VIDEO_ID: Ky2Z6kGp1Vk
URL: https://www.youtube.com/watch?v=Ky2Z6kGp1Vk
TIMESTAMP_FLAGGED: None
LANGUAGE: en
SNIPPET_COUNT: 373
================================================================================
Whether you like it or not, here is the truth about AI voice agents. Whilst everyone else is hyping them up, we've actually sold them for over $12,000 each and have run our voice AI agency for two plus years, building for many different industries. I've seen exactly what works, what breaks, and what no one warns you about until it's too late. So, for this video, I'm walking you through five hard-earned lessons from realworld deployments. The stuff you just won't hear from the basic tutorial videos. Number one, do you have to disclose that it's an AI on the call or can you let it pass as human? Number two, why AI cold calling is banned and how to approach outbound without getting flagged. For number three, why your agents are probably too complex and how to simplify them without killing the functionality. Number four, why fallback systems are really critical and how to actually build them right. And number five, why these tools do not replace humans and what happens when you try to make them. So, if you're building or selling voice agents, this could save you from legal issues, annoyed clients, furious callers, and systems that just crash the moment that they go live. All right, so first up, we have AI disclosure. It's not something that people talk much about in this space, but it is super important. This is where at the beginning of the call, we mentioned to the caller that they are speaking to an AI or some kind of automated system. Now, this is only required if these calls are being made within a few US states of California, Colorado, and Utah. Otherwise, this is not required at the start of the call. You don't have to mention that it's an AI specifically, just that it's some form of automated or virtual call. Additionally, you need to make sure that this first message is a static first message. This means it gets read out every single time for every call. If you don't set this to static and you let the AI decide how to read out the message, you will be introducing the chance for it to hallucinate and forget to read out the disclosures. If this were to happen, you would then be liable for not having that disclosure. Alongside the first message, it's always best to ensure that you've prompted the agent to always mention itself as an AI or automated system. It's definitely possible that people don't hear the first message clearly, and if they ask later, it's best that the agent is clear to the caller about being an AI or automated system. And if you really want to make sure you are compliant at all times, what I would recommend doing is setting up a couple of QA alerts that specifically track if that disclosure was announced on the call or not. This is something that Retail AI actually recently introduced. I'd recommend setting something like that up. Now, my advice from deploying a lot of these AI calling systems is that it's best to be honest, mentioning that it's an AI at the start of the call, regardless of which state the system is taking the calls for. We've honestly seen no negative impact from upfront disclosure. If anything, the results have improved. People just appreciate knowing who they're talking to from the start. Trying to hide that it's an AI definitely damages trust far more than being transparent ever could. As long as the agent can handle that inquiry and help them out as much as possible, then there is no reason for someone to honestly care how that was done. Now, another part to disclosure is mentioning that these calls are being recorded. If you are recording the calls, it is required in a few different US states to mention directly at the start of the call that it is being recorded. You can go ahead and Google this and find out the specific states around it. Missed calls means missed revenue. If you are losing leads because your team can't answer the phone 24/7, I've got something for you. My team at Inflate AI builds inbound and outbound calling systems to automate voice conversations across your business. We're talking 24/7 receptionists, speedtole outbound callers, appointment booking, whatever your business needs. And if you want to see it in action, we've built a demo voice agent on our website that you can test right now. So call it, ask it questions, and just experience how natural these conversations actually are. Once you see what's possible, book an introductory call with us by using the link in the description. All right, for number two, we have AI cold calling. This is pretty much banned in the US. There is specific legislation now from the Telephone Consumer Protection Act, which means that we are not allowed to send any individuals who have not directly consented to receive them any AI or automated calls. Violations can result in fines of $500 to $1,500 per call. We can send calls to businesses, although we must make sure that we are calling the direct landline for the business and not a personal call to someone at the business. Nowadays, most people just use their mobile numbers for everything. So, it's becoming increasingly difficult to make these calls, even with the intention of a business call. Calling a mobile number with AI is still unsolicited and something we cannot do. Now, to be super clear, this doesn't mean we can't send outbound AI calls. It's just the cold part that we can't do. There are certainly some great outbound use cases that are far easier to set up. For example, outbound speed of lead callers, calling leads right after form submissions, client reactivation systems, so calling existing consented customers for repeat services or appointment reminder calls. Obviously, you still need to be very careful when doing these outbound use cases. Consent needs to be clearly captured and stored properly for it to be valid. For example, you need to ensure that you've got the exact date and time that it was collected and the method of consent clearly recorded. When capturing consent, it must be very clear and not hidden away. it should be as close to the registration button as possible. For number three, why creating complex agents is just a really bad idea. So, there definitely does seem to be a tendency to try and make things more complicated than they should be. We try to add new features because they might seem cool, but in production just becomes a nightmare. For example, we've had a few clients reach out to us for AI reception systems, and the booking systems that they've used have had no APIs or very few endpoints to actually build the booking system. Whilst we certainly can try and build out this entire booking system and string together an approach that might work, offering a simpler approach, like texting the user an online booking form might just save you weeks of work. Now, whilst we and the client might prefer to have the booking system over the phone, these AI systems are not perfect. So, if we combine that with a non-perfect API setup, we're going to be putting ourselves into a lot of trouble. We