প্রতিলিপি 🔗
WEBVTT
00:00:00.080 --> 00:00:05.759
Yeah, I mean the ability to have an MCP or like a CLI for a technology has now become like a litmus test for us, I think.
00:00:05.839 --> 00:00:14.640
Like if it could be the greatest solution on the planet, but if it doesn't have an MCP or a CLI, then it's gonna become a lot harder for us to embrace it and kind of roll it out across the company.
00:00:14.800 --> 00:00:21.920
Cause now you've been you've created a walled garden where I have to now leave our agent harnesses to go into your tool to do actions that I want.
00:00:22.079 --> 00:00:24.399
And that's just going to be a new friction point.
00:00:24.480 --> 00:00:32.000
When I want to try and instill in our engineers and our people here at VIXO, is like try and do your entire job without having to leave the harness.
00:00:32.240 --> 00:00:38.240
If you have a solution that does not work with that, then that's probably an inflection point or something for us to kind of discuss.
00:00:48.560 --> 00:00:49.439
Hi everyone.
00:00:49.600 --> 00:01:03.840
Welcome to Deep Learning with PolyAI, our continued podcast series where we introduce new technology, speak to business leaders, and encounter all the exciting developments and challenges in the world of AI.
00:01:04.000 --> 00:01:09.439
I'm Damien Sasso, group product manager at PolyAI, focused on our developer platform.
00:01:09.599 --> 00:01:33.519
And we've recently had a lot of exciting releases and technological developments that we've done in our agent studio developer platform this year, with everything from a REST API stack, where we've allowed customers to develop against our Poly AI agent studio functionality using APIs, an agent development kit for more development through command line interface tools.
00:01:33.680 --> 00:01:37.439
And recently we've launched what we're calling our builder MCP.
00:01:37.599 --> 00:01:55.760
And so on that note, I brought some of our contacts and leaders, one of our key clients, to talk a little bit about how they inspired us to build out our builder MCP and how they're encountering all the challenges that come with AI and new technologies like MCP.
00:01:55.920 --> 00:01:57.599
So just a quick introduction.
00:01:57.760 --> 00:02:10.319
Joining me is Derek Neighbors, CTO and Bobby Honeycutt VP of Information Technology at VIXO Facility Solutions, one of the largest facility solutions providers in the United States and a Poly AI client.
00:02:10.479 --> 00:02:12.000
Derek, Bobby, welcome to the show.
00:02:12.240 --> 00:02:13.280
Hello, thanks for having us.
00:02:13.520 --> 00:02:26.639
Derek, why don't you start just quickly by telling us a little bit about yourself, a little bit about VIXO, and then we'll start getting into some of the you know cool technological developments that you guys are facing and that we're working on at Poly AI?
00:02:27.360 --> 00:02:27.599
Sure.
00:02:27.680 --> 00:02:41.520
I've been uh in the software industry for 25 plus years, largely helping private equity or VC companies either get ready to IPO or basically turn turn their business.
00:02:41.759 --> 00:02:53.039
In the last two and a half years have really been focused on AI and AI transformation, originally mostly in the engineering space, but fanning out into the organizational space.
00:02:53.199 --> 00:03:08.240
I think one of the things that I've seen in the last two years, year and a half in particular, is AI is starting to impact a lot more than just engineering workflows, engineering processes really, you know, basically being pervasive with inside the organization.
00:03:08.479 --> 00:03:15.680
And so there needs to be kind of a technical bent or engineering bent to help the business transform into that.
00:03:15.759 --> 00:03:18.400
Uh, but it certainly isn't just about engineering and code.
00:03:18.479 --> 00:03:20.319
It's about a lot more than that.
00:03:20.479 --> 00:03:25.360
And so from Vixo's perspective, we've been around for 20 plus years.
00:03:25.520 --> 00:03:26.960
We're one of the top leaders.
00:03:27.039 --> 00:03:37.680
If you if you think of a retailer, a convenience store, a QSR, like a think of like a Starbucks, think of a uh Circle K, uh Quick Trip, you name it.
00:03:37.840 --> 00:03:43.120
We we're probably in doing their facilities or part of their facilities work, or we have in the past.
00:03:43.280 --> 00:03:50.560
Any kind of retailer, you think of a Macy's, you think of, you know, anything you see in a mall, big box store, grocery store.
