So imagine hiring a brand new, just incredibly eager sales rep. Oh, the dream employee. Right. Exactly. Like they work 24/7, they never take a break, and they reply to every single customer inquiry in seconds. Sounds amazing. Until the bill comes. Yes, so the end of the month rolls around. And you realize they spent, I don't know, 45 minutes talking to a wrong number. Or they got stuck in an infinite conversational loop with a spam bot. Which happens all the time. All the time. And suddenly you've racked up a $10,000 phone bill because in the world of uh AI voice agents and automated chatbots, that is not a hypothetical nightmare. No, if you don't know what you're doing, that is just a regularly Tuesday. Exactly. So, welcome back to another deep dive. We are super thrilled to have you here today because before we even start, you know, pulling apart today's material, I have something really highly actionable waiting for you. Something you'll definitely want to use for this one. Oh, for sure. If you check the show notes right below this, you're going to find a link for a completely free 30-day Go High Level trial. Which is huge. Yeah, that is double the standard trial length. So you have a full 30 days to get into the software, test the waters and actually apply the precise agency level strategies we are breaking down today. So make sure you click that link. You're really going to need it to follow along with the dashboard stuff we're talking about. Right. So let's get into the mechanics of this. Our mission today is to completely deconstruct High Level's internal documentation on their AI reporting and pricing structures. We are looking at how digital marketing agency owners can, you know, stop treating AI like a shiny toy. And start controlling it like a business, like a heavily optimized, highly profitable unit. Exactly. Because this shift from just a casual user to a real orchestrator, that seems to be the biggest hurdle right now, right? It absolutely is. I mean, look, when you hire human employees, the financial modeling is completely predictable. Right, you know what you're paying them. Exactly. You negotiate a salary, you sign a contract, and it just sits there as a fixed number on your spreadsheet. But deploying generative AI, that shatters that predictability entirely. Because it's all variable. Right, you are entering this wild landscape of paper-use APIs, fluctuating token costs, these compounding microtransactions. If an agency owner is scaling up their AI deployments without a granular understanding of that data, they are flying blind. Flying blind into a total margin killer. Exactly. So to stop flying blind, you need a command center. And the documentation points immediately to the AI usage dashboard. But, uh, there's a really critical distinction here. Yeah, the access level. Right. This dashboard sits strictly at the agency level. It is built exclusively for agency admins, meaning your clients, the sub account users, they cannot see this overarching view. And that's a very deliberate architectural choice by High Level. Because you need everything centralized, right? Like client status, token usage, transaction counts. Yeah, exactly. As an agency owner, you cannot manage your infrastructure by logging into 50 different sub accounts and like manually checking individual bot logs. That would be a nightmare. It's impossible. You have to see the entire macro environment first just to track where the capital is flowing. You're looking at this time series chart that lets you toggle between aggregate costs, sheer token volume and specific event triggers. Right. But let's look at the mechanism here, though. Having a macro view is sort of like having a smart electricity meter for a massive 100-unit apartment building. Okay, I like that analogy. Right. Like you look at the dial, you see a huge spike in power usage on a Thursday, and you know something happened. But the smart meter doesn't tell you why it spiked. No, it just shows the surge. Yeah, it doesn't tell you if someone uh left a space heater on or if a crypto mining rig just booted up in apartment 4B. So a spike in token cost on a dashboard doesn't inherently give an agency owner actionable intelligence, does it? Well, it doesn't unless you scroll past aggregate charts and actually interact with the transactions table. Oh, yeah. That is the forensic layer. The aggregate charts are really just the warning system. The transactions table is where you diagnose the anomaly. So you see the spike, and then you dig in. Exactly. When you spot that spike in costs, you drill down into a single row on that table. You are looking at the exact timestamp, the specific product that triggered the cost, whether that's conversations AI or agent studio, the billing mode, and the precise token breakdown of one isolated event. I really want to pause on that token breakdown actually. Because this is where bots usually go rogue. When a chatbot burns through a massive amount of tokens unexpectedly, it's rarely because it had a thousand normal conversations. No, it's almost always a context window issue. Right. The bot gets confused, starts looping, and every time it replies, it has to re-read the entire previous conversation history. Yep, compounding the token count exponentially with every single back and forth. It's terrifying how fast that burns money. It is, but that is the exact mechanism of a runaway cost. In a large language model, you know, the bot doesn't possess inherent memory. It has to feed the entire transcript back into