Late checkouts, AI agents and the future of guest communication with Cole Rubin of Conduit

[00:00:00] Cole Rubin: They wouldn't even offer or allow the $40, $50 late

checkout upgrade because it would cost about $40 worth of human labor to

triage and figure out. Can we actually do that late checkout? We pretty much

lose money orchestrating and facilitating these checkouts. Let's not even offer

them. But now with AI, that process can entirely be automated.

[00:00:31] Matt Welle: Hi, everyone. Welcome back to another Matt Talks

Hospitality. My guest today is Cole Rubin. He is the co-founder and CEO of

Conduit. Cole is not your typical tech founder. Before starting Conduit, he was a

hospitality operator managing a $19,000,000 residential portfolio of around 200

properties across Airbnb and booking a Colony Lyft with the messages and the

missed calls and the inbox chaos, and that really led to Conduit today. I recently

found Cole on LinkedIn. He posted a really interesting post about hoteliers

wanting to connect Claude to their PMS, and that really got me into him and

figuring out what Conduit does. And I'd love to just talk about his AI platform

and how it can transform the lives of hoteliers in a modern kind of life today.

Thank you for joining me today.

[00:01:19] Cole Rubin: Thank you for having me. Excited to be here.

[00:01:21] Matt Welle: So, you were a hotelier from day one? Like, was it always

your dream to become a hotelier or, like, in the hospitality world, or did you just

roll into that?

[00:01:29] Cole Rubin: So, I grew up really interested in real estate and

construction and knew that that was something I wanted to do in my career.

And, you know, out of college, I was working for a private equity group buying

apartment buildings. And, you know, I love that job, but my passion was, like,

construction design. And on the weekends, I started just building spec homes

out in Joshua Tree, California. Back then, you could buy land for about $12,000,

$20,000 for an acre of land, and I was building these purpose-built homes for

Airbnb investors. So, kind of started doing that on the side, and then eventually

that took off. Quit my W-2 job, about a year out of college and just went all in on

developing homes for purpose built for vacation rentals and then built up a

management business alongside doing that where a lot of times when I would

sell them, people just wanted, like, a totally turnkey package where it'd be

furnished on Airbnb with the forks in the drawer, management in place, and they

could just close and have guests in there the day after they close on the house.

[00:02:35] Matt Welle: I love that. And then as you started managing Airbnbs,

I'm imagining in the beginning, it was just you and you're logging in to the

platform to respond to messages, etc. So, I'm assuming it was very manual until

it wasn't sustainable anymore.

[00:02:48] Cole Rubin: It was entirely manual in the early days. A lot of it just

run by myself, and then at a certain scale, I started bringing on a team,

outsourcing a lot of the communications, day-to-day stuff, overnight stuff,

overseas to a team, but you know, the team was great, but it's constant back

and forth, finding new people, filling gaps in their knowledge. And, you know,

when AI was really coming around, it was evidently, like, very clear to me that

this was gonna change the way businesses talk to their customers. And with

hospitality, like, the communications kind of being the entire part of the

business, I saw this as a massive opportunity and really dove headfirst in and

entirely stopped developing and just built AI now. Developing AI, not real estate

anymore.

[00:03:34] Matt Welle: Why didn't you just buy a platform that was on the

market? Was there a problem that you're like, no, I actually have a different idea

about the platform that actually is going to solve this problem?

[00:03:42] Cole Rubin: So, when we started Conduit, there were no other

hospitality AI agents in the space, really. There were a couple, like, very, very

people in the early stages, like, similar, maybe a little ahead of us when we

started, but there was no product that really, like, took off and resonated. And

the biggest thing we saw too was for the AI agent to be good, to be useful, it

needs access to all the channels, all the communication channels you talk to

your customer on. If you can't feed that communication to the agents, it's not

gonna be able to automate anything. So, really, the first year when we were

building was focused on just building the best unified inbox for the hospitality

space with, you know, the PMS integrated. So, when you're texting or

WhatsApping a guest, you have their booking details integrated to the OTAs.

