WEBVTT 00:00:00.000 --> 00:00:03.020 I want you to just imagine a sound for a second. 00:00:03.919 --> 00:00:07.500 It's Tuesday morning, 10 a .m. You're in a dental 00:00:07.500 --> 00:00:10.119 clinic. The receptionist is busy. The dentist 00:00:10.119 --> 00:00:13.019 is working. And in the background, a phone is 00:00:13.019 --> 00:00:17.280 ringing. It rings and rings and then it stops. 00:00:17.820 --> 00:00:20.160 Silence. Just silence. And to the average person, 00:00:20.199 --> 00:00:22.879 that silence is, you know, just a missed call. 00:00:23.059 --> 00:00:25.140 Right. It's an annoyance, but it's part of doing 00:00:25.140 --> 00:00:27.059 business. But we're looking at a case study today. 00:00:28.010 --> 00:00:31.070 of a dr martinez he's a solo dental practitioner 00:00:31.070 --> 00:00:34.250 and for him that silence was the actual sound 00:00:34.250 --> 00:00:37.250 of money evaporating okay his clinic was missing 00:00:37.250 --> 00:00:40.789 34 of its calls during the day and what 100 of 00:00:40.789 --> 00:00:43.189 calls after hours i'm guessing 100 of course 00:00:43.189 --> 00:00:45.009 but they didn't audit they put in the solution 00:00:45.009 --> 00:00:47.390 we're going to discuss today and in six months 00:00:47.390 --> 00:00:52.189 they recovered 1 .3 million dollars 1 .3 million 00:00:52.189 --> 00:00:54.829 all from calls that were just going to voicemail 00:00:54.829 --> 00:00:57.420 before That is a staggering number. I mean, it 00:00:57.420 --> 00:00:59.500 completely reframes the problem. We're not talking 00:00:59.500 --> 00:01:01.140 about better customer service. We're talking 00:01:01.140 --> 00:01:04.060 about plugging a massive hole in the bottom of 00:01:04.060 --> 00:01:06.900 the boat. And the plug wasn't a second shift 00:01:06.900 --> 00:01:10.219 of employees. It was an AI voice agent. Right. 00:01:10.340 --> 00:01:12.159 And I know what everyone's thinking. Oh, great. 00:01:12.280 --> 00:01:15.099 Another robotic phone tree. Press one for appointments. 00:01:15.359 --> 00:01:18.299 That's not what this is. Not at all. The source 00:01:18.299 --> 00:01:21.260 material we're breaking down is a guide. And 00:01:21.260 --> 00:01:25.099 the promise here is, it's pretty wild. It claims 00:01:25.099 --> 00:01:29.260 you can build a full digital employee, one that 00:01:29.260 --> 00:01:33.140 thinks and speaks and even negotiates, in under 00:01:33.140 --> 00:01:36.439 18 minutes. 18 minutes. For something that agencies 00:01:36.439 --> 00:01:38.579 are currently charging small businesses, what, 00:01:39.040 --> 00:01:41.939 $5 ,000 to $25 ,000 to set up? That's the opportunity 00:01:41.939 --> 00:01:44.519 right there. So here's our plan for this deep 00:01:44.519 --> 00:01:46.620 dive. We're going to strip away the hype and 00:01:46.620 --> 00:01:49.000 really look at the engineering. First, we need 00:01:49.000 --> 00:01:51.219 to define what actually makes an agent different 00:01:51.219 --> 00:01:53.359 from, say, a chatbot. Then we'll dissect the 00:01:53.359 --> 00:01:56.200 stack. We're talking retail, AI, cal .com, and 00:01:56.200 --> 00:01:58.980 Gemini. And then walk through the build process, 00:01:59.239 --> 00:02:02.480 the brain, the memory, and the tools. And I want 00:02:02.480 --> 00:02:04.939 to be clear for everyone listening. We are ignoring 00:02:04.939 --> 00:02:07.400 the spaghetti code. The complex flowcharts. This 00:02:07.400 --> 00:02:10.340 is the zero code approach. The guide claims if 00:02:10.340 --> 00:02:12.500 you can fill out a web form, you can build this. 00:02:12.659 --> 00:02:15.259 Which makes it accessible. Exactly. Okay, let's 00:02:15.259 --> 00:02:17.599 start with the philosophy. The guide uses a phrase 00:02:17.599 --> 00:02:20.840 that stuck with me. Scalable labor. We usually 00:02:20.840 --> 00:02:24.180 