What Are AI Agents and How Will They Transform the Travel Industry?

In this first episode of the AI Compass podcast, Jason Elkins and Jason Halberstadt dive into the world of AI agents — what they are, how they differ from old-school chatbots, and how travel companies can start using them to automate real business processes.
 
 💡 Discover how integrating AI into your internal systems can turn your assistant into an autonomous agent — one that works while you sleep, evaluates your website, finds flaws in your sales process, and never forgets a policy.
 
 🧭 Whether you're just getting started or already using tools like ChatGPT, this conversation gives you a clear direction to start building the future of your travel business.
 
 👉 Subscribe for weekly episodes about AI, automation, and innovation in travel.
 
 🔗 Learn more at https://mytrip.ai
 

Book a FREE AI strategy call with host Jason Halberstadt at MyTrip.AI.

Transcript

Jason Elkins (00:00)
Welcome everybody to episode number one of the AI Compass podcast. My name is Jason Elkins and I'm here with Jason Halberstadt. We're gonna spend about 30 minutes kind of boiling down some really cool stuff involving AI and how we use that in the travel industry. Jason, so happy to see you again.

Jason Halberstadt (00:20)
Likewise, Jason. Nice to be here.

Jason Elkins (00:22)
So

tell our listeners, what are we gonna be covering today?

Jason Halberstadt (00:26)
we're going to be covering the amazing new world of AI agents. I mean, there's been a lot of buzz and news and everything about AI agents. I think there's a lot of confusion about what they exactly are and how they can help travel companies to automate processes, communications. And so yeah, we're going to dive deep into this brave new world.

Jason Elkins (00:53)
Very cool, and as we do that, Jason, I'm gonna be trying to ask those questions. Any of those questions that pop into my head that I'm sure our listeners might be wondering as well. And let's give them a little teaser. By the end of this conversation, what are we gonna be talking about? Why should somebody stick around and listen to this call?

Jason Halberstadt (01:12)
Well, hopefully by the end, we'll kind of talk through what the technology is going to look like, you know, coming up in the next six months, year, year and a half, and give you, yes, hopefully some guidelines and a bit of a compass, hence the name of the podcast, of how to lead your company ⁓ in the direction where it's going to be positioned well to thrive in this new world.

Jason Elkins (01:38)
very very cool so let's let's jump right in what what are we what is AI agents chatbots what's where do we go with this conversation where do start

Jason Halberstadt (01:50)
Yeah, well, I think a lot of people nowadays think of AI primarily being chat GPT or Gemini, where they go to a web page, they type in what they want, and then it gives back increasingly better answers ⁓ as the technology gets better. But we're going to talk about what happens when you take it further, beyond

AI just giving you a response and giving you, you know, say an email, how to write an email. We're going to talk about how you integrate it into your company, in the processes of your company, with the software of your company, and then how that will enable your AI to more than just generate a document for you, but it will actually enable you to, the AI will actually be able to do things autonomously. Yeah.

It will.

Jason Elkins (02:49)
So

what I envision there is, because I use, you know, chat GP and some of those tools quite a bit, and I'll go in and, okay, help me write a blog post, help me come up with ideas for titles for podcasts or whatever, right? ⁓

But when I stop typing, it stops working. It's just kind of a put in, I get something back out. What I hear you saying or kind of touching on is how do we let this AI stuff do stuff while we're out on the beach or we're sleeping or whatever and it's actually taking care of stuff. Is that kind of where we're going with that?

Jason Halberstadt (03:23)
Yeah, absolutely. That's absolutely it. Yeah. Yeah. And so I guess maybe we can run through a scenario, some examples, common examples of things that would happen in a travel company. ⁓ And so ⁓ say you have somebody come to your website and you have a chat bot on your website that can interact with your customers and say help them to plan their trip.

And so essentially in that scenario, you kind of have created almost like your own chat GPT for your own company. And so the AI can pop up and start asking questions, start interviewing the user, similar to how a human travel agent would interview where they want to go, who they're traveling with, where they want to get out their trip, et cetera.