You can natively message through platforms like Airbnb. So, once we felt like we

had built the best unified inbox in the space, we came around and started

working on the AI. We felt like we actually had the playing field for the AI agent

to go ahead and really focus on the AI technology. So, kind of the first mission

we set out to build was just getting everywhere we're talking to these guests, all

these people, into one platform. So, it's just one thing the AI agent can sit on top

of us.

[00:05:01] Matt Welle: Got it. So, it's a unified inbox, and then you started

building AI on top of it. So, what are the AI elements, like, what are you

automating, or what reporting tool functionality that you offer?

[00:05:10] Cole Rubin: So, the core of what we're doing is just automating the

guest communication. When a guest is looking to, you know, add a late

checkout, book a reservation, get directions, we can instantly respond to that

message most times better than a human would be able to. If there's ever a case

where the AI can't respond, it brings in a human in the loop through our unified

inbox. What else people are doing is now with, you know, like, the MCP

landscape and the tools that you can give to agents now, things are going

beyond just sending a response to the guest. It's actually automating the

workflow that happens after, when you need to add a late checkout for a guest.

Okay, we need to send them the payment link, we need to make sure they pay,

we need to update the checkout time, update the cleaning team that this

person's coming later, so on and so forth. So, things have really gone from just

being able to automate the back-and-forth communication to the actual back-

office work that happens after that communication transpires. And the

communication is really just turning into triggers for all these agents in the

business that do these back-office tasks.

[00:06:18] Matt Welle: And how does that work with different PMSes? So,

Mews has an open API, so it's quite easy to integrate with our platform. How

does that work with legacy PMSes? Is it possible to integrate these agents on

top of that?

[00:06:31] Cole Rubin: So, one of the requirements for the agents to work is

really just having, like, an accessible API. So, if it's not one that the PMS exposes

openly, we'll do direct integrations where maybe they, like, OAuth into our app,

or they just have some other way to integrate. We've experimented with, like,

browser automations for, like, very, very old systems that don't even have an

API, where it's just like, basically automating what a human would do over the

screen.

[00:06:58] Matt Welle: Like Cowork, like, Claude Cowork kind of situation, right?

[00:07:01] Cole Rubin: Exactly. It works. But, obviously, having an open API with

as much endpoints is gonna give the end customer the best solution.

[00:07:09] Matt Welle: Yeah. So, what's, like, a very specific example of a

problem that you've been able to solve that, honestly, you'd needed multiple

humans for before?

[00:07:17] Cole Rubin: So, one of the biggest requests is these late checkout

requests when the guest wants it, so front desk, or if it's Airbnb, they have to

look, okay, they have to look at the calendar. Is there actually the ability to

accommodate this? Okay, can the cleaning team actually go do it? So, what

would take, like, four steps of communicating after the guest requests that, the

AI can just automate entirely. And it's funny, like, there are customers too that

they wouldn't even offer or allow the $40, $50 late checkout upgrade because it

would cost about $40 worth of human labor to triage and figure out, can we

actually do that late checkout? So, it's just like, hey, like, we pretty much, like,

lose money orchestrating and facilitating these, like, checkouts, like, let's not

even offer them. But now with AI, that process can entirely be automated.

[00:08:09] Matt Welle: It is such a genuine challenge that I often see. And I

travel every week, and it's like when you ask the question, like, “Oh, can I have a

late checkout? I can see them, like, go to the back office, and then something,

they come back, like, 10 minutes later, and then they have negotiated. And

often, like, no, it's not possible. I'm like, but I'm willing to pay. How much is it?

And there's no system to tell them how much it is. So, this is a genuine, complex

workflow, I think, that you're solving for hotels.

[00:08:32] Cole Rubin: No. And, you know, when it's right there integrated to

maybe even like a Mews portal where they can just click there and, like,

sometimes the portal endpoints, like, can just push right back to us, and you

know, an action they request in the portal that would've just went back to a

human to orchestrate, the agent can orchestrate the behind the scenes of that.

[00:08:51] Matt Welle: Yeah. What's the messaging inboxes when it's written

text is easy. I've seen that happen, but voice is the thing that I've heard not great

feedback on in the industry. Where do you think voice is moving, and do you

have any voice capabilities as well?

