Subscribe
Share
Share
Embed
Real business owners. Real bottlenecks. Fixed on camera.
Each week Paul sits down with a guest who's stuck — in manual work, disconnected tools, or processes their business has outgrown. They diagnose it live, map the fix, and start building. No theory, no fluff, no generic AI takes. Just the exact solution Paul would build for a paying client, handed over for free on camera.
Every episode is a case study your ICP can't stop watching — because it's their business on the screen.
PAUL KORTMAN
(00:00):
Paul Kortman (00:00):
You're dealing with a multimillion dollar contract here, and if Delta catches wind that one of your agents said, yes, we are Delta Airlines, that would shut down your whole business tomorrow. I'm just using Delta as an example, but an entire contract can be lost if your call center agent says the wrong thing. Every business has inefficient processes, limited hours, and technology that really isn't pulling its weight. This is Connex the Dots, where real problems meet real solutions used by real businesses. So I'm Paul Kortman. Let's connect some dots. Welcome back to Connex the Dots. Today, we're going to cover one of our great case studies about a travel agency call center with a lot of complexities, a lot of information, a lot of data that needed to be reviewed and analyzed and automated. And you might think, okay, well then the solution's going to be complex because the problem's complex.
(01:08):
But actually, even our client was surprised at how simple the solution is compared to the complexity of the problem. Stay tuned though, because towards the end, we uncover this snafu that really complicated everything even more. So first off, this travel agency, just to give you an idea, you may think of travel agency in the past where you would call your travel agent to book things. And why would anybody use a travel agent today in the era of Google Flights and online bookings, et cetera, that are just so simple? However, there are still people that want to pick up the phone and talk to somebody to solve their problem. So the core supposition of this travel agency is that they run ads for when people are trying to get ahold of say Delta or trying to fix their booking with United. And some folks who would just prefer to talk to somebody to have them find the right booking and make the booking for them.
(02:12):
So this travel agency has deals with lots of different airlines. And so they're able to offer similar pricing to online, if not sometimes cheaper, but the travel agency themselves pays cheaper than that. And so they make a margin, a profit in the difference there. So the user is still getting a good price and the travel agency, the call center is making money off of this. So that's kind of the basic profit idea how this works. The problem is they get so many calls and because they have a relationship with each one of these airlines, they can't say certain things. And so just give you an idea, each call probably takes five to 10 minutes looking at all the options, what is it that you're trying to do? You're trying to fix a flight that you already have or you're trying to rebook or you're trying to start a new flight.
(03:11):
All of these are very long and complex conversations. Well, now imagine a supervisor working over a call center and they're trying to analyze how well their staff is doing. Are they going to listen to every call? Are they going to listen to every call from every agent who's working with them? What if they just hired a new agent? How many of those calls do they listen to? And so with each call being somewhere between five to 10 minutes, it just complexifies really fast if that's a word. But take five to 10 minutes, you've got five employees that are working all at the same time. As soon as they get off that call, they pick up the next call and now five to 10 more minutes. How is a supervisor ever going to keep on top of that? And when we came in, they were doing this manually.
(04:05):
These supervisors were listening to spot checking essentially, listening to calls. And every now and then they would catch problems, but they couldn't, as Pokemon liked to say, they couldn't catch them all. And any one of these issues could have brought down the entire agency. So you're dealing with a multimillion dollar contract here. And if Delta catches wind, one of your agents said, yes, we are Delta Airlines, that would shut down your whole business tomorrow. So you can't just let these slip through. How do you do it? So let's take kind of the stakes, the table stakes here. If you're reviewing somebody who's doing a sales call, let's say, and what's the risk? If they say the wrong thing, well, they promised the wrong thing to the wrong person. That's a minor mistake. But in this situation, the stakes are much greater. An entire contract with a major corporation, I'm just using Delta as an example, but an entire contract can be lost if your call center agent says the wrong thing.
(05:20):
And so there's a dozen different things that you have to look out for. These supervisors, they're training these call center agents, but they're human. And they get paid based on how much revenue they bring in. And so they're trying to get people off the phone and get to the next one and that sort of thing. So there's this whole quality control aspect and stress on the supervisor that if you let three or four of these slip, one of those gets back to Delta and then your business is shut down. That's hugely significant. So the stress on these supervisors, not to mention when you show up, roughly there's six and a half hours of recording from the previous shift that you have to review, that's 80% of your shift dedicated to reviewing those previous calls. So the stress is through the roof mentally on each one of these supervisors, not to mention the actual pedium of the work of listening to these calls, reviewing them.