itself to understand the context of the next reply. So if a user sends like 50 short, confused messages. The context window inflates massively. But the transactions table exposes this immediately. You can see clear as day that a specific conversation burned 80,000 tokens in five minutes. Wow. And then you can actually do something about it. Right. You isolate the sub account, identify the rogue bot, and intervene before it eats the client's entire monthly retainer. And I guess it also completely removes the ambiguity from client billing, too. Oh, totally. You aren't guessing why their invoice is higher. You have a definitive timestamped receipt of the compute power that you utilized. Which is great. But, uh, uncovering that single looping error inside a massive transactions table assumes an agency owner actually has the hours to stare at rows of raw data. Which nobody does. Right. Realistically, on a busy Monday morning, nobody has the bandwidth to manually parse a sprawling analytics dashboard. Which is why the architecture relies on utilizing an LLM to analyze its own outputs. AI reading AI. Exactly. The documentation details this dashboard summary AI feature. It's literally designed to synthesize those raw dashboard numbers and time series charts into plain language executive summaries. So you just what click the summarize dashboard with AI button in the header? Yep, it's right there. To me this acts like having a forensic accountant or like a highly competent chief of staff. Someone who crunches the 50-page analytics report overnight and hands you a three-bullet point memo over a coffee. That's a perfect way to look at it. But I am inherently skeptical of software that claims to, quote, summarize data. Is this just running a basic rule-based script? Like, is it just going to tell me sales are down 10% which I can clearly see with my own eyes? No, it is much deeper than a simple percentage readout. It actually uses anomaly detection algorithms combined with a language model to provide contextual narrative. Okay, so it gives the why. Exactly. It won't just say token usage dropped. It analyzes the applied filters and the specific date ranges to provide causality. It might synthesize that a precipitous drop in usage on Wednesday correlates directly with a specific sub account deactivating their inbound voice AI channel. Oh, wow. Yeah. And then it will recommend whether that is a trend you need to investigate or just an expected baseline shift. And you can actually query it directly. Right. Using natural language. You can. You can type, "What are the key driving trends behind this week's token usage?" And it will literally converse with you based on the live data. But wait, the documentation explicitly notes a major caveat here. Those natural language follow-ups only work on custom widgets you've built yourself, right? Yeah, that's a crucial detail. Like if you were using the standard out-of-the-box dashboard widgets, you cannot ask it conversational follow-up questions. Right, which forces agencies to become proficient in building custom data views tailored to their specific KPIs rather than just relying on the default templates. Makes sense. And speaking of hidden functionalities, I should mention this entire summarization suite is not active by default. Oh, yeah. I saw that in the notes. It's kind of a hidden gem. Very hidden. You have to navigate to agency settings, go to the company tab, and explicitly toggle the AI employee switch to the on-on position. So if you skip that step, the summarization features simply do not exist in your UI. They're just gone. Good to know. So, okay, assuming we have the dashboard configured and the AI summarizing our macro costs, we really need to shift our focus to the actual assets generating these metrics, right? Yeah. The frontline bots talking to the end customers. Right, the employees themselves. That evaluation happens inside the conversation AI agents dashboard. Which is separate. Entirely separate from the financial usage metrics. This is purely about performance and engagement tracking across every single bot in your ecosystem, all consolidated into one view. So you can filter by date range, specific channels, like SMS versus Instagram DMs, or by individual agents. Exactly. Let's talk about the metrics here though. Because counting raw chat volume is kind of a vanity metric, isn't it? Like if you tell a client their bot processed a thousand messages, they might assume they have a thousand warm leads. Which is almost never true. Right. But the dashboard specifically tracks total unique contacts. How does the system actually deduplicate those contacts under the hood? Because it's not just doing simple string matching on a phone number, is it? No, it's way smarter than that. It relies on session windowing and intent recognition. Okay, what does that mean in practice? So, if a customer texts, "Hey," then 30 seconds later texts, "Are you there?" and then texts, "I have a question about pricing." A normal system counts that as three separate interactions. Right. But the AI dashboard utilizes natural language processing to recognize that this is a single, continuous thought process from one unified identity within a specific time frame. Oh, so it compresses that entire flurry of messages into one unique contact interaction. Exactly. It provides a surgically accurate picture of actual human engagement, stripping out all the noise of double texting. That's brilliant. And it also tracks total actions triggered. Which implies the bot is operating