[00:09:05] Cole Rubin: So, we have voice agents at Conduit. We do voice and

chat. So, really, like, any modality that your customers are talking on, whether it's

Instagram DMs, Facebook Messenger, that little widget that you'll embed into

your booking website, all of that comes into Conduit. The voice has actually

gotten really good these days. I'll review, like, calls weekly, and you know, most

of the time, it ends with the customer saying thank you to the AI because it did

what they called to. Maybe they knew it was an AI, but they resolved what they

needed, so who really cares? And then the way I think about, like, the voice

agents too, is there are two ways to do it. You can have it as offense or defense.

So, offense is, like, think of it as, like, glorified IVR. Instead of the guest hitting

like a press 1 for this, 2 for this, blah blah blah, the voice agent just answers,

“Hey, Matt. Looks like you're staying at Property 123. What's going on?” “Oh! I'm

just looking to, I need the house car. I wanna go to the mall.” “Oh, let me send a

note to the team, and they'll text you when the car is ready.” “Great.” Now,

human didn't need to go in the loop. Or maybe this call would've gone to

voicemail instead of because a human wasn't gonna get it. So, hey, let's just

have that call fall back to a voice agent now. They could figure out what the

customer needs, maybe resolve their issue end-to-end. But at the least, the

customer is talking to someone, the company is hearing what they need, and

then the agent could, you know, maybe collect what would have been a missed

call and send a recap to a Slack channel for the team to come in and action it.

[00:10:42] Matt Welle: Can it take actual bookings, which is a very complex

workflow usually, and then figuring out how to take the payments and send

confirmations?

[00:10:49] Cole Rubin: So, the payments, actually taking, like, a card over the

phone. Like, we have, you know, PII redaction where we could collect the

information. We don't have the ability to actually process it in real time over the

phone. That's something we're gonna be adding this year. But how the booking

workflow with the voice agent normally works is they'll call in, we'll block it on

the calendar, and then send them the link over text, WhatsApp, or email to

actually go check out and process the booking.

[00:11:16] Matt Welle: Nice. So, I found you on LinkedIn because of a social

media post that you wrote about asking hoteliers if they're interested in

connecting Claude to their PMS. Like, why do you think there was so much

interest in that?

[00:11:30] Cole Rubin: The Claude post is, I mean, they're literally, like, very en

vogue right now. So, I did see that as, like, I thought that'd be an opportunity

that would go pretty viral. And, like, I think people are seeing it with like, once

you connect Claude or ChatGPT or whatever coding agent AI tool you're using

to an external tool, like, out of the box, like, to have those connectors, like your

Gmail, your Google Calendar, your Notion. And when you start seeing go

beyond just, like, using the ChatGPT or Claude as just like a chat interface for

back and forth Q&A and have it actually connect to other tools and go automate

work, that's, like, the biggest, that's how I'd say, like, I really started using AI. You

unlock, like, new workflows, new productivity that you didn't know was possible,

and these PMSes don't have the native connectors to get it into Claude. So, like,

most people, they don't even know that this is possible. So, that guide was just a

simple way, where it's like, hey, like, your PMS has an API. We have

documentation on how the API works. We can feed both of these to the agents,

and it can automate a lot of the work that, you know, you would do by

controlling your PMS with the mouse. So, I mean, I think it was just, you know,

very en vogue right now, and people wanted to do this, but there was no easy

way that anyone had put out on how to yet.

[00:12:52] Matt Welle: Can anyone integrate Claude with their PMS? I think

that's the thing that people were wondering. Or should they wait for the PMS to

make that available?

[00:13:00] Cole Rubin: It's gonna be a better experience. If the PMS has it

natively, it'll just be easier to set up. Usually, an API versus an MCP. Like, they'll

make the MCP, like, purpose-built for this use case. But, I mean, if you have a use

case and you wanna connect your PMS before they have it natively, like, there's

no downside or risk to just, you know, connecting it right with the API.

[00:13:22] Matt Welle: I love it. No. It was just really good to see the

engagement because it makes me excited that hoteliers are clearly using

Claude, and clearly excited about integrating it with a PMS because, you know, a

lot of hoteliers are like, you know, I'll just build a PMS. But if you think about your

token spend, it's probably not worth building the core infrastructure, but it's

incredibly cool to do things on top of the PMS. And I was very happy to see the

excitement of hoteliers because I've always been worried whether they're not

catching up with what's happening in the world that we're experiencing. We're in

tech, so we see you know, I love Claude, and it's so exciting. But the thing that

really triggered me was like, oh, there is a real interest from hoteliers in this, and

that really is exciting.

