's in. So come in and say, hey, sheet one. I want you to pull the first row in that sheet or the row that's in this spreadsheet so that I know who to call. And what you can do is you can actually test it. You can say, all right, I'm going to execute that node and kind of see what happens. But we're going to wait. We're going to keep building while we go through this. So you get that row, right? And then the next thing you're going to want to do is call the patient, right? I've looked at the. I've looked at this thing. I need to now call the patient, right? So we're going to just do that next step. We're going to use something called an HTTP request, which you'll actually see. If you go to here, it'll tell you that you need to kind of do this thing called an HTTP request. All that means is that it needs to go to VAPI now. Right. And, and tell Eileen here's who to call. Right. So that's, that's all you're doing in this, in, in this thing. I'm just going to take some of the parameters that it has in here. It kind of tells you exactly what to do in here. I'm just going to cut and paste.
- [41:42]
A
These things in here for, for those listening. We're on. We're in. Claude. Claude. Yeah. OB is building all of this inside of N8N, which essentially is just connecting nodes. So if you're picturing what his screen looks like, it's just a box that says, I'm going to go look at my Google sheets. And then he's connecting that box to another box which is, which he's giving some instructions to, to go essentially go connect that to vapi.
- [42:08]
B
Yep. And Claude's basically kind of guiding me the entire way. Right. It's kind of giving me like the, the instructions that, that, that I would need.
- [42:17]
A
Right. And it could be Claude, it could be Chad, GPT, it could be Copilot, whatever.
- [42:24]
B
Could be, could be anything you like. And that's the flexibility of these tools is they work with pretty much any of these assistants that you might be familiar with, even Grok, you know. So.
- [42:37]
A
Any particular reason you like Claude?
- [42:39]
B
I like Claude for a couple reasons. One, I actually think Claude is like just visually very minimalist. It kind of puts me at ease for some reason. I don't know if it's the color choices or the font choices that it's picked, but it makes me feel more creative than ChatGPT or others. But the other thing that I found in practice, like, practical, is that I believe it's a better writer than ChatGPT. Like, when I'm. One of the things I do a lot is I use it to brainstorm on content that I want to write or documents that I'm creating. And I feel like the writing is more human than ChatGPT and others. But these models are getting better every day. Like literally, you know, you can kind of put them head to head now, and it's, it's hard to tell the difference. The other thing I like about Claude is for these engineering tasks, like kind of where you're writing code, you know. So, for example, here there's some code that's being written here. And this looks complicated. It's actually not all that complicated. All you're doing here is telling Vapi, like, I need to talk to the Eileen assistant and I need to give it the information from the sheet, right? To be like, hey, here's the phone number to call, here's the patient's name, here's the appointment time. That's what this code is doing. And to write and generate this code, in my experience, Claude does a better job than the other tools, like a writing code. So that's all I'm doing in here now is I've cut and pasted kind of that, a directive, and it's in this format called JSON, which you don't really necessarily need to pay too much attention to, but it's code that you just cut and paste in here. And all it's telling Vapi is here's who to call, here's the patient's name, here's the appointment day, here's their phone number. Right. So I go back to the canvas now. So you're going to make that phone call. You're probably going to have to wait, right, for kind of that call to complete. So you can actually put a little wait thing in here and have it like, wait for a period of time to kind of complete that. Complete that phone call.
- [44:49]
A
Right.
- [44:50]
B
And you can give it, you know, like. Like a time period here to like, I kind of wait for a response, you know, so, because remember, Eileen has to pick up the phone and make the call and, you know, do all that. So. So you can, you can, you can have it. You can have it do that. So after you kind of wait, you're going to want to do what. You're probably going to want to get the calls result. So what happened on the phone call, you know, is OB actually going to show up to this thing or not? Right. So you're going to do another HTTP request here and you're going to bring in that particular call that was made.
- [45:29]
A
Right?
