Episode Summary
David Cady, Netfor's CTO, joins Ian Medley to break down why traditional IVRs fail and how AI-powered IVAs fix the intake problem. David traces his path from a six-figure-call broker career at Charles Schwab into building Netfor's BI department and eventually leading its AI strategy, and argues that most IVRs fail because they assume callers arrive prepared, when most don't. He explains how retrieval augmented generation (RAG) lets Netfor feed an LLM the right knowledge article at the right moment instead of relying on training data, why hallucinations can be reduced but never fully eliminated, and how pre- and post-generation guardrails keep AI from making things up. The two dig into intelligent routing and prioritization, why current transformer-based AI is nowhere near AGI, and why companies that try to build AI in-house without an experienced partner see far worse outcomes than those who don't. David closes with his standing test for any AI vendor: if it isn't solving a real problem, don't use it.
Full Transcript:
Ian Medley: So this is the first Netfor, currently unnamed, podcast where we have brought in David Cady, our CTO, to talk a little bit about IVRs and IVAs and the reasons behind why they're failing, what's coming in the future, and just the general backstory of the CTO's role: what you do, how you do it, and how you're pushing Netfor forward. So I'll let you have the floor. Tell me a little bit about you and Netfor. I know you and I started working together in cubicles next to each other about five or six years ago, back pre-COVID, and it was a fun time. We've grown a lot, we've changed a lot. So tell me about where you started and where you're at now.
David Cady: Thanks, Ian. It's great to be here, I appreciate the invitation. So I've been with Netfor about five and a half years now, almost six years. I started just as the senior analyst. When I started here we really didn't have a very mature BI department, so my first role was setting up the BI department. From there I grew into being the BI manager, then the Director of Data Operations. And then with the advent of AI, I took the lead role introducing Netfor to AI and how it's used, back in 2022-ish, which led to my current role as CTO, where I mostly design our AI products that marketing can then take to market and solve real issues.
Ian Medley: That's the goal.
David Cady: Yep. And the main focus so far has been... for a little bit of backstory, I spent my first career as a broker for Charles Schwab. We worked in a call center environment, and I probably, without exaggerating, answered a couple hundred thousand phone calls during my career. So I built a lot of soft skills doing that, which when I moved into the tech side of things was a little unusual. Most people who come up through the tech industry, we can say they lack some soft skills, to be generous. So I came into the tech industry already having those soft skills built, and then I built the tech knowledge on top of it, which gave me a certain advantage. I also understand IVRs extremely well.
David Cady: I was around back in the mid-90s in the call center when they were first being introduced. They operated on a premise that your callers know what they want and are prepared to answer questions like account numbers and things like that. They presumed a certain level of caller readiness in order to work correctly, and occasionally they did. Occasionally people would call in with all their documents and questions ready, and they could navigate the IVR properly and it would help them. But most callers, over the years, don't operate within those bounds.
Ian Medley: And that's a lot of assumptions to make about callers being ready right off the bat.
David Cady: But if you think about it from an engineering perspective, you have to start somewhere. So the obvious thing to start with is that they're going to know why they're calling. But if you've ever worked in a call center, you learn pretty quickly that they absolutely don't.
Ian Medley: Yeah.
David Cady: Or at least they don't know the specific area they need to be talking to, or they don't know the specific question they should be asking, especially around things like insurance products or equities. Even something as simple as, Comcast famously has a terrible IVR, and I happen to know a bunch of people who work there. I've always wanted them to run the experiment of just getting rid of their IVR entirely and having people answer the phones from the get-go, something like a classic receptionist. Train them on which departments people need, train them to answer some very entry-level questions. The cost of operating those humans may be higher than operating the IVR.
David Cady: But my thesis is that in the end you'd save money because you get so much less rerouting of phone calls. You can train those frontline individuals to answer simple questions, hours of operation, address, things like that, and save your call center agents for the more complicated, nuanced questions. When ChatGPT came along, I realized we could actually implement this experiment, because now we have the ability to scale. I don't need to hire hundreds of employees to run that receptionist operation, I can train an AI to do it and get a lot of success.
David Cady: So now we have the opportunity to take these ideas I've had for the last 30 years and actually implement them at scale, getting the best of both worlds. The front line isn't "press 1 for this" or scream "operator" until you're blue in the face hoping something good happens. The front line is a natural conversation, so it's a good customer experience, but you still get the benefits of automation.