autonomously within the CRM, right? It's not just acting as a conversational sounding board. Right. It is executing workflows, updating contact records, transferring conversations based on user intent. Okay, but let me push back hard on some of these performance indicators. Go for it. As an agency owner, I look at the time saved metric on this dashboard and I immediately cringe. Oh, absolutely. That is an incredibly fluffy metric. You cannot deposit time saved into a bank account. If I am charging a client $2,000 a month for an AI infrastructure, they do not care if I saved their receptionist 12 hours. They want to know the hard, undeniable ROI. Which is exactly why you pivot the client's attention entirely away from time saved and anchor it to the hard conversion metrics. Okay. The dashboard meticulously tracks total appointments booked. But hold on. Even a booked appointment can be a negative ROI if the bot is just aggressively jamming completely unqualified leads onto a busy client's calendar. Oh, for sure. Like if a bot books 10 meetings for a high-end consulting firm, and all 10 people have zero budget and no idea what the firm does, the bot just wasted 10 hours of the client's actual time. And that highlights the necessity of strict prerequisite gating, which thankfully the system handles natively. The bot doesn't just hand out a raw calendar link to anyone who says hello. Okay, how does it gate them? When configured correctly within High Level, the bot integrates directly with the calendar ecosystem and initiates a qualification workflow. It asks the specific qualifying questions you've programmed, budget, timeline, specific needs. So it's actively screening them. Yes. Only when the LLM confirms the user's intent matches the qualification parameters does it execute the appointment link shared event. That is huge. And because everything happens within the same closed loop ecosystem, the dashboard cross-references the bots' link shares against actual calendar submissions, giving you an airtight, mathematically proven conversion rate. See, that integration is what turns the bot from a fascinating tech demo into a concrete revenue generator. But, you know, we also have to account for failures because bots hallucinate, they get confused, they hit the limits of their system prompts. They absolutely do. Which brings us to the most vital diagnostic metric on the entire dashboard. The human handover count. The human handover. Right. The system actively monitors how many times the AI reaches a confidence threshold so low that it stops replying and pings a human staff member to intervene. Okay, let's contextualize how an agency actually uses that metric to optimize. Totally. Imagine you built a bot for a local dental clinic, right? Okay, I'm with you. The bot is trained on all the standard clinical terms. Yeah. Cavities, root canals, whitening, but in the real world, a patient texts the clinic and says, "My chopper is absolutely killing me. I need someone to look at it ASAP." Oh, right. The bot's NLP model might not map the colloquialism 'chopper' to 'tooth.' It gets confused, hits a dead end, and triggers a human handover. And historically, without the dashboard, that failure just disappears into the ether. You'd never know it happened. Exactly. But because the dashboard flags that specific handover event, the agency owner can review the exact transcript. You identify the conversational failure point, the word 'chopper,' and you immediately update the bot's custom prompt. So you just add a directive that says like interpret words like chopper, sweet tooth or grill as referring to dental anatomy. Precisely. You deploy the updated model and you've instantly closed a conversion leak. You are engineering continuous optimization based on definitive failure data. I love that. So look, we've covered how to track the overall cost, summarize the analytics and optimize the bots on the front lines. But all of this strategic deployment hinges entirely on understanding how High Level actually charges for these tools. The money part. Yeah, I was digging into the pricing documentation and the financial architecture is incredibly specific. It is, and it requires a really thorough understanding to maintain healthy profit margins. The baseline offering is the AI Employee Unlimited plan. Right, the unlimited plan. It's a flat $97 a month. I view this like the ultimate all-you-can-eat buffet for an agency. It is pretty much. It covers unmetered use of conversation AI, reviews AI, content AI, ask AI, and specifically inbound voice AI. But just like any buffet, there is a rigid fair use policy. Right. You can't walk in, dump the entire tray of prime rib into your backpack and walk out. Exactly. If a sub account starts processing an absurd enterprise-level volume of data that strains the system, High Level will throttle the account or require an upgrade. So, understanding the boundaries of that $97 tier is crucial, especially regarding what is explicitly excluded. Right, what's not in the buffet. Agent Studio, the voice AI widget for websites, and critically, voice AI outbound calling. Those are all paper-use features. If you are building a system that cold calls a thousand prospects a day, that is not covered by the flat fee. Right. So let's break down the math on the paper-use structure, specifically for voice AI. Because when you look at it initially, the pricing feels insanely complicated. It scares a lot of people off. It isn't just a flat 10 cents per minute. The billing is actively split into