(06:27):
And then it's not just listening to them minute by minute, but then also making notes in the CRM, making coaching recommendations to each agent so that they can do their job better. Needless to say, it's a lot on each one of these supervisors, but it gets even more complex. You may have assumed that we're talking about just English, but they also serve Spanish speaking crowds. And so now you have agents who are bilingual and you have a system that needs to be able to adapt to, well, is this an English or a Spanish call? And what am I going to do if it's Spanish? I have to handle that completely different. Even though we're trying to do the same thing, we have to look for different keywords in the transcription. So now your supervisor needs to also be bilingual and also figure that out.
(07:26):
As if it weren't stressful enough, then they expanded into the Spanish market and it just blew up their systems. Needless to say, our client knew they had a problem, knew they could not scale their way out of this. They couldn't afford to hire five supervisors for every five call center agents. And so they attempted to do some things on their own. And one of the things they were doing was turning to AI for transcription and transcription analysis. However, they ended up spending significantly more money with those services than it would take to hire a human to do that. So they came to us basically scratching their heads going, where's the promise in all this efficiency? How come we're spending so much money and we still don't even have our problems solved yet? So let me break down parts of their system that they were trying so that you can understand.
(08:25):
First of all, they had two call systems. And unfortunately, one of them actually did a transcription, whereas the other one didn't, but they had to pay for that transcription at that first service. The second service, we had to then send, or they were sending the transcription to a service called Whisper that was then they were end up paying like 10 cents a minute or something for transcription services. And that, then they ended up having two transcripts of the same call because of the way their telephony system worked. They couldn't even match those IDs up. And so they just treated both these transcripts as separate calls. So already we've multiplied, we've doubled the cost of each call because now you're paying to transcribe it twice. Then they ran that through an AI analysis tool. But these were back in the days. We've come a long ways just in general.
(09:27):
Most people do more than just prompting today, but back then it was just prompting. And so they would prompt an AI service to analyze these transcripts, but the prompt response was formatted poorly. It was just natural language. And so each call, depending on how long the call was, it could be six pages of transcript. And so that prompt plus the response would cost 100, 210,000 tokens. And so they were losing money because then they were actually doing it twice for each call. Not only the data they were getting back wasn't usable. So that's kind of the lay of the land that when we stepped in, it was like, okay, you're getting natural language back and just putting it into a CRM of like, okay, here's a call, here's the transcript, here's the AI summary. The supervisors were still having to read that summary and the client was ending up paying through the roof for the transcription services and for the AI services, the token utilization.
(10:42):
So combine all that, they're spending a lot of money and the supervisors still haven't. Their problem still has not yet been solved. They're still having to read this paragraphs of information to determine, "Hey, is this a call I need to pay attention to or not?" And so that's when we stepped in and significantly optimized things for them. So even though the client knew, "Hey, we've got a problem," and they knew AI was going to be there to help them, they actually came to us because they're like, "We're spending way too much money on this AI and we're not actually solving our problem." And that's where we stepped in. So kind of for visualization for you, really what they were doing was shoving everything into a drawer, kind of like your digital junk drawer in your kitchen or whatever where it's like, "Hey, I just got this stuff.
(11:33):
I'm going to put it in there. I know I need to keep it. I don't know what I'm going to do with it, but someday I might need that." And so they just kept putting it in there and putting it in there. And they even had to pay more for the storage. Their drawer got expensive, not just putting the data in there, but then they even had to pay for the drawer to be larger. So after they brought us in and after they helped us understand their problem and the lay of the land and the situation, we took our normal approach to this of what's the one thing that we can do? We'll call it low hanging fruit. What's the one thing we can do that will help today? And we could have reduced costs. We could have changed their transcription service right away. We ended up doing that later, but that wasn't the thing that was going to make the supervisor's job easier today.
(12:30):
Sure, that may save them a thousand dollars a month, but the supervisors still were dying on the vine as it were. And then with the complication of the multiple languages and all of this, what we ended up doing is just saying, okay, let's drill down to just English. And yeah, we know we have duplicate calls coming in. We're going to deal with that later because that doesn't help the supervisors. We needed to change that prompt so we're getting back structured data. Instead of just this natural language, we got it back in JSON for those of you who are nerdy and listening, but we got it back in JSON format so we knew, hey, was Delta mentioned? And we were able to ask specific questions. Did the agent claim we were Delta Airlines? Yes or no? And so with this, now we could put that into the CRM into specific fields and then have reports that were like, "Hey, this call needs to be paid attention to because the AI determined that we said we were Delta Airlines on that call.