can always push this to be done at a later date in some sort of a phase 2 implementation once we've learned how the users interact with that first version of the system. For us at least, a lot of the potential complexity comes from scoping with clients. They will always want to see what's possible and it is our job to set realistic expectations for this first version. So it doesn't mean trying to build as much as you can. It means building an MVP to a minimal viable product to see the potential and get early results before committing to more complexity. So more complexity can be in the form of more workflows, handling more scenarios, adding new automations, new APIs, etc. These are just new failure points that we then have to manage. Trying to get an AI to correctly capture a complex email address or a specific booking code over the phone line is always difficult. Speech to text models still struggle with distinct spelling. By pivoting to SMS, you aren't just simplifying the build. You're actually improving the user experience as well. And also from a prompting perspective, the more instructions that we are stuffing into the system prompt, the dumber the agent is going to get. An agent with one clear goal will always outperform an agent trying to juggle five different tasks all at the same time. Moving on to number four, critical fallback systems. If there is one thing that I've learned from running our agency, it's this. AI will fail. It's not a matter of if, but when. Maybe the caller has a thick accent, maybe the connection is bad, or maybe the LLM just hallucinates and gets stuck in some kind of loop. If you build your system, assuming the happy path where everything goes perfectly, you are going to have issues the moment you go live. You need to build for that unhappy path. So for us, the first part of a good fallback system is always going to be the live transfer. Every single agent you build needs to have some kind of an escape hatch. If the caller gets frustrated or if the AI detects sentiment turning negative, it needs the ability to transfer the call to a human immediately. We script our agents to recognize phrases like, "Can I speak to a real person?" or "Connect me to support." When the AI hears this, it shouldn't try to argue or deescalate that. It should simply say, you know, absolutely, let me get someone for you and then just transfer that call. So, this safety net gives your clients the confidence to let the AI handle the bulk of the traffic, knowing that if things do go south, which they eventually will, a human is going to be able to get transferred. The second part of this is how you handle function calls and automations. So, a major mistake that we made is trying to get every single automation to run during the call. This could be updating the CRM, sending a confirmation email, notifying the sales team via Slack, all while the customer is still on the line. And this causes a big latency issue. Every time the AI triggers a tool, there is a delay. The AI says, you know, one moment while I send that email, it'll then go silent for 5 seconds while that API is processing and the caller will just think the call is dead. They're going to then, you know, say hello, which will interrupt the AI, breaks the function call, and just ruins the experience. Instead, we need to rely on the end of call automations. If your agent's only job during the call is to collect the data. Don't make it do heavy lifting live. For example, if the user wants a confirmation email, the AI should just say, "Great, I'll send that to you right after we've hung up the call. Then once the call is finished, obviously we can use a web hook to trigger the email, update the CRM, and then notify the team." So, by moving these heavy processes and functions to after the call, we are keeping the conversation more fluid, eliminating the awkward silences, and just ensuring that the data is processed correctly without the user hanging up in frustration. It's easier to track this as well and rerun the end of call automations if ever needed. If you're interested in scaling your own AI voice agency and becoming a better AI developer, we've just released our AI Launchpad school community. Inside, you're going to get access to a complete A Toz course on how to build AI voice agents alongside an A toz course on how to use NAND to build AI workflows and agents. We've also partnered with Bookedin, which is a voice and chat AI client dashboard software, which you'll get as a bonus within the community. We've got over $3 million in software discounts for some of the top AI and business tools, including NN, Make, Stripe, Web Flow, Lovable, and so much more. I guarantee you will make your money back in the first day using these discounts. Alongside all this, I've hired the top voice AI and automation developers to work in the community to provide unlimited technical support. So, if you ever have a question or problem you need solving, we'd love to help you out. I'll have the community linked in the description. Finally, for number five, this is probably the most important lesson of all. AI voice agents do not replace humans. Sales in particular is still best done by humans. People want human connection. They want empathy, nuance, and the feeling that someone actually understands their situation. The real job of a salesperson isn't just answering a phone call. It's building [music] trust and establishing a relationship. That's something AI simply cannot do at a high level and honestly shouldn't be trying to. The underlying core goal of a sales call or onboarding call is to establish a connection and relationship with someone, not just answer questions. So, there's no chance an AI could ever replace this connection because it's not human. AI is obviously excellent at handling the repetitive high volume tasks of qualifying leads, answering common questions, routing calls, capturing details, and just filtering out noise so that when a human does get involved, they're speaking to someone who actually needs them. Another mistake we see all the time is trying to use AI for low volume calls. If a business only gets a handful of calls a day, AI often makes things worse for this, not better. In those cases, a human picking up the phone will almost always deliver a better experience with less friction. If you're a small business and have a low call volume, it's probably best to use that low volume to your advantage to take these calls yourself. There's no need to try and automate those low volume calls. And you should certainly never deploy an AI voice agent into a workflow that no human has ever done manually. If no one on the team has taken those calls themselves, doesn't understand the objections, the edge cases, the weird questions, or how callers actually behave, then you have no chance of trying to automate it. I've seen firsthand that it's best to use real world data of past calls and examples that the AI can help mimic. [music] We always ask our clients for existing call recordings, scripting templates, and resources. Really anything that will help to replicate the existing core processes with our AI.