[00:14:05] Cole Rubin: There was good demand. I'm curious. Is a native MCP for

Mews on the horizon?

[00:14:11] Matt Welle: I hope so. Like, definitely, we've moved everyone at Mews

onto Claude ourselves yesterday. So, we've, like, migrated 1,500 employees at

Mews onto Claude. We've given everyone significant token access because

that's the thing that surprised us. So, we're making it really come to life across

our organization. So, I think it's a natural next step that we'll actually add that as

well now that everyone sees, like, I really want to make sure that everyone at

Mews understands, like, no, no, the world has changed since January, when

Claude Cowork came out and Claude Code last year. The world has shifted very

rapidly, and I want everyone at Mews to understand that this is the new world,

and we've gotta figure this thing out fast. And we're creating lots of enablement

and training internally. There's not a human at Mews that doesn't have access to

it now and is expected to do something with it. So, I think it's just a matter of

time before we get there.

[00:15:03] Cole Rubin: I totally agree. How was that process, like, to move that

many people? Was everyone using ChatGPT before, and you migrated everyone

to Claude?

[00:15:12] Matt Welle: Yeah. So, like, the developers were using Cursor before,

and then we realized that Claude Code was significantly better, so we migrated

them. And the rest of the organization, we use Glean for enterprise, and it's very

good, but it's good for internal knowledge bases. And you had access to the

GPTs, but they were always a model later. So, it just wasn't catching up with

what our teams wanted. And then we just said, “Okay. We need a real platform

that's the best in the world”, and today, Claude is just the best in what we need

as a business. So, we retained Glean, but we're adding Claude on top of it. Now,

the next step will be all these apps that we're building, they're sitting locally

hosted, so we need to figure out how do we lift them into the cloud. So, we're

now looking at additional platforms like a Lovable so that we can lift things that

people build into the cloud and host it so that we can have other people use

those apps as well. And you know, this is what we do at Mews, but actually, I

think a lot of hotels will replicate the same model, especially if you're a larger

hotel. But even if you're a small hotel, there's so much to automate. Like, I always

talk to hoteliers and, like, can you give me your checklist? I was at a hotel

yesterday, and I said, “I'm sure your front desk has a checklist. Right?” And they

do. And they're just manually checking things. They go through reservations and

check that the segment is correctly set on the reservation. Like, why do you do

that? Why don't you just automate that? But it's because we've always done it.

And today, that all changes with AI, and we just need to take hoteliers on the

journey that maybe they look at you and me, and we've embraced it. And they're

like, yeah, but you're the tech companies. I'm like, no, no, no. It's all of us. In every

manual checklist, there should be no paper with things that you have to check

by hand because the computers can do some of that stuff and automate some

of those workflows, so that's why I think it's relevant to talk about it so openly.

[00:16:59] Cole Rubin: No. I mean, I totally agree with the sentiment there. And,

yeah, it's like the magic of, like, when you actually just, like, connect Claude or

whatever you're using to the actual tools that you use in your day-to-day, and it

goes to, it becomes literally, like, an expert in every tool that you use. Like, it

builds better models than I can in Excel. It's gonna do better research about

something that I'm gonna be able to. So, I mean, the people that are using it

today, it's just such an edge. Like, it's 1000% the future, and, like, you're not

gonna be replaced by AI entirely, but I would say you would be replaced by a

human that is really good with AI that can do your job, you know, four times

more effectively in half the time.

[00:17:47] Matt Welle: Yeah. Absolutely. Yeah. So, if we go back to Conduit, how

does it learn about the property? Because normally, what we used to do with

these kinds of bots was, like, you gave it a knowledge base, and then that was

kind of what it feeds off. Like, it knows your opening times, etc. Does it gradually

learn as well? Like, is it, like, a semantic layer that builds over time?