(13:36):
And that's a no-no. And so the supervisor needed to review that call and address that with the actual agent." That was the start and it just expanded from there. But day one, once we did that, then suddenly supervisors were like, "Okay, here's the 10 calls I need to listen to. And here's the three agents I need to talk to today and what I need to talk to them about." Right away on day one, when we were able to get that data back, push it into the CRM, and they were able to have cleaner reporting, on day one, those supervisors, their jobs changed where they could actually catch problems and not just listen to six and a half hours of calls before we get into the results of this case study. If your team is spending time on manual process checks like these supervisors are, Connex Digital actually offers free process audits where we'd love to help.
(14:31):
The link is in the show notes. So what we did, instead of, we'll use the phrase boil the ocean, instead of like, "Hey, analyze this whole call and tell me should a supervisor review it? And here's the 25 things I'm looking for." Instead, we said, "Okay, AI, I just want you to answer the question for each one of these 25 things. That's all you're analyzing." It actually improved the token utilization right out of the gate, and it gave the supervisors more insight and information, more data that they could use instead of a generic statement of, "Hey, this call needs to be reviewed." We actually were able to have specific yes and no filters as to why this call needs to be reviewed and what was going on there. So instead of trying to eat the whole elephant in one bite, it was taking little bites of like, okay, yes, AI can do amazing things, but let's just boil it down to, was Delta Airlines mentioned?
(15:40):
Was it claimed that we were Delta Airlines? Was a credit card received? Did they charge the credit card? Did they upsell this? Et cetera. Those were the 25 questions that we had the AI come back with an answer, structured answers to, that then we put in the CRM and then it becomes an easy filterable view to be able to say, "Hey, which calls actually sold tickets?" And so they could review those. Which calls did not have upsells in them? And so they could review that. And it became more of a thematic thing of like, "Okay, for the next hour, I'm going to look at..." Anytime there were a problem, anytime there was a threshold tripped on a certain problem, an alert would go to a supervisor. That was the next step of like, "Okay, great. Now all these calls are happening. Now we actually have actionable data." When there's an issue, a contract potentially devastating issue, the supervisor needs to know about that call within seconds of that call happening so that they could address it.
(16:51):
So by boiling it down to just one fix of like, okay, let's just get structured data back. Remember, there's Spanish and then there's we've got multiple transcripts coming in for a single call. All of this information is getting lost and weird and eventually we'd get to that. But by starting with, let's just get this data structured, it quickly went into, oh, now that we have that, we can get alerts. And those alerts became more of a priority than reducing cost because that was mission critical that could have devastated the business. So massively complex system. As we talked about in the beginning, all these different issues, different problems and factors that made it even more complex. And by solving one problem, and I keep talking about structured data, but by getting it to where basically it's true/false or a five-star rating or a three-star rating. Yes, no, maybe, something like that.
(17:58):
By getting that data back, we saved on tokens because there's a difference between a paragraph response and a yes response. We boiled the paragraph down to yes or no. So when I say structured data, that's what I meant is like we got it back to just did this happen? Yes or no? Or even an I don't know. So that's the three ratings of sometimes the AI could not determine whether certain things happened. And so we gave it the ability to say unknown. So anyways, all this, by bringing it down to these true, false or very simple answers, that junk drawer analogy that I was talking about, we could actually right size the junk drawer because we're not storing copious amounts of paragraphs of content and data in there. We're just storing yes, no information. And ironically, their CRM had some difficulty with character limits. One of their field types had 2,000 characters and another field type had 10,000 characters, but they didn't have anything beyond 10,000.
(19:05):
So when I talked about expanding the junk drawer and right sizing it, we went from 2,000 to 10,000 and then we couldn't go beyond that. And so it was very much like we needed these just very bite size information that ended up saving them costs right out of the gate and made their junk drawer more organized. So it became less of a junk drawer and it became more, here's the drawer for good calls, good examples. Or here's the drawer for calls that are potentially dangerous. Here's the drawer for calls that it was the wrong number or they weren't interested in buying a plane ticket. Those became easily identifiable, easily sortable. And then we could actually remove ones that we don't need anymore because they're not examples. We don't need them for discovery or anything like that. It's just, oh, that was a wrong number. Get rid of it.