And then your chatbot can actually, in many cases, access other systems. For example, access all the information that's on your website. So all of your trips. And say you can help match this person that wants to take a trip with a specific trip on your website. And then it'll ask for, say, their email, contact information. And then it will send that.

information to the sales team at your company. And so in that example, you know, we've actually got an AI that's kind of on 24 hours waiting for somebody to initiate the conversation with it. And it's then able to understand what the person needs. It kind of gets a context of what the person needs as the context of what your company can do.

products you have and it then can ⁓ help the person, make progress on planning their trip. then, yeah, when it sends the email to your sales team and inserts it into your CRS system, then it's actually doing something. So we've got there kind of a basic AI agent where it's

independent. It's on all the time. It kind of makes decisions by itself based upon what the user wants. And then it does something. It collects the email and it sends you the email to your sales staff.

Jason Elkins (06:02)
Okay,

so a few questions. One thing is I recognize that chat bots or what I'm not even sure is the right terminology. You can help us with that. But I remember five, six, seven years ago, you could go on a website, something maybe would pop up and it would ask you some questions and you would give it some information and it would ask some other questions. And frankly, most of them were horrible.

Jason Halberstadt (06:05)
Yeah, sure.

Yeah.

Jason Elkins (06:32)
And someone was explaining to me, well, that's because they're really trained, a trained response thing where the programmer try and guess what people might say and get as close to it as possible. And it might ask you like, ⁓ you know, I remember in real estate was an example. It's like, what area do you want to live in? And you'd say anywhere except for Tolleson.

And the response would come back and say, oh yeah, Tollison is a wonderful area. You know, that type of thing. And so anybody that's listening to this, especially to our operators, they've been around for a while, maybe closer to my age. And they're like, yeah, I've seen those things before. Those are horrible. That's embarrassing. I would never, ever want my customers to deal with that. So maybe you can tell us a little bit about what's different about those compared to the agents that we're using now.

Jason Halberstadt (07:01)
Yeah, yeah. ⁓

Yeah.

Yeah, that's a great example. And I think those are very analogous to the phone trees that we get. Press 1 for customer service. Press 2. And so now that's

Jason Elkins (07:32)
yeah. Our menu has changed,

so please make sure you listen to every item. Yes, I remember. ⁓

Jason Halberstadt (07:39)
That's right. Yes.

Absolutely. ⁓ they still exist. But that's a really, really good example of a non-agentic system that is made to almost seem agentic because, they're kind of defining what the conversation is going to be about at the beginning and then giving you options of different things to do. So, there's not really any...

thinking or any autonomy that's going on. It's just a strict algorithm. If they ask this, then do that. If they ask this other thing, then do this other thing. ⁓ And so, yeah, that we can qualify that on the scale of agentic is barely being agentic whatsoever because it's really not so, it's not making decisions. It's not independent necessarily in any way. It's just presenting options to the user.

Yeah, and so, ⁓ yeah, and then when we go further on the scale of being, you know, what is an agent, being agentic, yeah, so we start off at the very basic, you know, like the old style telephone menus, yeah, and the old style, yeah, chats as well. And then, yeah, we go up to the next level of, ⁓ like the chat,

chat bot situation we were talking about where things were itself, that is really, I guess, kind of an assistant level where it's able to be a thinking assistant to you, semi-autonomous, and could actually do things for you. And then kind of the next level of agentic is when an AI can actually coordinate and orchestrate.

multiple things at once. ⁓ For example, if somebody wrote in that they want to cancel their trip, it would go to the script of what you're supposed to do for a cancellation and follow the script. Maybe you have to go into your booking system or the AI would have to go into your booking system, mark this as a canceled trip, then inform the correct people. And so

Yeah, at that point, you have an AI that is, yeah, it's not only agentic itself, but it's also doing different things. It's coordinating with other systems.

Jason Elkins (10:21)
So it sounds like what I hear you saying is one level, maybe the level right below that one is you come to the website and that agent pretty much knows everything on the website and is able to kind of answer questions based on what's on the website. And they're kind of reading you the website, helping you narrow it down, whatever. And then the level that you were just mentioning is the more like, okay, I know everything on the website. And if you wanna...

Book a trip, I know the next steps to that as well and I can actually check space and I can come back and say, yes, we actually have space available in this departure. And is that right? Am I hearing that right? Okay. And the other thing I really wanna ask you about, cause I know the word script sometimes can trigger people. So you said that that agent, if somebody wants to cancel, the agent knows what the script is. But when I hear you say script and then you described it, it's...