00:03:50.639 --> 00:03:55.120
We're probably in there servicing the air conditioner, the refrigerator, their coffee machine.
00:03:55.439 --> 00:03:56.319
You name it.
00:03:56.639 --> 00:03:59.680
A pretty kind of blue-collar turn-to-wrench type of business.
00:03:59.759 --> 00:04:01.680
You wouldn't think super high-tech.
00:04:02.000 --> 00:04:21.600
But the reality is it's a lot of moving parts around when something breaks, what information is around that asset, how to find a person a service provider that's in proximity to be able to fix that, fix it efficiently, effectively, to be able to get the parts needed for it and to coordinate all of that work.
00:04:21.759 --> 00:04:22.800
There's a ton that goes on.
00:04:22.959 --> 00:04:35.040
And that's, you know, we've had technology that has helped with that some in the past, but there really hasn't been a forward-thinking look about how to leverage technology to really go to the next step.
00:04:35.199 --> 00:04:53.680
And, you know, in the last year and a half, I think it's become very apparent that assets, machines, facilities are built for agentic processes, meaning self-referential learning loops within the facilities management process really can make a fundamental difference.
00:04:53.920 --> 00:05:00.959
The thing that's difficult is because we're kind of in a blue-collar world, none of the systems surrounding us are very agentic by nature.
00:05:01.040 --> 00:05:10.399
Meaning if
Load full transcript
we're in the SaaS world, it's super easy to go find everybody, their brother says, go try out my new MCP, go grab my new REST API, go, you know, get my CLI.
00:05:10.720 --> 00:05:12.079
You can really have that conversation.
00:05:12.160 --> 00:05:15.519
It's really hard to have that conversation with Fortune 50s.
00:05:15.600 --> 00:05:23.519
It's really hard to have that conversation with a facilities management group that, you know, is is dealing with, you know, some hardware that's 30 years old, right?
00:05:23.600 --> 00:05:40.000
And so, you know, I was kind of brought in to say, how do we transform VIXO into a modern facilities organization and produce facilities OS, which is the agentic form of managing facilities assets and everything related to outcomes for a facility?
00:05:40.240 --> 00:05:40.879
No, that's great.
00:05:40.959 --> 00:05:54.240
And I think I think you hit it hit the nail on the head in how critical your business is to the operations of businesses that you know are important to the customers of all different types of industries and consumers across this country.
00:05:54.560 --> 00:06:14.399
Bobby, like just to throw it to you, like when you look at how critical like an operations business like Vixo is, what considerations did you have to go through when you start even getting to the point of rolling out like an AI solution, something like Poly AI or, you know, any urgentic processes across your firm?
00:06:14.480 --> 00:06:16.879
Like what considerations and planning had to go into that?
00:06:17.120 --> 00:06:17.279
Yeah.
00:06:17.360 --> 00:06:23.839
So I I think a lot of the industry has been having a really hard time trying to figure out where the ROI with these AI tools is at.
00:06:24.079 --> 00:06:26.959
ROI seems very straightforward in the engineering space.
00:06:27.040 --> 00:06:34.879
You know, you're dealing with code, you're building apps, but a lot of the industry was trying to roll this out to non-technical users and we're kind of having some struggles there.
00:06:35.040 --> 00:06:36.800
And that's what we're trying to solve for.
00:06:36.959 --> 00:06:42.319
How do we put agents in the hands of people that are not in engineering, that are not technical focused at all?
00:06:42.560 --> 00:06:47.360
You've got these incredible models with these incredible brains, but they have no hands, they have no tools.
00:06:47.519 --> 00:07:00.240
And that's where MCPs and these other skills kind of come into effect, where they allow the end users to work with the data and the systems that they want to be in on a daily basis, driving these agentic use cases.
00:07:00.480 --> 00:07:07.920
So that's kind of what I'm helping Derek out with right now is how do we roll out agents to every single individual in the company?
00:07:08.079 --> 00:07:20.319
How do we build agent skill sets and teach them how to use these harnesses to do the type of work that they're already doing, but at a much, much faster and more proficient level using AI agentic tools?
00:07:20.560 --> 00:07:21.120
That's great.
00:07:21.199 --> 00:07:21.439
Yeah.