[00:18:12] Cole Rubin: So, it does. So, step one is feeding it all the context we

have available. So, like, maybe it's scraping your website. You have a Notion

database about the properties. Maybe it's even in Glean that you just connect

natively to Conduit. And then there's also another component of, like, tools that

the agent needs. So, the stuff that comes up that is not just, like, finite forever,

like nightly pricing, availability of certain team members. So, there's some stuff

that the agent needs to dynamically go find in time. So, to do that, we give it

access to tools and whatever tools it would need to go figure out everything

that your customers ask. You just connect it, give it the context, and then

everything it knows, it will automate, send the response, and close the

conversation. Anytime we're unsure of what to do, need a human in the loop, or

it's just something you specified, we don't want AI to handle this; it comes back

to you as an escalation. And then all these escalations, these things where the AI

couldn't respond, we have a dedicated section of our app where we, during

onboarding, we really push like, "Hey, you need to dedicate someone who every

week goes to the escalation center and teaches the AI what it didn't know.”

Because it's like a human. Like, a human gets a one-on-one every week, every

other week to fill the gaps in their knowledge and become a better employee, so

an AI is doing the majority of your conversation. It's not done with the training

after you kick it off with onboarding. There's always gonna be edge cases that

come up and come up, and you need to, you know, dedicate someone to

unblocking and making the AI smarter in every situation that it gets blocked. And

for our companies, we actually kind of coined a new term, a new role. We call

this a conversation engineer.

[00:20:02] Matt Welle: I love that. Like, and I think what you're saying is a really

critical thing to get right because so often we're like, right, so it's giving me an

output, and I'll just fix it because there was a mistake in there. But I'll just fix the

email that it's sending because it didn't get that one line right. And what you're

saying is saying, no, no, no, don't fix the output. Go back to the inputs and teach

it, and figure out why it got it wrong in the first place. Because if you fix it every

time, yeah, sure, you save 90% on the thing, but you still, every time, if you get

involved in the output, you have to go back to the inputs and constantly teach it.

You wanna never touch the output because you're constantly doing more and

more automation, and it does require upfront quite some, you know, specialist

that you hire that always goes back to the inputs and fixes the prompt, and

that's a skill that you need to either learn or you hire into the team. But once you

get that culture in, you start to really reap the benefits.

[00:20:57] Cole Rubin: Definitely. Yeah. Like, you just wanna go, you wanna fix

the underlying issue that we're pulling this information from to generate this

response. We don't wanna just edit the one message one time because you're

just solving it right there. And to your point of, like, knowing how to, like, prompt

and do all this, we've actually built an internal agent in Conduit. So, anytime you

wanna teach the AI, you just, in natural language, pretty much exactly, like, the

same way you would tell a human, “Hey, when that happens, do this or don't do

this.” You would just tell our internal agent, and that agent looks at what you're

saying and figures out, “Hey, is this just a piece of knowledge we need to add?”

“Is this a new skill we need to teach the agent?” “Is it a guardrail we need to

add?” There are different types of ways you can teach the agents. So, we don't

wanna have to make every user feel like they need to become, like, a prompt

engineer. So, you just tell our AI, like, you tell a human, and our AI figures out

how to, you know, achieve that outcome with the settings available.

[00:21:59] Matt Welle: When you're going to Claude and you're working on

Claude Cowork, and I use it to set up all kinds of automations. And I use it, like,

for simple things. Like, you know, when an employee is here at 3 months, I send

them a nice video saying, “Right, you're three months and, like, let me talk to you

about what your next phase looks like.” But it does get it wrong, and I always go

back to the right. Let me explain to you what you got wrong, agent, and, like,

you're constantly teaching it because the prompt is never perfect the first time

around, and you gotta figure this thing out. And that's why I would say any

hotelier out there that is listening that made it to this part of the episode is, like,

get Claude installed on your desktop and just build something. Like, really get

your hands dirty because I think a lot of them are like, I'm too busy or, you know,

I'm nervous about trying it. And I'll wait for a perfect training moment to come

up, and that moment may never come. But I so enjoy my weekends now. I've

moved entirely from stop watching Netflix, and I spend my time in Claude

because I find it much more interesting than whatever Netflix produces as

content nowadays.