Ian Medley: Yeah, the intake is something that's always lacked within IVR spaces. Like you just said, pressing 1 or screaming just to talk to somebody, even when you get on the phone it can automatically tell you what the experience is going to be like. Something you mentioned was the cost savings, and I think those savings will also come from an increase in customer experience, because I can confidently say I've left an unnamed internet service provider because of how horrible their customer experience was to go through. And you're right, sometimes I don't know what I need, I'm just trying to fix something and I don't know what the issue is.
Ian Medley: All I know is the internet's down, call this number, and then I have to go through hoops or phone trees, or I'm Googling Reddit trying to find a real customer service line instead of the one they throw on the box.
David Cady: Yep. And when you train a whole industry that the moment they hear an IVR they start immediately pressing 0 or saying "representative" or pressing 1, that's not a failure mode, that's systemic. You're not getting any of the benefit you were hoping to get out of the automation, and you're negatively affecting everything around it. My personal opinion is that a lot of the people who design these systems have never worked in call centers. They don't have the soft skills necessary to understand how a conversation should flow from a good customer experience standpoint.
Ian Medley: Yeah, I think that's huge, and a big differentiator for you since you've taken those hundred thousand calls and know exactly what to expect. Is there anything positive about an IVR anymore?
David Cady: Honestly, no. Cheap, maybe, though not even necessarily cheap. I happen to know an unnamed internet service provider pays about $35 million a year for their IVR contract. I know the company that builds it for them, so it's not exactly cheap. Is it cheaper than deploying however many humans it would require? Probably. But there's an opportunity cost that most companies aren't willing to recognize. I can't imagine anybody in ISP leadership not understanding that people despise their IVR.
Ian Medley: Yes.
David Cady: I think it's just never risen to the level of "oh my gosh, we have to fix this."
Ian Medley: And now, I think it's coming down the pipe, the IVR's time to shine is now, because they're about to get cut.
David Cady: I don't really see it that way necessarily. I think in a couple of years, once there are fewer rough edges in AI and ML, how you deploy it makes all the difference. We take a lot of care in guardrails and making sure we have very tight controls over what the AI can and cannot do.
Ian Medley: Yeah, that's important.
David Cady: Unfortunately we've seen a lot of products in the market that don't do that. Every time you read one of these stories about some random AI suggesting adding glue to your pizza to keep the cheese on, did you read about that one?
Ian Medley: That would be considered a hallucination.
David Cady: Yeah, and it makes my job so much harder, because AI naturally does occasionally hallucinate.
Ian Medley: Yes.
David Cady: But a lot of the hallucinations people run into in production, if I deploy an IVA there's generally some kind of knowledge base you're pointing it at, and if your knowledge base is some shared drive with no management over it, of course the IVA is going to hallucinate, in the exact same way that if you hired me and said "somewhere in this shared drive is your answer," there's a good chance I'd pull up an old or incorrect document. It doesn't matter how smart I am as a person, you wouldn't just throw the smartest person in the world into your business and say "have fun." You'd put structure around them.
David Cady: That same kind of structure is what these AI products are lacking right now. A lot of products on the market don't have what would be called the correct harness. They're not guardrailed correctly, they're not shown what to do when they can't answer a question, so you want to prevent them from just making one up.
Ian Medley: Right, and I've seen that in my GPT and Claude experience. I'll just tell it, you don't need to answer me, if you don't know the answer you can just say you don't know. You don't need to make something up or best-guess it. I'd much rather have no answer than a totally made-up wrong answer, or glue on pizza.
David Cady: That doesn't sound fun. Yeah, and there are tons of examples out there we can learn from. It's fascinating that back in the 70s we started building all these frameworks around how technology is supposed to be built, it's supposed to be QA'd, it's supposed to meet these standards. Then AI came along and as an industry we just got amnesia. We were like, just throw it out there, it's fine, I'm sure it'll work great. We didn't test it, we didn't QA it, we just started throwing things at the wall hoping it would work, and then when it didn't, we're like, gosh, I don't understand why it didn't work.
David Cady: Uber spent hundreds of millions of dollars testing their product before releasing it, because it's a very complicated product.
Ian Medley: Do you think there's a lot of companies out there prematurely launching? Like, what do you think the ratio is, 95%?
David Cady: I don't know, I think that ratio is coming down recently. I think people are starting to understand it's not an autonomous worker. It will never be an autonomous worker.
Ian Medley: No.