00:07:21.519 --> 00:07:27.759
And I mean, it's such a critical consideration with the people that are involved in using any agent tools.
00:07:27.839 --> 00:07:35.439
So yeah, I mean, coming into this year, you know, we've seen so much technological development in just how AI tools get embraced.
00:07:35.600 --> 00:07:43.040
We spent a considerable time at PolyAI looking in how we can enable developers through things like agent development kits, which I touched on as we kicked off.
00:07:43.279 --> 00:07:49.360
But MCP has really been, you know, it initially launched as part of uh a protocol a few years ago.
00:07:49.519 --> 00:07:55.120
And it's it's this cool little like, you know, tech thing that that like, oh, is it going to really be embraced?
00:07:55.199 --> 00:07:57.519
Is it really, but it's really started to catch on.
00:07:57.759 --> 00:08:01.439
What does it mean to your organization in the practical sense?
00:08:01.600 --> 00:08:08.079
Because it's like, you know, something that we can always think of, and you touched on a little bit as being a technological protocol.
00:08:08.160 --> 00:08:10.879
But what does it mean to the people that work at VIXO?
00:08:11.040 --> 00:08:14.160
What does MCP like you know look like in the day-to-day?
00:08:14.319 --> 00:08:18.959
Mary, Derek, maybe you want to touch on that or or or Bobby, but I'll throw it to you to start.
00:08:19.279 --> 00:08:19.519
Yeah.
00:08:19.680 --> 00:08:28.240
I mean, I think at the end of the day, agents are fairly useless if they don't have headless access to everything that a human would have access to, right?
00:08:28.480 --> 00:08:38.960
And so the the bets that we made is that giving AI and agents to engineers is fantastic, but it only scales so fast and so far.
00:08:39.120 --> 00:08:44.240
And so the reality is we really needed to say, how do we give agents to the edge?
00:08:44.399 --> 00:08:49.600
How do we give it to the end user, down to the technician out in the field, everything in between?
00:08:49.759 --> 00:08:53.759
And what we really landed on is that there's some fantastic harnesses out there already.
00:08:53.840 --> 00:09:00.480
You know, you could think of codex, you could think of cursor, you you know, you think of Claude, all of those are harnesses in their own right.
00:09:00.639 --> 00:09:24.799
And we and we said, look, we don't really have the resources to go build our own harness like a block does, but we could certainly take one of these existing harnesses, we could do a lightweight company type of brain that sits inside of that, load a number of skill sets, uh, skills that can load into that and load all of the MCPs that give everybody access to all of the data that they normally have access to via UI, only with their personal agent.
00:09:24.960 --> 00:09:35.919
And then we can teach them how to write their own skills
and how to build their own processes and workflows with the skills as part of that to customize their agent and that harness to themselves.
00:09:36.159 --> 00:09:40.799
And immediately the first thing that comes comes in is somebody has a workflow.
00:09:41.039 --> 00:09:44.399
They have to be able to talk to the systems that they're dealing with every day, right?
00:09:44.639 --> 00:09:46.480
So for a poly AI is great examples.
00:09:46.639 --> 00:09:54.639
Hey, I've got a service request coming in here, I've got this happening, I've got a service provider, I need to reach out and call that service provider, I need to get some information, right?
00:09:54.799 --> 00:10:09.120
You know, currently that's you know, hey, I have or you know, before that was I have to go input something or go into a UI or I have to ask engineering to go create something that I can interface with then calls the API, right?
00:10:09.200 --> 00:10:10.720
And so now I'm I'm waiting.
00:10:10.799 --> 00:10:18.000
If I if I want new functionalities for my process or my workflow, I'm waiting on engineering, I'm waiting on their backlog, I'm waiting on product, right?
00:10:18.080 --> 00:10:21.679
And when we gave people the skills, they're able to now get in and do all of these things.
00:10:21.759 --> 00:10:27.360
They said, well, why can't I just have my agent initiate a call on my behalf, the poly?
00:10:27.440 --> 00:10:29.519
Why do I have to to tell you how to do it?
00:10:29.600 --> 00:10:34.399
And you've got to go do a bunch of stuff and write about, you know, stuff I don't understand and it takes six months.
00:10:34.480 --> 00:10:35.759
How come I can't just play with that?