Jason Halberstadt (10:51)
Yeah.

Yeah, yeah, that's it. That's exactly it.

Jason Elkins (11:18)
more of a procedure as opposed to a script because when I hear the word script I think of, thank you so much for contacting us today and it's very scripted but what you're really saying is a script in a sense of this is what you do. You need to go check the terms and conditions, apply them and actually notify the vendor and do all these different things. So by script you kind of mean a procedure, right?

Jason Halberstadt (11:21)
Yeah. ⁓

Yeah, and procedures, yeah, and processes. Yeah, they're not necessarily a word by word script. You could give it examples. Say, so if somebody wants to make a cancellation, then state our cancellation policy. And here it is.

Jason Elkins (11:45)
Okay.

You

know what came up though when you were describing that is I was imagining somebody you know because some of our listeners have teams that work closely together and I used to work in a team that was closely together where if you had a question you looked over your shoulder and you asked the person behind you and I'm imagining someone sitting there and they get a message or maybe a call and somebody says hey we need to cancel my trip.

And it's a newer person, they're like, my gosh, I'm not sure. Okay. Hold on. Let me, I need to talk to my boss. ⁓ they don't really have much of a script. If you want to call it that. And they put the person on a hold. say, Hey, this is so-and-so they're supposed to be leaving in three days. And they're calling to cancel their trip. And what do we do? And, and their colleagues like, well, I think we have to do this. I think we need to do that. Well, the boss just stepped away from the desk. So we're not really sure. So they get back on the phone and say, you know what? I'm not sure. I got to check with.

my boss I'm not sure what we got to do and can I call you back in an hour and so let's compare and contrast that which I think anybody listening to this can relate to and what you're talking about having an agent that can actually do that

Jason Halberstadt (13:02)
Yeah.

Yeah, yeah, really actually AI agents, you actually train them in a way that's almost similar to training a human, you know. You need to give them training documents. So here's our standard operating procedures. And here's what we do if there's a cancellation. The big difference is that people sometimes forget. People don't read them first.

Jason Elkins (13:31)
that they read it? Because I know we've got employees working on our teams. You say, here's

our standard operating procedures. Make sure you read that in the first week you're here, right? OK, good.

Jason Halberstadt (13:43)
Right.

Yeah, yeah. First problem, they don't read them. Second problem, they don't remember them. then, yeah, then that's, yeah, the AI very rarely forgets. It is possible. You there's the famous hallucinations out there. It's not, you 100%, but it is probably 99 % these days, especially, you know, if you're working within...

Jason Elkins (13:46)
You

Let's talk about, think we need to touch

on hallucination. When you say that, some of our listeners aren't sure what you just said.

Jason Halberstadt (14:16)
Yeah, okay. Yeah, so, Hallucinations, like, especially in the early days when ChatGPD just came out, you could ask it to, I don't know, write a blog post for my website. And sometimes it would say, make up things. Or essentially, really what it does is it kind of tells you what you want to hear.

rather than something that is actually accurate and precise. And then they call that a hallucination. Hallucination, it makes up something. And so, yeah, that was much more common, much more of a problem in the early days of AI, and it's continually improving. And then actually, the way to take kind of a quantum leap to eliminate hallucinations is to make sure that the AI is relying on your

source of knowledge, your documents that you upload, your database, rather than having to go to the large language model where it's going to all the information on the internet and trying to come up with the answer that you want to hear. ⁓ so, yeah, hallucinations although not completely eliminated, they're greatly reduced now to say versus a year or two ago.

Jason Elkins (15:35)
Well, and I suspect that, you know, maybe I don't want to say the typical listener, but someone like myself, when I first got started, you know, you just putting in stuff, you don't really know how to avoid hallucinations. You don't know how to tell the system, please don't make stuff up. So anybody that first sat down when chat GPT came out and started typing stuff in, was like, this is fun. This is pretty cool. But it didn't take long before you got some weird kind of response. You're like, well, okay, I know that's not right.

Jason Halberstadt (15:50)
Yeah.