David Cady: And that's the point, it's not trying to be an autonomous worker. It's an exceptionally good worker, very scalable, and it can do what it's told pretty well most of the time. As long as you structure it the same way you would a live human agent, give it feedback, give it the proper structure, I think most people would find the AI errors somewhere in the same neighborhood as a human would.
Ian Medley: Yeah.
David Cady: Every human on earth is going to make a mistake on occasion. Out of those several hundred thousand phone calls I took, I guarantee I made a few mistakes. The AI would be on that same level if we treated it properly, in my opinion.
Ian Medley: Yeah, and I don't think expecting perfect results from anything, AI or human, is ever the correct path forward. I know we wrote a blog a few years ago about AI SLA compliance, about why every SLA can't be 100%, because we're acknowledging nobody's perfect, never has been, never will be. And I think that's a part of Netfor that may be different, we've been doing this a while, we've got people like you who know the skills, know the environment, know everything there is to know about taking a call and what's required of it. That's why we can build SLAs that are really accurate and have good first-call resolution rates, because of our knowledge base.
Ian Medley: Which is another big point I wanted to bring up, knowledge bases and AI. Give me a general sense of how AI is influencing companies' knowledge bases and helping surface questions or answers sooner when customers call in with a repeat or new edge case.
David Cady: Yeah, so I think the main lesson the market is learning right now is that your knowledge needs to have lifecycle management applied to it. Everything from creation to using that knowledge to retiring it, there's a specific lifecycle it goes through. And for thirty-some-odd years that's kind of how Netfor has operated, we've always built knowledge articles for virtually everything. We have a knowledge article for every possible question our agents can answer.
Ian Medley: A knowledge article for a knowledge article.
David Cady: Yes, we have a knowledge article for what to do if you get dead air on a phone call. We even have knowledge articles for what to do when you can't find the correct knowledge article.
Ian Medley: But you can find that one, it's kind of like a tongue-twister knowledge article.
David Cady: What we realized is that if we take that same lifecycle management and apply it to... for anybody in your audience not familiar, there's a term called RAG, retrieval augmented generation.
Ian Medley: Yes.
David Cady: You can think of it as, when you're talking to ChatGPT, I insert myself between you and ChatGPT and take your question, attach some kind of knowledge article to it, and when I pass it to ChatGPT I say, please answer this question using this knowledge article. So we have a way to refine the knowledge the AI can use to answer your questions. By doing this we can insert company-level knowledge, very specific answers, verbatim wording, and all kinds of things like that. What we're doing is taking our thirty years of lifecycle management practice and applying that same concept to the retrieval augmented generation we're building.
Ian Medley: Say that again, just slower for me.
David Cady: Retrieval augmented generation.
Ian Medley: You lost me on all three of those words.
David Cady: Let me think of the best way to explain this. We are augmenting, we are changing the retrieval of information for the LLM.
Ian Medley: Okay, that makes sense, so it's changing where it's grabbing information from. It could be grabbing from the internet, from a knowledge article, or from an attached document.
David Cady: Well, typically, if you don't give it any context at all, the only context it can use is whatever its training data is. If you imagine an LLM trained on Reddit, you're probably not going to want to just ask it random questions, you're going to want to be very specific about the questions so it uses the correct training data to answer them. And we don't actually know what any of these LLMs' training data is.
Ian Medley: Right.
David Cady: So what we can do is, say you have a point of sale and you know the model number, and someone calls in saying that point of sale is down, we can look up a knowledge article for that specific point of sale and say, follow these steps to determine why it went down and how to fix it.
Ian Medley: Interesting.
David Cady: If I didn't give the LLM that knowledge article, it might still answer correctly based on its training data, but you'd be hoping that somewhere in its training data it happened to be exposed to that very specific point of sale, which probably didn't happen. So we have a way of teaching the LLM. The way I first described this to my boss was, it's like every time somebody asks a question, we pause time and quickly train the LLM on how to be a Netfor agent, and then unpause time and let it answer. This happens in milliseconds, but every time someone asks a question we give it a quick training update: here's a knowledge article, learn this, then answer that question.
Ian Medley: Cool, I think I understand it at a high level, but that's interesting, that's a lot of information.
David Cady: Yeah, and you can do this on your own too. Think about, maybe you recently graduated college, not that...
Ian Medley: Long ago, like six years ago. I don't know if that counts, compared to you.
David Cady: Say you were assigned some chemistry document and you just flat out don't understand it.