hear automation. Why is scalable labor a better 00:02:24.180 --> 00:02:26.800 way to think about this? Well, think about a 00:02:26.800 --> 00:02:30.240 human receptionist. Let's call her Sarah. Sarah's 00:02:30.240 --> 00:02:34.219 great, but she's linear. She can handle one call 00:02:34.219 --> 00:02:36.599 at a time. Right. If three people call at once, 00:02:36.680 --> 00:02:39.539 two go to voicemail. Which, as you said, is basically 00:02:39.539 --> 00:02:41.979 a black hole where leads go to die. Yeah. An 00:02:41.979 --> 00:02:44.740 AI agent is parallel. If 100 people call at the 00:02:44.740 --> 00:02:48.520 exact same second, the AI just. It spawns 100 00:02:48.520 --> 00:02:51.419 instances of itself. It answers every single 00:02:51.419 --> 00:02:53.419 call. Every single one. It doesn't sleep. It 00:02:53.419 --> 00:02:54.900 doesn't get sick. It doesn't get overwhelmed. 00:02:55.539 --> 00:02:57.840 That's a kind of scalability that biological 00:02:57.840 --> 00:03:00.340 labor just can't compete with. It fundamentally 00:03:00.340 --> 00:03:02.419 changes the economics of the front office. But 00:03:02.419 --> 00:03:04.840 I have to push back a little. We've had automated 00:03:04.840 --> 00:03:07.620 systems for years. I call my bank. I talk to 00:03:07.620 --> 00:03:09.580 a robot. I get frustrated. And I just start yelling, 00:03:09.719 --> 00:03:11.960 representative, how is this really different? 00:03:12.240 --> 00:03:15.139 That is a critical question. That old system 00:03:15.139 --> 00:03:19.150 is an IVR. An interactive voice response. It's 00:03:19.150 --> 00:03:21.430 just a decision tree. A script. A rigid script. 00:03:21.650 --> 00:03:24.969 If you say balance, it plays recording A. If 00:03:24.969 --> 00:03:27.550 you say transfer, it plays recording B. It's 00:03:27.550 --> 00:03:32.069 a maze. An AI agent is a brain. It's got a large 00:03:32.069 --> 00:03:34.810 language model and LLM at its core. It's not 00:03:34.810 --> 00:03:36.650 following a script. It is understanding your 00:03:36.650 --> 00:03:39.729 intent. So if I just put a talking chat bot on 00:03:39.729 --> 00:03:42.650 a phone line, is that an agent? No. Without tools, 00:03:42.830 --> 00:03:44.849 it's just a brain in a box. Agents have to actually 00:03:44.849 --> 00:03:46.930 do work. Okay, let's unpack that. The source 00:03:46.930 --> 00:03:49.590 lays out a formula. Brain plus memory plus tools 00:03:49.590 --> 00:03:52.050 equals agent. That's the holy trinity right there. 00:03:52.110 --> 00:03:54.349 The brain is the reasoning engine. In our case, 00:03:54.449 --> 00:03:56.349 that's Google's Gemini model. And the memory? 00:03:56.569 --> 00:03:58.569 Memory is just context. It's knowing that you 00:03:58.569 --> 00:04:00.129 just told me your name is Mike so I don't ask 00:04:00.129 --> 00:04:02.110 for it again five seconds later. Yeah. You know, 00:04:02.169 --> 00:04:04.729 it avoids that digital amnesia. So the tools. 00:04:05.169 --> 00:04:08.300 The tools are the critical piece. The tools are 00:04:08.300 --> 00:04:10.719 the hands. It's the API connection that lets 00:04:10.719 --> 00:04:12.919 the brain reach out into the real world and do 00:04:12.919 --> 00:04:14.280 something. Like this check at Google Calendar? 00:04:14.439 --> 00:04:16.879 Exactly. Check the calendar, see an open slot, 00:04:17.079 --> 00:04:19.560 and then write Mike 2 p .m. into the database. 