Jason Elkins (16:05)
So now I question it and I doubt it or they've seen social media posts or stuff that are just ridiculous. ⁓ So people might be walking around with this idea, you can't really trust it. But what I hear you saying is by really training it, especially on your content, because if you as a tour operator feel like, my website is good. The information on my website is good.

Jason Halberstadt (16:06)
Yeah.

Mm-hmm.

Jason Elkins (16:30)
And I know that the internet as a whole has a lot of crummy information, but if I can restrict the responses to the information I have on my website to some certain level, then I've got some sense of control over what my chat bot's gonna say. So I don't, if I do bicycle tours in Italy, I don't have to hopefully worry too much about my chat bot saying, ⁓ know, January is a wonderful time to go bicycle riding in Northern Italy.

Right? Because there's nowhere on my website doesn't say that. ⁓ It might be somewhere else on the internet. Is that kind of what you're saying when you're talking about kind of limiting it?

Jason Halberstadt (17:08)
In a way, yeah,

if you say I want to go bicycling in northern Italy in January, yeah, it may say it's a great time because it's kind of giving you what you want. And really, the key to this is actually there's one super important word, even when you're working in CHGP or whatever with AI, context. So let's go back to the scenario where you have a new salesperson

Jason Elkins (17:21)
huh, yeah.

Jason Halberstadt (17:37)
staffing you need to train them. When they start day one, they don't have the context to understand what you sell, all your policies and procedures, your standard operating procedures, which they don't read anyway. AI ⁓ is very much the same. It's like a new employee every time you open up a new conversation in ChatGPT. You have to give it context. You have to say, I am a salesperson.

or act as a salesperson for this company, use my catalog of products to offer to the person. Here are my standard operating procedures. Here's the way that you should initiate the conversation. And so when you give it all this information, the right context, then just like the new employee.

when after they have the context, they have their training, they can understand how to do their job correctly. Yeah. And then you have people making up many fewer things that you don't really know about if you give the eye the context that it needs.

Jason Elkins (18:54)
Well,

so then the other thing is, you I remember when I worked in a close space with my team, if I overheard someone on my team having a conversation with a client and I realized they were saying, yes, January is a great time to go to Northern Italy to go bicycle riding. There was an opportunity for me to, you know, give them some coaching after the call and we could discuss that and I could educate them a little bit more about what's going on in Northern Italy in July, in January. I don't know, I've never been there in January, but I'm guessing.

⁓ And so with the agents that people are using, how do you notice when the agent says something crazy without...

running into a client six months later and they say, yeah, I almost booked a trip with you once, but I was on your website and your agent said I should go to Northern Italy in January. I went and found someone else to work with. How do you capture those mistakes and then correct them so they don't keep happening over and over again? Because I think anybody that's listening to this that's thinking about it, putting an agent on their website is thinking about those mistakes.

Jason Halberstadt (19:52)
Mm-hmm.

Jason Elkins (20:03)
and it's never gonna get better, but the reality is people make mistakes too. But unfortunately, you train up a person and then they might leave and go somewhere else, but at least the chatbot, as you train them, they stick around, right? They don't quit because they're having a family crisis.

Jason Halberstadt (20:08)
Yeah, absolutely, absolutely.

Yeah. Yes.

Yeah. Yeah. And exactly. So also, if say, a person, if you overheard this person made a mistake, you, yeah, you'd usually have, a supervisor that's supervising to make sure, especially if they're new, to make sure they don't make mistakes. And you can do the exact same thing with AI agents. You create an agent that is a supervisor. And it's also called like LLM as a judge.

who is kind of like a quality control agent. So they can listen in on conversations or can ⁓ revise the conversations after they're completed. And it can point out any errors that were made, ⁓ any ways that could make the customer more satisfied. It could say, yeah, you need to offer more personalized options, for example.

or you need to ask the customer about their preferences before offering them a package so that you can match their needs and desires with what you have. so, yeah, so we actually have, and actually have had for about a year now, an AI that comes in, analyzes each of the chat conversations, and then gives this full feedback report.

like a quality control report, and gives feedback to the AI of how it can improve the next conversation.