Ian Medley: Right.
David Cady: You're just not grasping it. A very common thing students do nowadays is simply drop that document into Claude or GPT or whatever and say, please explain this document to me. This is the exact same thing.
Ian Medley: Really?
David Cady: This is the exact same thing. The order of operations is a little different, but we're basically taking a knowledge article and giving it to an LLM and saying, please answer questions based on this article.
Ian Medley: Okay, that makes sense. So what happens if one point of the workflow is totally broken, you've got an IVA or an IVR with broken knowledge but you want to band-aid it with AI, is that even possible?
David Cady: Absolutely. Yeah. At a high level, the concept is called failing gracefully. If you don't do anything and ask an LLM a question it has no idea about, it's literally going to make things up, whatever sounds good in the moment. You'll never get the same answer twice, it will make up something brand new every time, and even if you tell it, hey, that's not correct, it'll say, oh sorry, and make up new stuff. It becomes a never-ending cycle.
Ian Medley: So...
David Cady: What you want is what's called guardrails. Right now, the market mostly uses what are called post-generation guardrails. You ask GPT a question, it answers, and before I show you the answer I look at it and say, no, that's wrong, and send it back. Or I look at it and say, hey, don't swear at this guy, and send it back. I can put guardrails on what's allowed to happen, and if I ever reject a message I can substitute a different one, something like, I'm sorry, I don't have the proper knowledge to answer that, let me get you over to a human who can.
David Cady: There are also pre-generation guardrails, which we're exploring here at Netfor, that we've found extremely successful in reducing hallucinations. And reduce is the key word. MIT proved, I think about a year ago, that mathematically you cannot ever fully get rid of hallucinations, they will always exist. To eliminate them entirely you'd essentially need infinite knowledge in finite space, and logically you can't do that. But with proper pre- and post-generation guardrails you can drastically reduce them, and hopefully catch them before they're exposed to the user.
Ian Medley: Okay, another...
David Cady: A lot of information, yeah. The thing that fascinates me about this whole industry is, even prior to 2023 when ChatGPT came out, I knew about it and played around with LLMs, and it was very difficult to interact with them, basically a command line where you had to know what to type, and getting the generation, the retrieval, was always a pain. What ChatGPT did, their innovation wasn't creating LLMs, it was creating the interface by which we interact with them. That was their huge contribution. When they came out with that, I knew it was a watershed moment. I was also alive when the internet first became a thing.
Ian Medley: I think I was too.
David Cady: When did it come out? Technically the late 80s, but it didn't really hit popularity until the early to mid-90s. I was born in the 70s, so it was pretty easy to recognize early on that this LLM thing was just as impactful as the internet.
Ian Medley: That's what everybody's saying, it's like the second coming.
David Cady: It is, absolutely. I think in the end it'll be more impactful than the internet. The only analogous relationship I can think of is the Industrial Revolution, where we automated manual labor.
Ian Medley: Yeah.
David Cady: This AI revolution, we're automating thinking labor, and after that there's really nothing left to automate. This is going to be the last of these revolutions we'll see for a long time.
Ian Medley: Do you think AI and everything surrounding it is going to come for people's jobs, or make people's jobs easier? I know it's a split divide, some people think it's going to take over the world and nobody's going to have a job, others think it's going to make everybody's job a lot easier and more productive. I personally think it makes my life more productive. I can do things quicker than I was able to before, I've been able to replace certain workflows because now I can automate them, turning a 10-hour task into a 10-minute one. It changes the game. But I want your input on where you think it's going.
David Cady: I think, yesterday I had a coding thing to do that normally would have taken about half an hour, and instead I got it done in about 30 seconds with Claude. So I'm the first to recognize that the scale of this thing is beyond anything we can really prepare for yet, whether or not it replaces people's jobs.
Ian Medley: I don't want to fear-monger people here.
David Cady: The products we're building aren't designed to do that. They're designed to enhance people's ability, like you said, to let them be more efficient, to let them address the things that require a human brain rather than the repetitive things that happen again and again. Those repetitive things are where we're targeting AI.
Ian Medley: Yeah.
David Cady: AI is not ready to do a lot of different things.
Ian Medley: Right.
David Cady: If I ever called a hospital, I wouldn't want to call a psychologist and go through psychotherapy with an AI, let's just put it that way. You have to have...
Ian Medley: It's going to start making things up again.