00:04:19.959 --> 00:04:22.439 Without the tool, it's just a nice chat. With 00:04:22.439 --> 00:04:25.139 the tool, it's a completed task. So the tool 00:04:25.139 --> 00:04:27.379 is really the bridge between digital thought 00:04:27.379 --> 00:04:30.899 and physical reality. That's a great way to put 00:04:30.899 --> 00:04:33.319 it. Tools turn conversation into completed tasks, 00:04:33.579 --> 00:04:36.000 like booking a seat. so let's get into the build 00:04:36.000 --> 00:04:40.439 itself we're using retail ai as the uh the platform 00:04:40.439 --> 00:04:44.220 or orchestrator and gemini 2 .5 flashlight as 00:04:44.220 --> 00:04:46.420 the brain and that choice of flashlight is really 00:04:46.420 --> 00:04:48.480 important why light wouldn't we want the biggest 00:04:48.480 --> 00:04:50.860 smartest brain possible if you were writing an 00:04:50.860 --> 00:04:55.379 email yes but for voice no in voice your number 00:04:55.379 --> 00:04:58.800 one enemy is latency. The pause. The awkward 00:04:58.800 --> 00:05:02.180 pause. Right. If you say hello and the AI takes 00:05:02.180 --> 00:05:04.560 three seconds to think and say hi, the illusion 00:05:04.560 --> 00:05:07.040 just shatters. You think the call dropped. Or 00:05:07.040 --> 00:05:10.720 you talk over it. Gemini Flashlight is incredibly 00:05:10.720 --> 00:05:13.680 fast and incredibly cheap. We're talking milliseconds 00:05:13.680 --> 00:05:16.040 of delay. It feels like a natural conversation 00:05:16.040 --> 00:05:18.519 rhythm. So speed is more important than pure 00:05:18.519 --> 00:05:21.800 IQ in this specific context. 100%. The AI doesn't 00:05:21.800 --> 00:05:23.959 need to write a philosophy paper. It just needs 00:05:23.959 --> 00:05:26.829 to check a calendar and be polite. Speed wins. 00:05:27.110 --> 00:05:30.110 So step one in the guide is just setting up the 00:05:30.110 --> 00:05:32.910 agent and retail standard stuff. But step two 00:05:32.910 --> 00:05:35.709 is the knowledge base. This is where it gets 00:05:35.709 --> 00:05:37.829 interesting for me. Yeah. This is basically how 00:05:37.829 --> 00:05:39.870 you teach it about your business. Think of it 00:05:39.870 --> 00:05:43.110 like giving a new employee. an employee handbook 00:05:43.110 --> 00:05:45.569 you're not retraining a whole model no god no 00:05:45.569 --> 00:05:47.470 you don't send it back to college you just hand 00:05:47.470 --> 00:05:50.290 them a pdf with your hours your prices your cancellation 00:05:50.290 --> 00:05:53.230 policy so you literally just upload a pdf or 00:05:53.230 --> 00:05:55.889 a text file that's it a pdf a word doc whatever 00:05:55.889 --> 00:05:58.610 when a customer asks a question the brain doesn't 00:05:58.610 --> 00:06:01.490 guess it reads that document in a split second 00:06:01.490 --> 00:06:03.930 to find the correct answer so why is a simple 00:06:03.930 --> 00:06:07.930 static pdf better than say fine -tuning a model 00:06:07.930 --> 00:06:11.980 from scratch Simplicity. The AI references the 00:06:11.980 --> 00:06:14.740 document dynamically. It makes sure the answers 00:06:14.740 --> 00:06:16.920 are accurate and always up to date. It's the 00:06:16.920 --> 00:06:19.319 difference between memorizing the textbook and 00:06:19.319 --> 00:06:22.379 just keeping the textbook open on your desk. 00:06:22.459 --> 00:06:24.500 That is a perfect analogy. Okay, so we have a 00:06:24.500 --> 00:06:26.860 brain. We have the handbook. Now we hit step 00:06:26.860 --> 00:06:29.779 three. And I have to admit, whenever I hear the 00:06:29.779 --> 00:06:32.689 phrase API key, I usually... I freeze up a little. 00:06:32.790 --> 00:06:35.750 You are not alone. The API key is usually that 00:06:35.750 --> 00:06:38.550 scary monster under the bed for people who don't 00:06:38.550 --> 00:06:40.870 code. It sounds like the part where everything 00:06:40.870 --> 00:06:43.610 is supposed to break. It does. But trust me, 00:06:43.649 --> 00:06:46.170 this is the boring part that pays off. It's much 00:06:46.170 --> 00:06:48.129 simpler than it sounds. All right, walk us through 00:06:48.129 --> 00:06:50.389 it, the impossible part, integrating the calendar. 00:06:50.629 --> 00:06:53.290 So we're using cal .com. It's free. It's powerful. 