- [45:30]
B
So there's a call that kind of is made as an ID for that call. It automatically kind of brings that stuff in here and you're going to bring in the. You're going to bring in the details of that call into here. That's all I'm doing here is just saying bring in those details of that call. And then last but not least, you know, we wanted to update the spreadsheet and mark it as called, right? So you're going to just append or update a row and you're going to tell it to just use that document that we made and you're going to tell it to update that sheet and you're going to tell it to update the particular column, you know, that you're going to want to update. So, yeah, so that's high level, just kind of like really simple workflow. Right. Where it's going to go through and make. Make those. Make. Make that update. Right.
- [46:24]
A
So what we did basically, again, if you're just following along on the audio, what we did is we basically have a Google Sheets document where Obie put patient information. And then we said, hey, go pull this patient information. Hey, N8N the brain that that OB described, go ahead and pull this patient information from this Google Sheet. And then next step, now that you have that information, now feed it over to vapi. And VAPI has been trained as to what to do with that information with that prompt that we designed earlier. And then VAPI calls that patient, and then the call is pulled back into N8N to say, this is what happened. And then based off of that, the Google sheet is updated.
- [47:12]
B
Yep. And it's a. It's a very simple workflow. But I think what you're. What you're trying to do here is again, if your staff is making these phone calls, and maybe it takes them an hour a day to do that or whatnot. Right. You're just. You're just trying to take that work and offload it. Right. Give them like a way to just spend their time differently. So. And yeah, it's not perfect, obviously not integrated with the EMR and all of that, but you're able to at least offload.
- [47:44]
A
Right.
- [47:45]
B
Some of that work, which is useful. So that's a quick demo that.
- [47:51]
A
Should we get that thing to call us real quick?
- [47:53]
B
I would love it to call. Yeah, I don't know if maybe we can call you this time. I'm gonna put you in here. And I'm assuming, Eugene, that everybody has your phone number, but you could obviously. You could obviously blur this out if you don't want people to have your phone number here. I got it. That's what I got in there. Cool. So I'm going to have it give you a phone call, see if it works.
- [48:23]
A
Yeah, let's do it. Oh, I have spam blocker on. I wonder if it'll work.
- [48:29]
B
Oh, you'll have to make sure that. Make sure your do not disturb is off.
- [48:33]
A
Yeah, let's see what happens. Hello, I'm async with Eugene speaking. Hi, Eugene, this is Vision Care Optometry calling. I'm reminding you about your eye exam appointment tomorrow at 2pm can you confirm you'll be able to make it. Oh, I thought it was at 2:30. I have it scheduled for 2pm tomorrow. If you need to reschedule, please give us a call. No, 2pm is okay, I think. Perfect. We'll see you tomorrow. If anything changes, please give us a call at 555-0100. Have a great day, Eugene.
- [49:21]
B
Bye.
- [49:23]
A
She.
- [49:24]
B
She, she got, I think she was a little turned off by your spam agent because she was not friendly, you know, when she got on there. Yeah, that's funny.
- [49:34]
A
That's okay. That's. But that's, that's where the iteration comes in. This is the, this is the, the cool part of this, right? Like we just built something in a matter of minutes that's generally practical and at the same time it's going to have bugs. And I encourage people to think about this as like, the cool thing about AI is that it's not just software that's off the shelf. It's like a, it's like an employee that you can continue to train and you can say aha. In that particular case, if that happens, let's, you know, let's write a rule for that. Let's tell it what to do in that particular case.
- [50:09]
B
100%, I think. Think about it like your first day. This is like Eileen's first day on the job. She spent maybe 30, 45 minutes on the job. That would have been okay for a first call. I wouldn't have like, you know, fired her after that. Right. So, yeah, that's it.
- [50:26]
A
And then you tell Eileen how she can be better.
- [50:29]
B
That's right. Give her the feedback. Nice thing about VAPI is, by the way, is like, all the calls are like stored in here. You can kind of see the trans scripts. You can see all the logs. It can tell you exactly what, like you can see right here what, what you said. Right. You know, that was a 230. And this allows you to evaluate the AI and kind of, you know, keep kind of making it better. So if this is thing, this thing's out in the wild making calls that there's an importance of evaluation and iteration. Right? That you're evaluating what the AI is doing, you're figuring out how you can make it better. And you, you kind of iterate. And again, you can always use your friend to help you iterate. You could literally, which I love to do, export all these transcripts, feed them to Claude and be like, claude, improve the system prompt. You don't even have to get in there and tell it exactly what to do? You can say, give me like exactly what I need to do to improve these calls.