David Cady: Yeah, it's nowhere near that, and I don't really foresee a time when it would be. The basic technology by which an LLM operates, at a very deep level, is these things called transformers, and this basic transformer technology, in my opinion, precludes the possibility of AGI, artificial general intelligence, human-level AI. Foundationally, these transformers, it doesn't matter how nifty we get with the math, in the end it's still just math.
Ian Medley: Yeah.
David Cady: In the end, all we're doing is running a very long, complicated calculation and showing the results. So we can mimic reasoning, but we can't actually reason. If it doesn't have the proper context, it can't even mimic reasoning, whereas a human can take their whole life's worth of context into account, and even when presented with a situation they've never encountered before, use their past experiences to inform how they react. AI is never going to get there, not with the current level of technology. Having said that, there are so many use cases, intake for an IVR is such a good example, in my opinion.
Ian Medley: Is that considered an IVA at that point?
David Cady: It is, yeah.
Ian Medley: So an IVA replaces the IVR with smarter intelligence and smarter intake.
David Cady: Yes.
Ian Medley: Okay.
David Cady: Because you're solving a problem. If your AI isn't solving an actual problem, you're using it incorrectly.
Ian Medley: Yeah.
David Cady: I can think of tons of examples. Famously, Meta just announced a couple of days ago that they're shutting down the metaverse. They spent something like $80 billion, and you don't even know what it is.
Ian Medley: The VR thing, yeah, okay.
David Cady: They spent $80 billion on it, and your face just now told me you don't even know what it is. That's how unsuccessful it was. The entire reason it was unsuccessful is because there was no problem it was solving.
Ian Medley: It was a thing that existed for no reason.
David Cady: Yes, there's no point to it. And that translates into everything we should be doing with AI. If you're not solving a real-world problem with it, you shouldn't be using it. The problem we're solving is the IVR stage.
Ian Medley: Right.
David Cady: The intake stage. We're not saying everything that happens after that is also solved by this product, because everything after that is more nuanced, harder, we're not there yet with this technology.
Ian Medley: That's usually handed off to an agent right now.
David Cady: Yeah, absolutely, as it should be. But things like, what are your hours of operation, where do I send my bill, what's your website address, tier-zero stuff.
Ian Medley: So repetitive, and it's just information.
David Cady: Yes, so now you have the ability to remove all of that and let your human agents focus on the more emotionally or technically nuanced questions.
Ian Medley: Yeah, which is what they want to do. I don't think people want to answer the phone and go through 95 password resets, I think they want to be challenged.
David Cady: Yes, absolutely.
Ian Medley: I think they want to deal with new edge cases that come across, things AI will just gladly hand off because it doesn't have a clue what it's getting into, since we've set up guardrails to not let it solve a problem if it doesn't know what's going on.
David Cady: Yes. Intrinsically, the way the market played out, we built these LLMs to really want to be helpful. On one hand that's good, but on the other hand it creates the scenario where, when it can't answer a question, it just won't tell you "I don't know" unless you tell it to do that.
Ian Medley: Yeah. Now, good intake from an IVA standpoint, I think its best use case right now, I could be wrong, I'm not deeply knowledgeable in the IVA space, is that it can save callers about 30 seconds by immediately accepting the call and prompting questions like, what's your name, where are you calling from, who's the client, getting 30 seconds of information, and then handing off to an agent who now has a handoff packet, information on why they're calling, who they are, whether this person or store has called before. So I think that saves a lot of time, and 30 seconds saved over 5,000 calls a month, I don't know the exact math, but that seems like a lot of minutes, 2,500 minutes.
David Cady: And I'd say realistically you could be looking at far more than 30 seconds, more like two minutes. Think about it, can I have your name, email, phone number, account number, whatever identifying information, then on the back end you're looking that up against a database to see if they have access or have called before, then asking why they're calling, determining whether it's a tier-zero issue the IVA can knock out so the caller can get on with their life. And the key part you mentioned was the handoff packet.
David Cady: When it does need to escalate to a human, everything that was said and done needs to transfer along with that. And at the same time, there's no reason it can't already be looking up the proper knowledge article to answer these questions.
Ian Medley: For you, on the back end, it can just have it ready to go so you can speak to it.
David Cady: Yes. And you can go as far as, while the agent is on the phone with the caller, listening to that conversation and suggesting certain things they should be saying, or giving them notice that, hey, the sentiment on this call is turning negative.