(20:01):
And that saved on storage. So we were able to, when you bring it down to the simplest thing, then you can expand and solve problems along the way. As well as like I was talking, the alerts, have extra features that you hadn't even thought about or hadn't even dreamed of because the system was so complex, there was no way to get an alert. By bringing it down, getting it, solving that first thing of like, what's the first thing? And the client actually thought the first thing was doing the transcriptions. And then the second thing was analyzing those with just a prompt. Yes and no, but the first thing was getting that data back as a true, false, yes, no structured format so that we could solve lots of problems. Again, they came to us to solve this problem to make sure that they didn't have compliance issues.
(20:59):
But one of the benefits that came about from this was that those supervisors now could become coaches. And the supervisors weren't necessarily really good at coaching, weren't really good at advising an agent to be better at selling, but the AI has unlimited knowledge and information. And so on those potential calls where it was like, "Hey, this could be a coachable moment," which was that true false stuff that we had the first analysis go through. If that came back as true, then we fed that to a second prompt to say, "All right, give the supervisor bullet points on what you would recommend for this agent to be coached on." So now we took it to a second tier that didn't need to happen. The client hadn't even though of that before, but because we had that raw data, we're able to go to an AI, train it to be the best coach of the supervisors so that they could coach the agents.
(22:13):
And so again, just all these little features that come out, once you have the kind of core problem solved, then you can build on that. And so a lot of times clients will come to us with, "I need," in this case, "I need my supervisors to be better coaches." And it's like, "Okay, well, we need to break this down and get down to these small blocks," like think Lego blocks of like, "Okay, you want a castle? Yep, I do too. But we have to know where the walls are and what the bricks are to be able to build that and give your supervisors the ability to be better coaches." And that's kind of like a, "Oh, and one more thing that kept happening with this client and this solution of once we got those basic fundamental bricks in place, we could build all kinds of things upon that.
(23:06):
Okay, let's review the before and after, but in numbers in detail." So before they were working with us, a supervisor needed to spend roughly 400 minutes per shift, which works to be about six and a half hours or 80% of their shift, reviewing calls from the last shift. We now reduced that down to 50 minutes, about an hour. So only reviewing flagged items, things that needed attention or things that needed coaching. So doing the math, whereas a supervisor would need to spend six and a half hours on each shift, we reduced that by 29 hours running the math across their average pay scale, et cetera. We saved $66,000 per supervisor per year for this client. And that's just the raw numbers there. There's a whole bunch of soft numbers that we saved as well, being their mental health, the concern of the missing mission critical information that leaked out.
(24:16):
And then there's also just the general environment of knowing, yes, our calls are being supervised. Yes, nothing is falling through the cracks. And that whole soft number stuff of, I don't want to say touchy-feely, but just helping improve quality of life for the agents, for the supervisors, and for the owners of the company of just knowing this is taken care of and our supervisors actually can do their jobs. The other numbers, just to kind of throw that out there, is they were spending 40 to 50,000, let's say, tokens in a given time period, and we were able to reduce that down to 10,000. So a drop by 75% roughly of just hard costs with their AI, with their transcription services. We did eventually work through reducing it. So we only transcribed each call once, and that went through a matching process to make sure we got the right call.
(25:17):
And then also that we were only analyzing the things that we needed to analyze. So we actually had a pre-analyzer, and then that would determine, hey, was this call longer than a minute?
(25:33):
Was it the right call to be analyzed, et cetera. So we're able to just continually chip away at dropping the token usage. And so you can do the math. I don't remember the current cost. It kind of feels like fuel prices, but 10,000 tokens, say that's $500. We were able to take them from 40,000 tokens down to 10,000 tokens. So save 75% of their money that they were spending every month with the AI tools. If you're listening to this and you run some sort of high volume operation or business, here's my one takeaway that I want for you is to drill down to what is that one thing in there that you could solve? The tiny little thing that would have a domino effect on improving everything. What's that one little thing? And if you are having trouble identifying it, reach out. It is one of the things we do is help you identify what is that one piece, the low-hanging fruit, the thing that would make everything more efficient.
(26:40):
So if you're listening to this and you want to understand more about how to practically improve the efficiencies inside your business, check out our other episodes. That's what they're here for. But if you're ready to start, reach out. We offer a no-cost consultation and I'd be happy to help you.