00:10:36.000 --> 00:10:40.960
That's, I think, where we really approached and said, look, we really need to look at what does it look like to have MCPs?
00:10:41.039 --> 00:10:42.639
And I think you guys have been absolutely fantastic.
00:10:42.799 --> 00:10:47.360
We said, hey, we can take the APIs and roll our own MCP fairly quickly to get people access.
00:10:47.440 --> 00:10:51.840
I think you guys said, hey, we're already on the stack and we were able to do something very quickly.
00:10:52.000 --> 00:10:53.840
But that's really the game changer, right?
00:10:53.919 --> 00:11:06.639
Is it it really gives universal access to the users, to the data in the systems that they already have access to, only now you're giving it to their agent as well, right?
00:11:06.720 --> 00:11:10.080
You're you're making that data a first class citizen to the agent.
00:11:10.320 --> 00:11:12.960
When I say agent, I'm gonna say harness, I'm gonna say personalized agent.
00:11:13.039 --> 00:11:14.320
I don't mean like generic agent.
00:11:14.399 --> 00:11:23.200
I mean, you know, literally a personalized agent that is doing work on their behalf now has all of the same access that they have access to, right?
00:11:23.279 --> 00:11:34.879
And I think that's one of the things is as ugly as MCPs can be and as you know inefficient as they can be, the reality is they're very simple to connect and to get out to the edges.
00:11:35.039 --> 00:11:37.279
And the end user doesn't have to understand a lot.
00:11:37.360 --> 00:11:44.320
If I give somebody, hey, here's the SDK, here's the the API, go write some code to do this, like it's just not happening.
00:11:44.480 --> 00:11:55.039
If I say, hey, go create a skill or just talk to your agent and tell you you want to do things and you already have the MCP hooked up, stuff just happens and it just starts to work and the agent can figure a lot of the things out itself.
00:11:55.279 --> 00:12:01.120
I think that's one of the things we're really seeing with MCPs that is fantastic, is they do a ton of self-discovery, right?
00:12:01.200 --> 00:12:03.440
They'll try to do something, they say, I can't really do this.
00:12:03.519 --> 00:12:04.320
They look for a different thing.
00:12:04.399 --> 00:12:05.519
Hey, I've got a different way to do it.
00:12:05.600 --> 00:12:09.440
Maybe if I do it in a different order, I can kind of circumvent and do what you want to do, right?
00:12:09.600 --> 00:12:23.840
I mean, I I think they they kind of expose things in real time that a developer using an API would have to do a whole lot of work work through to do and have to understand a lot.
00:12:23.919 --> 00:12:26.080
And the end user is never gonna do that, right?
00:12:26.159 --> 00:12:34.720
And so it kind of abstracts that and lets the the actual AI or the agent decode a lot of what's in there to get the data that's expected.
00:12:35.039 --> 00:12:36.320
Yeah, that may makes a lot of sense.
00:12:36.399 --> 00:12:48.960
You you touched on something that was really interesting, which is the act of having like MD files and skill files on behalf of certain personas, certain, you know, people working on certain accounts.
00:12:49.200 --> 00:13:04.960
You know, Bobby, is that something that you've seen kind of this transition in the world to making onboarding new new people easier, making you know, tailoring, I think, technology to your entire workforce more streamlined.
00:13:05.120 --> 00:13:15.279
Is that is that something that this is solving, or is it is there other even other elements to it that that's making you know, sort of the MCPs enable enable your workforce?
00:13:15.519 --> 00:13:21.519
Yeah, and it I mean, the ability to have an MCP or like a CLI for a technology has now become like a litmus test for us, I think.
00:13:21.600 --> 00:13:30.399
Like if it could be the greatest solution on the planet, but if it doesn't have an MCP or a CLI, then it's gonna become a lot harder for us to embrace it and kind of roll it out across the company.
00:13:30.559 --> 00:13:37.679
Because now you've been you've created a walled garden where I have to now leave our agent harnesses to go into your tool to do actions that I want.
00:13:37.840 --> 00:13:40.159
And that's just going to be a new friction point.
00:13:40.240 -->
00:13:48.000
When I want to try and instill in our engineers and our people here at VIXO, is like try and do your entire job without having to leave the harness.