00:06:54.170 --> 00:06:56.990 First, you create an event type. Let's call it 00:06:56.990 --> 00:06:59.269 a 60 -minute meeting. Okay. You grab the event 00:06:59.269 --> 00:07:01.829 type ID right out of the URL. You just copy it. 00:07:01.850 --> 00:07:03.769 Copy, paste. Got it. Then you go into settings. 00:07:03.829 --> 00:07:06.629 You generate an API key. Quick tip, make sure 00:07:06.629 --> 00:07:09.589 you set it to never expire so your AI receptionist 00:07:09.589 --> 00:07:11.750 doesn't quit in a month. Right. And you paste 00:07:11.750 --> 00:07:14.509 both of those into retail under functions. That's 00:07:14.509 --> 00:07:16.750 it. That's it. You've given the AI permission 00:07:16.750 --> 00:07:20.029 to see your schedule and, crucially, to write 00:07:20.029 --> 00:07:23.069 to it. So the AI actually sees my real Google 00:07:23.069 --> 00:07:27.100 calendar. Yes. And it syncs instantly. Right. 00:07:27.160 --> 00:07:28.920 It prevents double bookings without a human ever 00:07:28.920 --> 00:07:32.060 having to look at it. That is, that's profound. 00:07:32.139 --> 00:07:34.279 It's not a script anymore. It's a dynamic system. 00:07:34.439 --> 00:07:37.379 It's alive in a way. It's reacting to the real 00:07:37.379 --> 00:07:39.920 world in real time. So we have the brain, we 00:07:39.920 --> 00:07:42.800 have the tools, but we still have a robot. We 00:07:42.800 --> 00:07:45.060 need a personality. Right. And that's steps four 00:07:45.060 --> 00:07:48.240 through seven. This is the fun part. Prompting, 00:07:48.240 --> 00:07:52.319 voices, and very importantly, handoffs. What's 00:07:52.319 --> 00:07:54.939 the handoff? The human handoff is your safety 00:07:54.939 --> 00:07:57.800 valve. Sometimes the AI will get stuck or a caller 00:07:57.800 --> 00:07:59.879 is just angry and wants a person. Let me talk 00:07:59.879 --> 00:08:01.959 to your manager. Exactly. So you set up a call 00:08:01.959 --> 00:08:05.019 transfer function that just routes the call to 00:08:05.019 --> 00:08:07.360 a real human's phone number. Okay, that's critical. 00:08:07.540 --> 00:08:09.860 And the prompt? The prompt is the secret weapon. 00:08:10.980 --> 00:08:14.139 And the source has great advice here. Don't write 00:08:14.139 --> 00:08:16.269 it yourself from scratch. Yeah. Use a prompt 00:08:16.269 --> 00:08:18.430 generator. You define the persona. You were a 00:08:18.430 --> 00:08:20.889 helpful receptionist named Henry. You set the 00:08:20.889 --> 00:08:23.569 guardrails and you let the generator create the 00:08:23.569 --> 00:08:26.009 structured instructions the LLM actually needs 00:08:26.009 --> 00:08:28.089 to follow. And the voice itself. We're using 00:08:28.089 --> 00:08:30.370 11 Labs voices here. But the guide had this one 00:08:30.370 --> 00:08:32.769 tip that I just loved. Adding background noise. 00:08:32.990 --> 00:08:35.490 Yes. Light coffee shop background noise. Why 00:08:35.490 --> 00:08:37.009 would you add background noise? Doesn't that 00:08:37.009 --> 00:08:39.350 just make it harder to hear? It creates psychological 00:08:39.350 --> 00:08:42.269 realism. Yeah. Total dead silence feels digital. 00:08:42.750 --> 00:08:45.519 Ambient noise feels human. It's a texture. It 00:08:45.519 --> 00:08:48.080 creates a sense of presence. That's it. And then 00:08:48.080 --> 00:08:49.940 you just set the welcome message to dynamic. 00:08:50.240 --> 00:08:53.179 So the AI generates its own greeting based on 00:08:53.179 --> 00:08:56.039 the prompt. Hi, I'm Henry from AI Firestore or 00:08:56.039 --> 00:08:58.679 whatever it is. OK, so we've built it. And now 00:08:58.679 --> 00:09:02.340 the moment of wonder. I'm looking at this transcript 00:09:02.340 --> 00:09:05.379 of a demo call from the guide for a made up Sarah's 00:09:05.379 --> 00:09:07.820 salon. This is the aha moment. It really is. 00:09:07.960 --> 00:09:11.139 A caller asks for a cut and blow dry. The agent 00:09:11.139 --> 00:09:15.220 quotes the prices. 