Jason Elkins (21:51)
So now you take that and I remember a conversation I had with you in the past where you're like, we can actually have an AI do a secret shopper type of thing. So we can have an AI come into the website, have a conversation with the chat bot ⁓ or the agent, and then you've got a supervisor that reviews that conversation. And then the secret shopper reviews that conversation and you can tie that together so it can just continuously kind of get better with limited personal.

Jason Halberstadt (22:18)
Yeah.

Jason Elkins (22:20)
Direction is that right?

Jason Halberstadt (22:21)
Yeah,

yeah, yeah, yeah, yeah. The way we do this is we create yet another AI agent that has a persona, like an ILO customer profile. So we say, know, Betsy from Washington who likes to quilt flags, wants to go to wherever and give Betsy a travel goal. And then Betsy, the AI virtual Betsy.

goes and it interacts with our chat bot that's on the website. And so it's essentially simulating almost like a video game or like SimCity. It's simulating hundreds or even thousands of different conversations. And then you have the evaluator, the judge, the quality control AI come in and judge those conversations. And then you give a feedback and say, the AI.

In this conversation with this fake traveler, you could have handled the situation in a better way, and this is the best way to do it. so that gives the feedback and gets smarter. So you can really simulate thousands of conversations before your customers actually interact with the chatbot.

Jason Elkins (23:26)
Wow.

you've got some I presumably if the either the the judge or the secret chopper or anywhere in there come to a kind of like ⁓ yeah we're not really sure how to handle this because it's not in the information that we were given then I can say to maybe someone that someone on the team the owner of the company it can say hey we realize that

Jason Halberstadt (23:54)
Yeah.

Yes.

Jason Elkins (24:03)
people are potentially asking about this product and we don't really understand this product well enough to give them a good response and what that can do to trigger the tour operator. Well, maybe that's because I the way I've got that product described on the website is flawed, which means if a live customer went to my website did not engage with my chat bot, they're going to run into the same problem.

Jason Halberstadt (24:09)
Right?

Mm-hmm.

Jason Elkins (24:32)
So now what you've done is you've surfaced a problem in your marketing materials and your sales materials or in your operations that you wouldn't have otherwise surfaced until a customer got on the phone and said, hey, I'm on your website. I can't find the answer to this question, which I know anybody that's listening to this has had that. And they're like, ⁓ yeah, you're right. Our dates for that trip are five years old. Maybe we should fix that. And now you've got the chat bots.

Jason Halberstadt (24:36)
Yeah.

Jason Elkins (24:59)
Not just helping people come to your website, but they're helping you figure out areas of your operation that are lacking. that an accurate statement? That's cool. ⁓

Jason Halberstadt (25:06)
Absolutely. Yeah, yeah,

in the agentic world, that's called human in the loop. basically, any issue that it comes across that doesn't know how to handle it can escalate it to owner or a supervisor to human so that they are advised and can take care of

Jason Elkins (25:15)
Mm-hmm.

So if somebody, right,

so if a tour operator put an agent on their website, even if it never engaged with a person, never turned into booking, but was doing all this behind the scenes work helping you figure out gaps in your system and problems on your website, there's some benefit there even if it wasn't engaging with clients, right?

Jason Halberstadt (25:51)
Yeah, good point. I hadn't really thought about that, but yeah, it can...

Jason Elkins (25:54)
I mean, obviously, hopefully

they're engaging with clients and booking trips.

Jason Halberstadt (25:59)
But

yeah, you could uncover issues with your website, with your policies and stuff during the simulations. And they could be escalated and identify these problems before your customers discover them.

Jason Elkins (26:13)
Because I recognize that somebody might be listening thinking.

Well, I don't have a lot of stuff on my website or I already know that my website's out of date. So if I hear you, Jason, saying, yeah, the chat bot is great because it's gonna use the information on your website and you're thinking, well, the information on my website is not very good. So I better wait to take any action and tell my website's better and then I can go talk to someone about putting an agent on there. And really what's come out of this conversation is...

Maybe having an agent earlier in the process would help you identify those areas on your website that you need to focus on. Even if you never turned it on for the public, you could almost install the chat bot and get the secret shoppers and they'd come back with a report and basically saying, yep, your pricing is out of date on these 14 trips and these trips have poor descriptions and it almost tells you where to go and do it, right? And then, I don't know, that's I'm thinking.