three distinct moving parts: the voice engine, the text-to-speech provider, and the underlying language model token cost. Why make it so complicated instead of just a flat per-minute rate? See, it looks complicated, but that modular pricing structure is actually the greatest hidden superpower an agency has for scaling profitably. Really? Absolutely. If High Level charged a flat per-minute rate, they would be forced to average out the cost of the most expensive models to protect their own margins. That means an agency would be massively overpaying for simple, low-level automated tasks. By decoupling the architecture into three billing layers, they hand the margin control directly to you. Okay, let's calculate exactly how that works. Layer one is the voice engine itself, just the infrastructure connecting the actual phone call. That's a flat 4 and a half cents per minute. Right. Layer two is the text-to-speech provider, the actual synthetic voice the customer hears. Yeah. If you choose 11 Labs version 2.5, which is highly realistic, that costs 3 and a half cents per minute. So right out of the gate, your baseline audio infrastructure is costing you 8 cents a minute. Exactly. But layer three is the brain. The large language model. You get to choose the brain you slot into this system. And the cost disparity between those brains is immense. I mean, if you want the bot to utilize Open AI's Chat GPT 5, which is a brilliant, heavy, deeply complex model capable of immense reasoning. It costs you $1.25 per million tokens of input and a massive $10 per million tokens of output. $10. But here is what blew my mind in the documentation. You don't have to use Chat GPT 5. No, you don't. If you're building a basic bot for a hair salon whose only job is to answer the phone, check business hours and say, "What time do you want to come in?" That bot does not need the reasoning capabilities of a supercomputer. Exactly. You can completely bypass the heavy models and slot in Google's Gemini 2.0 flashlight. And the token cost for Gemini flashlight drops to what? 7 and a half cents per million tokens. Wow, think about that difference. $10 for output versus 7 and a half cents. You are effectively acting as an arbitrageur of AI compute power. That's exactly what it is. The client pays you the exact same monthly retail retainer for a functioning voice bot. But by intelligently swapping out the underlying LLM based on the complexity of the task, your wholesale token costs plummet and your profit margin widens instantly. That is the definition of managing AI like a business. If a task requires deep, multi-step logical deduction, yeah, you pay for the premium tokens. If a task requires basic intent matching and routing, you use the lightweight models. That's just smart business. It is. Furthermore, the documentation outlines highly valuable free tiers to help you build these systems. Like the funnel and website AI provides a thousand free generative prompts per day per sub account. Which is plenty for most. Oh, yeah. And the workflow AI builder, the system that actually writes the complex automations inside the CRM is entirely free. Wait, really? You just type a natural language prompt like, "Build a workflow that texts the lead immediately, waits two days, and emails them a calendar link." And the AI physically builds the automation logic for free. It physically builds it, entirely free. That's incredible. So let's pull all of this together. We started this deep dive exploring the fear of uncontrolled, runaway AI costs that can completely sink an agency's margins. Right, the $10,000 phone bill. Exactly. But by systematically applying the tools provided, you know, using the usage dashboard for forensic financial tracking, utilizing the summary AI to parse complex trends, tracking hard conversion metrics on the frontline bots, and heavily exploiting the arbitrage opportunities in modular pricing, the fear of AI completely evaporates. Because it is replaced by absolute engineered control. You aren't guessing at your return on investment anymore. You are mathematically constructing it. Which is exactly where you want to be. Yeah. Now, to everyone listening, I want to remind you one final time. Hearing us break down token arbitrage and dashboard summaries is one thing. Actually getting your hands on the data is how you master it. You have to get in there and build. Right. So the link for a completely free 30-day Go High Level trial is sitting right there in the show notes. Remember, that is double the standard trial length. I highly encourage you to click that link, get into the platform, and start actively experimenting with everything we just discussed today. It really is the only way to transition from theory to execution, but, uh, as you start building, I want to leave you with a larger structural question to ponder. Okay, let's hear it. If High Level's AI infrastructure can now analyze its own anomalies, trigger its own qualification workflows, pinpoint its own conversational failures for retraining, and meticulously track its own operational costs, at what point does a digital marketing agency owner fundamentally shift from being a manager of human capital to an orchestrator of autonomous self-correcting systems? Wow, that's a huge shift. It is. And if that is the exact reality we are stepping into, what completely new, unforeseen skills will you need to develop to lead an agency in that landscape? Definitely something to think about as you start your trial. We will see you next time on the deep dive.