95 for senior stylist. 70 00:09:15.220 --> 00:09:18.039 for a junior. The caller then asks for tomorrow 00:09:18.039 --> 00:09:21.980 at 11 a .m. or 8 a .m. And this is where a dumb 00:09:21.980 --> 00:09:25.519 chatbot fails. If 11 and 8 are taken, it just 00:09:25.519 --> 00:09:29.399 says, sorry, unavailable. End of story. Right. 00:09:29.480 --> 00:09:31.840 But this agent says, I have an opening at 9 p 00:09:31.840 --> 00:09:34.340 .m. Would that work? Boom. The caller agrees, 00:09:34.679 --> 00:09:36.559 gives their name and number, and the booking 00:09:36.559 --> 00:09:38.779 is confirmed. It didn't just check a box. It 00:09:38.779 --> 00:09:41.000 negotiated. It negotiated the time. It didn't 00:09:41.000 --> 00:09:43.779 just say no. What does that imply? It implies 00:09:43.779 --> 00:09:47.220 reasoning. It understood the intent was I want 00:09:47.220 --> 00:09:49.080 an appointment. Yeah. And it offered a viable 00:09:49.080 --> 00:09:51.340 alternative. And to make this go live, to actually 00:09:51.340 --> 00:09:53.620 deploy it. You buy a phone number in retail for 00:09:53.620 --> 00:09:56.769 like two bucks. You click publish. And you can 00:09:56.769 --> 00:09:58.590 call that number from your cell phone five seconds 00:09:58.590 --> 00:10:01.409 later. It's live. Okay, so let's recap the big 00:10:01.409 --> 00:10:04.909 idea here. We moved from this sphere of spaghetti 00:10:04.909 --> 00:10:08.190 diagrams and code to a working system. We did. 00:10:08.230 --> 00:10:11.750 We used Retail AI for the framework, the body. 00:10:11.929 --> 00:10:15.330 We used Cal .com for the action, for the tools. 00:10:15.509 --> 00:10:18.009 And Gemini 2 .5 Flashlight for the reasoning, 00:10:18.070 --> 00:10:20.850 for the brain. And we did it all without writing 00:10:20.850 --> 00:10:23.259 a single line of code. You just built a system 00:10:23.259 --> 00:10:25.759 that, as you said, agencies are charging $10 00:10:25.759 --> 00:10:28.720 ,000 or more for. In the time it takes to watch 00:10:28.720 --> 00:10:31.159 a sitcom. It's the democratization of some really 00:10:31.159 --> 00:10:34.259 high end technology. It's power. It's putting 00:10:34.259 --> 00:10:36.960 enterprise grade automation into the hands of 00:10:36.960 --> 00:10:40.759 a solo dentist or a local salon owner. So what 00:10:40.759 --> 00:10:42.820 do you think this all means for the future of 00:10:42.820 --> 00:10:44.919 work? This feels like it's about more than just 00:10:44.919 --> 00:10:47.620 automation. It's about recovering lost time and 00:10:47.620 --> 00:10:50.899 lost revenue. It's about this idea of the 18 00:10:50.899 --> 00:10:53.940 minute employee. The 18 -minute employee. Yeah. 00:10:54.100 --> 00:10:57.240 If you can build an employee in 18 minutes that 00:10:57.240 --> 00:11:01.519 recovers $1 .3 million, the definition of hiring 00:11:01.519 --> 00:11:04.120 has just fundamentally changed. That is a very 00:11:04.120 --> 00:11:05.740 heavy thought. So I want to encourage everyone 00:11:05.740 --> 00:11:08.559 listening. Try the 18 -minute challenge. Even 00:11:08.559 --> 00:11:10.279 if you don't have a business. Just build it to 00:11:10.279 --> 00:11:13.039 see it work. Exactly. Build an agent that books 00:11:13.039 --> 00:11:15.879 time on your personal calendar just to see how 00:11:15.879 --> 00:11:19.299 that brain plus memory plus tools formula feels. 00:11:19.720 --> 00:11:21.720 Because once you see it work, you can't unsee 00:11:21.720 --> 00:11:24.120 it. Go build something. Go build something. Thank 00:11:24.120 --> 00:11:25.980 you for listening to The Deep Dive. We'll see 00:11:25.980 --> 00:11:26.419 you next time.