Jason Halberstadt (27:07)
Yeah. ⁓

Absolutely, that's a great idea.

No, that's a great idea. Yeah, yeah. And so actually, we're coming out soon with a platform where you're able to actually design these different individual agents. And so yeah, the point is that, if you wanted, say, a website quality control agent where

Jason Elkins (27:15)
Yeah, so interesting.

All right.

Jason Halberstadt (27:37)
Yeah, you haven't posed as a person and point out all these problems. Yeah, absolutely. That's for you.

Jason Elkins (27:41)
Yeah, we might have to come back

and do another episode specifically about using these agents to really help you get to the next level. ⁓ And I don't we wanted to touch on, what does, you know, the future a year from now, what is the tour operator that's fully leveraging the things that are available now? What does that look like? Let's just plant the seed before we wrap up.

Jason Halberstadt (27:50)
Yeah.

Yeah, so I think what's going to happen, it seems pretty clear, is this trend towards becoming more and more agentic, being able to do more and more things autonomously or semi-autonomously. It's going to spread into other aspects of the customer journey, of your company's processes and workflows.

Really, I think it's going to be a process of kind of automating kind of step by step, know, climbing the ladder of automating more process of being more independent. And so I think it's going to be a process of people creating more and more agents, kind of like our quality control agent. Yeah, we'll say maybe I need a finance agent that, you know, analyzes

how much I charge for my tours and how much I pay my providers, et cetera. And so I think what you're going to see is people able to build these agents that connect with their internal systems, with their accounting system, with their booking system, et cetera, and are actually able to do more and more roles and assist with more and more processes and eventually be able to do.

almost all of those processes with less and less human supervision.

Jason Elkins (29:30)
So what does this mean? What does this

mean for the person that's listening to this that has a tour operation business or they're an advisor? Does this mean fewer people on their team, fewer need for people, or does it mean they just get to go do the things they really get excited about?

Jason Halberstadt (29:46)
Yeah, think long term, yeah, it could mean fewer people on your team. But I really encourage people to think about how you can have AI agents kind of take over some more of the busy work or places where people, where AI can do a good job, a better job than people. And ⁓ then just get started.

and start off with StratGPT. ⁓

Jason Elkins (30:16)
Yeah and if you don't let's discuss what happens

for these operators that don't because there's people out there that just like oh I'm gonna resist the AI thing and I like it I saw some bad articles that used AI and da da da I'm not gonna do it so what's realistically and this is not me and you here pitching AI but what should people consider that you know if they don't take some sort of action

Jason Halberstadt (30:23)
Yeah.

Yeah. Yeah.

Yeah. So what we're hearing, have one client who call up and said, my competitors are now using AI to create their emails and their itineraries to respond to clients. ⁓ I want to start doing it too, or I risk ⁓ not making the sale. Because with AI, they're able to respond much quicker.

Jason Elkins (31:04)
No.

Jason Halberstadt (31:09)
and using these advanced techniques that we're talking about, actually, in most cases, can respond much better in a more personalized way than a human agent. I it's not going to be better than your top, top agent all the time every day. ⁓

Jason Elkins (31:24)
But if you're top agent, your

top agent's probably not available at three in the morning when someone, you know, across the country. Because I think we're getting to point where we're past the filling out the form on my website and we'll get back to you within 24 to 48 hours.

Jason Halberstadt (31:29)
There you go.

Jason Elkins (31:39)
type of thing, just people expect something totally different and we're going to come back and we're going to discuss that on another episode about the importance of timely connection and you know the person that connects first is usually the one that gets the deal. So Jason, you know, I know you and I have discussed keeping this kind of short and concise so everybody can just take these little bites every week. Is there anything absolutely that we need to that I forgot to ask you that you want to make sure we cover on this episode or should we come back next time?

Jason Halberstadt (31:39)
Yeah. Yeah.

No, I think we wrapped it up pretty well this time. Why don't you join us next time? And we'll continue.

Jason Elkins (32:11)
Very, very cool. We're going

to have a bunch of interesting conversations. Jason, I love your wisdom and I appreciate you letting me ask those questions because I'm learning a lot as we go as well. So thank you so much, everybody. We'll connect with you here on the next episode. Thank you so much for listening to the AI Compass.