- [51:24]
A
I would start with a what are the issues that you're seeing, Claude? And then make sure we're on the same page and then say, okay, if you had to correct these issues, how would you do it? Okay, great. Now if I agree with that workflow or if I agree with that recommendation, like, okay, go improve the system prompt. Right?
- [51:41]
B
Totally. Absolutely. Yep. And you can again use the act as there be like, hey, act as an expert AIML developer building voice agents and evaluating the performance of voice agents. You could even give it a, a role of like quality assurance. You know, like you're an expert QA for voice agents. You know, troubleshoot this, this based on all these, these transcripts. So yeah, very powerful.
- [52:07]
A
So I think that's awesome, Obi. I think that you're, that what we just did was demonstrate the ability to use a tool. Use a, use another tool. So we use Google Sheets simple tool that be a little bit more complicated and probably less known tool. And then we used N8N to build a brain to connect the two together and give the, give the AI an opportunity to autonomously do a task so it becomes an agent, not just, not just a tool or a prompt machine type of thing. So that's great. I think that that's a really good example. So I'm wondering if we realistically have time to build a second thing or if we should do that in a follow up episode.
- [52:53]
B
Let's, let's do a follow up because I like, I would love to kind of see the feedback. I mean, I mean, dude, you know your audience better than, than other folks. Like how do you feel like people would resonate with that? I'm just curious.
- [53:05]
A
Yeah, I, I mean I, I, I'm gonna ask the audience then right off the rip and say, hey, you listened to us build something or you watched us do it on YouTube. Either way, if you want to watch us build something. So this particular thing that we just, we just built, you could just take it off the shelf and you could play with it or you could at least get a most important thing is that you get the understanding that AI can have practical application. There are people working on things specifically, but you also don't have to sit idle to the side that there are things that you can do and play with these things. And especially, you know, I like Obi's point that you've got a buddy the entire time. Anything you want to accomplish, just ask your buddy and Whether that's Gemini, ChatGPT, or Claude, it's right there. So let's think about, you know, even if you're thinking, how can I have a really practical, meaningful application in my practice, start having that conversation with your buddy and then see if. We'll see what you can figure out.
- [54:09]
B
Totally. No, I like. I like that. And, you know, we could also take requests from the audience, like, you know, tag in, like, what, you know, you would like to build next. But the. The roleplay agent is actually much faster to build. It probably won't take that long. So if you wanted to record it, we could. Yeah.
- [54:26]
A
So, yeah, why don't we do that next? We'll just plan the next episode, and in that episode, we'll take the top request or the top two requests if we have time for it, and we'll just build stuff live. It'll be. It'll be really fun, and we'll do it in a few weeks based off of how I'm sure the audience will love it. But I also. This topic is so near and dear to my heart. My team is working tirelessly to make sure that there are tools available for optometrists and also thinking through a lot of these roadblocks and a lot of these guardrails and a lot of the challenges with the EHR integrations. So this is, like, stuff I absolutely love to geek out over. So. You're absolutely right. Let's get the audience to give us some feedback, give us some requests, and we'll build something live on air next time.
- [55:14]
B
Yeah, let's do it, man. No, it'll be fun. And I think there's. At least in this space, nobody's really doing anything like this, where they're actually showing, like, they're just talking about it. So I think. I think it could be a real differentiator if we just get on and just build. Just kind of show stuff, build in public, you know, and, like, I think it'll get people really excited about what's possible, not. Not just talking about it.
- [55:35]
A
I think that's great. All right, Obi, thank you so much for joining. I'm sure that we're going to. We're going to get some requests. Maybe not the ones we necessarily expected, but we'll get some requests from the audience to do the next one. Thanks for being on the Power Hour.
- [55:49]
B
Thanks, Eugene. Appreciate it. Talk soon.