Ian Medley: That's real-time interaction guidance.
David Cady: Yeah, and as long as those things are handled right, you don't want it to just be noise for your agent.
Ian Medley: Right, right.
David Cady: It has to be useful.
Ian Medley: Solve a problem.
David Cady: Yes, again, it has to solve a problem. I could think of lots of times, even back in the early 2000s, whenever I had to answer a question I had to find the answer, and the way the company I worked for let me search for that was very different from one company to another. In some instances you're just expected to know and memorize it from being told or having read it at some point, and sometimes companies would go as far as producing a wiki for you to look questions up on. This is the same kind of thing, we're just literally scaling it with an LLM.
David Cady: We're just letting the agent use the LLM to do the things they'd be doing anyway, but faster and easier.
Ian Medley: That brings me to another good point we're seeing through a lot of retailers and franchisers who are clients of ours, prioritization and escalations. We get a lot of calls because they have department stores, thousands of stores nationwide, and something that could be revenue-blocking, like a broken cash register, is way more important than a password reset. So AI can prioritize and escalate to a human way quicker if that case is deemed, via the IVR, to be an important revenue-blocking issue.
David Cady: Yeah, so with IVRs they attempted this with something called intelligent routing.
Ian Medley: Right.
David Cady: That became a fad maybe in the last decade or so, where instead of pressing 1 you could say "one," bringing in text-to-speech.
Ian Medley: Yeah.
David Cady: And then they thought, maybe we could use this text-to-speech stuff to help get the routing correct and reduce misroutes. It never really took off, because there again you're just routing a problem, not solving it. With IVAs you have the opportunity to actually solve the problem. With the priority thing, not all calls are the same. If I'm calling in to order more light bulbs, that's one issue, but if I'm calling in about a fire, that's a completely different issue.
Ian Medley: Right, yes, and it should be escalated a little differently.
David Cady: Right, but the queue we're going to is just one queue. So how do you handle that? With the advent of LLMs, we're able to understand context much better, so we can just have a conversation with the caller: hey, is what your problem is preventing you from making a sale right now? And if they say yes, we immediately shut up and get them to the next available person, and they move to the front of that line. We can't always say we'll immediately connect you with a person, because there's still going to be a queue, but no matter how many people, if there were 200 waiting, you're moving to the front.
Ian Medley: Yeah.
David Cady: And that's been entirely successful, which is something you couldn't do with an IVR. If you gave an IVR an option, "press star if you're not able to make a sale," within a week everybody would just press star. You'd train them to press star, and if everybody's an emergency, then nothing's an emergency. By being able to understand context, ask questions, and have callers answer them, we're able to actually pull off something like intelligent routing.
Ian Medley: And that's solving a huge problem.
David Cady: Yes, absolutely.
Ian Medley: That's awesome. So people are trying to implement this through their own companies, people like you, CTOs, IT leaders. Where do you think they're getting their wires crossed, and what advice would you give on implementing things like this to help with scalability, routing, immediate answering, and customer experience? Where's the failure, and where's the opportunity?
David Cady: The very first thing I'd tell anybody is do not try to do it on your own. Absolutely not.
Ian Medley: Not even if they're really smart and took 100,000 calls before.
David Cady: Working at Charles Schwab, we worked with a lot of partners to develop our solutions over the years. We have actual empirical evidence that when you're trying to implement AI, especially for the first time within your corporation, doing it on your own versus doing it with a partner who has done it before, you'll get a lot more success, something like four times the success working with a partner. There have been multiple studies showing this. I don't want to sound harsh, but all of us have dealt with or worked in an IT department that was an absolute catastrophe, very siloed, very little documentation, no knowledge base.
David Cady: There were just notes that one engineer who always wears the hoodie with the hood up kept, and he runs everything, and if he left your company, everything would blow up.
Ian Medley: Right.
David Cady: We've all dealt with those kinds of IT departments. So now you want to take a brand new technology that's far more complicated than anything they've ever done before, and let those same exact people try to deploy it, it's just not going to go well. You need somebody who's been in the trenches, who knows what to expect. If you didn't know going in that guardrails were a thing, or how to implement them, you might end up with a solution that lacks guardrails, and that's how you get dangerous products, and that's how you get all of these products on the market right now that aren't inherently good, because they were rushed. Nobody's an expert on everything, you can't be an expert on everything.
David Cady: So you want to bring in people who know what they're talking about, you want to listen to those people.