[00:23:00] Cole Rubin: No. I've been saying, though, I almost have to leave my

laptop at the office overnight sometimes because I'll bring home my laptop and,

you know, one more prompt, one more prompt. Next thing you know, it's 2:30 in

the morning, and we have a new internal tool. But for the operators, too, that are

feeling intimidated or a little scared by Claude, by the time this episode comes

out, Conduit will have self-serve onboarding. So, in about 10 clicks, you could

connect to your PMS and then just connect that to Claude right there. And

there's, like, some native prompts in our app that you can just click, and you

know, within two or three minutes here, you could be experiencing your first,

like, big connected tool AI moment.

[00:23:45] Matt Welle: I was gonna ask you what's coming up, but it sounds like

you've already got a busy roadmap, but any other exciting things that you're

developing and releasing soon?

[00:23:53] Cole Rubin: I think the most requested thing has been the mobile

app, which will be coming out towards the end of Q2 here. So, you know, as we

expand to other customer profiles that we're talking to, not just guests, but also

now, like, vendors, cleaners, maybe asset managers, these other stakeholder

conversations we're bringing into the Conduit platform. While these people are

more on the run across different properties in the hotel, when the agent pings

them, they'll just be able to, on the phone, you know, loop the agent in, or they'll

get assigned a ticket that they need to go handle and realize. So, super excited

about the mobile app. MCP, we actually launched today, so this morning,

natively, so excited about that as well. And then, an entire redesign of our

platform is launching at the end of May as well.

[00:24:45] Matt Welle: Wow. Never a boring day, it sounds like. And, like,

congrats on the business. I'm really excited to see you growing, and you're

integrated with Mews, so they can come into the Mews marketplace, and they

can connect it live there and then, right?

[00:25:00] Cole Rubin: Mhmm. Yep.

[00:25:01] Matt Welle: Amazing.

[00:25:02] Cole Rubin: And then by the time this is out, we'll have self-serve

onboarding, actually. So, just be able to go through the flow, check out Conduit

without having to talk to anyone, and just a couple of clicks through the onboard

flow.

[00:25:15] Matt Welle: If you had to give a hotelier one kind of piece of advice

to start tomorrow when they come back into the office, what would it be?

[00:25:23] Cole Rubin: Mhmm. So, I would say for, like, a week, start just kind of,

like, jotting down, like, repetitive stuff you're doing. You know, don't try to, like,

build the workflow and figure it out right then. Just kind of, like, take notes for a

week, and then take all your notes and put your notes in Claude or ChatGPT and

say, “Hey, this is all the stuff I was doing this week.” Also, put in the notes what

apps you did that work in. “These are the apps I use. Build me a guide. I'm a

beginner to AI. Build me a guide on what I could do and how I could use AI.” And

so my advice to you is to use AI to teach you to use AI, in short.

[00:26:05] Matt Welle: That is excellent advice. I really love it. And it seems so

simple, but it's genuinely that easy. And all of you who are still waiting for a

trainer to come in, this is it. This is the training because the LLM knows how to

explain it to you if you don't understand it. And if you get stuck, you just say, “I

feel like I'm stuck. Can you make it simpler?” And I think that's really as easy as it

is to build or to do anything exciting right now in the world.

[00:26:31] Cole Rubin: Yeah. And, like, sometimes I'll be building something, and

it gives me, like, this list of instructions for the next steps. I mean, that's too long.

It's too confusing. Make it simpler and more concise. And it just keeps telling me

I don't understand how to do that. Tell me how to do it. And, you know, people

get stuck, and then they, like, kind of ask someone else to go to Google. Just

like, sit there and keep telling the AI. You don't get it. And eventually, like, it's on

your team. It wants you to figure it out. So, it'll make it very easy.

[00:26:58] Matt Welle: And it will be frustrating, right? You will get things

wrong, and the LLM just won't understand or just doesn't get it right. But you

just gotta sometimes start a new prompt and try again with a slightly different

prompt. But once you get the hang of it, it's wonderful.

[00:27:14] Cole Rubin: It's life-changing.

[00:27:15] Matt Welle: Thank you so much for joining me today. It sounds really

exciting, and I'm really excited to do more together in the future. Thank you.

[00:27:22] Cole Rubin: I'd love to do more. And, no, thank you for having me. It

was great to be here.