Ian Medley: How do we separate the noise from fact? If somebody's got a large marketing budget and the loudest voice in the room, how do you differentiate what that person says from somebody like us, just starting off on a two-person podcast for the first time?
David Cady: I've started using this term I call demo-ware. There's a lot of products on the market right now, and I've personally seen many of them, where when we ask for a demo of their product, they can put together something that looks really, really good. Everybody knows the story of the first iPhone demo Steve Jobs did, he didn't have just one iPhone, he had like 12, and he'd switch one out because each iPhone could only do one thing correctly, there was no single one that could do all of them.
Ian Medley: I never heard of that.
David Cady: Oh, yeah. At that very first demo he had a podium, and in that podium there was a bunch of iPhones, because each one was set up to do a very specific function correctly, no single iPhone could do all of them. This is a very similar situation where these demos are built and brought to me and they look shiny and really good, and then when I try to productize them or put them into production, they fail at the slightest touch, because what happens in a controlled demo environment and what happens in production are two totally different things.
Ian Medley: So if you're seeing red flags, too good to be true, right up front with the demo, and then you get into onboarding and it starts to hallucinate or break at the slightest touch, that's a huge red flag, though some companies might brush it off and say, oh, it's new, it's the first year of AI, we're still working out some kinks. But do you think those are still just...
David Cady: When I asked, I'm not going to name the vendor, but I had a vendor where I said, why is this not working the way you guys demoed it, and their answer was, well, it's really hard.
Ian Medley: Okay.
David Cady: And I was like, yeah, that's not really the answer I was looking for there.
Ian Medley: No, that's an immediate DQ for me, partner.
David Cady: Yeah, and it is really hard, I won't doubt that. But don't claim you can do it if you can't. The marketing fluff is worse than I've ever seen it. Every company with an AI product says they've solved all the problems and everything else. I think the proof is in the pudding, if you can talk to other people who've used that solution, that's very helpful. If you're considering an AI vendor and have a project you want them to complete, talk to somebody else who's had a completed project from that vendor, that will tell you very quickly whether or not they can actually walk the walk.
Ian Medley: Yeah, that does a great job answering that question. So do you have any...
David Cady: Your knowledge base is the last thing. You absolutely cannot just turn AI loose in your environment, it is not an autonomous worker, please don't try to treat it like one. I think I wrote an article once, the example I gave was, if I was able to hire Einstein, it doesn't matter if I hire the absolute best person at that job on the planet, they're perfect, I'm still not just going to turn them loose in my environment and go golf for a while and say, please don't mess up. You're still going to train them on how to operate within your company, still going to check on them occasionally, see if they're doing a good job, give them reviews.
David Cady: You're going to incentivize them to act a particular way, because that's how your company culture wants it to be. There are all these things we've developed around, I hesitate to use a word like employee maintenance, but it's the structure around employees by which we want them to operate, because we've found that's how the business runs most efficiently, that's how we give the best customer experience, that's how we achieve our results in the best way. We incentivize and train our employees to act those ways. You cannot expect AI to do its job well without treating it the exact same way.
Ian Medley: Yeah, that's a really good point.
David Cady: Yeah.
Ian Medley: Any other parting notes before we close it out?
David Cady: No.
Ian Medley: Close it out.
David Cady: I'm very excited about the future. This industry is growing so fast and things are changing so much, every time I look at it, it's different. It's very difficult to keep up on all the new, exciting things, but...
Ian Medley: I know, but if we can solve the IVR problem with intelligent IVAs, good intake, good routing, I think that's a pretty big problem we can solve for a lot of people out there.
David Cady: And on a final note, about it replacing employees...
Ian Medley: Oh, yeah.
David Cady: No, it doesn't have to. There's no reason it has to. There's nothing inherently about this technology that says it has to be a negative for everyone, it can be a positive for everyone. It just has to be, we have to have the right goals and the right incentives.
Ian Medley: So defining goals and incentives should be something, besides looking at reviews of AI companies claiming to do things, that people and leaders bringing AI to their company could focus on, like what you just said.
David Cady: Yeah, if you're not solving a problem, you shouldn't be using AI.
Ian Medley: Boom, we'll leave it on that. Thank you very much, thank you for coming in.
David Cady: It's been great, I appreciate it.
Ian Medley: Yeah, of course, we'll talk to you soon.
David Cady: Awesome.
Full Episodes Available on YouTube and our Podcasts page