00:13:48.240 --> 00:13:54.159
If you have a solution that does not work with that, then that's probably an inflection point or something for us to kind of discuss.
00:13:54.399 --> 00:14:02.799
And to your other question regarding how do these markdown files or skills kind of propagate through the business and how do they help us bring onboard people?
00:14:03.039 --> 00:14:03.600
Absolutely.
00:14:03.759 --> 00:14:13.679
So at VIXO, we kind of have a skills hub that we've kind of built where everyone that has these harnesses are able to publish and build skills and then publish those to the skills hub.
00:14:13.840 --> 00:14:20.559
And then anyone at VIXO can go and pull down those skills, you know, fork them, riff on them, or just kind of use them as they're stated on the tin.
00:14:20.960 --> 00:14:28.720
And that's how I think you get these this kind of like groundswell activity of like, oh, I didn't know so-and-so in accounting had a skill.
00:14:28.879 --> 00:14:31.759
I kind of need, you know, 40, 50% of that because I'm doing something else.
00:14:31.840 --> 00:14:34.159
But over I'm over here in, you know, service part of management.
00:14:34.320 --> 00:14:38.879
How can they take that skill that has those built-in connectivities and tailor it to their own experiences?
00:14:39.120 --> 00:14:40.000
That's what we're trying to do.
00:14:40.080 --> 00:14:48.799
We're trying to help them figure out what is possible and have them share broadly so everyone can kind of be made aware of what's going on and what other people are doing with these same tools.
00:14:49.039 --> 00:14:50.559
Yeah, it's really, that's really interesting.
00:14:50.720 --> 00:15:00.320
I think one thing that I was really curious to get your take on, because we've started to think about this a little bit from the poly AI side of things, is is you know, you can build MCPs for a number of things.
00:15:00.399 --> 00:15:06.720
Like we have data analysis, we have customers that use our platform to analyze their call data and analyze their business trends.
00:15:06.799 --> 00:15:15.600
And then we have, you know, different roles at at companies like yourself that are actually configuring agents, constantly updating agents, want to be that human in the loop.
00:15:15.759 --> 00:15:20.720
And we've looked at MCP servers as a way to kind of almost ring fence those audiences.
00:15:20.799 --> 00:15:26.000
Like maybe we have a builder MCP for for certain roles that just want to build and tailor agents.
00:15:26.080 --> 00:15:31.840
And then we'd have an MCP server for people that want to analyze their call data, analyze their business trends.
00:15:32.000 --> 00:15:44.720
Do you kind of see that as kind of a an indirect use case for an MCP where you can almost lock down certain access by only giving access to certain MCP
servers that can reach certain capability?
00:15:44.799 --> 00:15:49.039
Or is that kind of too restrictive in a way for MCPs?
00:15:49.120 --> 00:15:50.000
What's your thoughts on that?
00:15:50.240 --> 00:15:51.840
For me, I don't want to say it's too restrictive.
00:15:51.919 --> 00:15:57.919
I think what we've kind of defaulted to is really kind of two modes, and our preference is a second.
00:15:58.240 --> 00:16:08.240
So the first mode is I'm gonna say, like kind of wide open, like, hey, you've got some kind of a key or a token or or similar, and it's giving unfettered access to that resource, right?
00:16:08.399 --> 00:16:09.519
There's a lot of danger in that.
00:16:09.600 --> 00:16:11.360
There's not a whole lot of governance, right?
00:16:11.440 --> 00:16:15.039
But it's the simplest, fastest way to get access to something, right?
00:16:15.279 --> 00:16:28.960
And then the second way is to basically defer all activity or access to the MCP to that of the user credentials of the person that created the token or access or similar, right?
00:16:29.039 --> 00:16:34.480
So most of those are happening through some form of OAuth cash credentials.
00:16:34.559 --> 00:16:37.440
So, you know, hey, I'm gonna hit, I'm gonna hit something like poly.
00:16:37.600 --> 00:16:42.080
I hit poly, it asks for my login, I log in, I authenticate, it saves that token.
00:16:42.240 --> 00:16:47.440
And now my agent that I created and bound up to that NCP has all of the same access I have.
00:16:47.519 --> 00:16:52.080
So if I have access to build, it's gonna give me, you know, all of the tools to build.
00:16:52.240 --> 00:16:56.960
If I've only got access to to read data or to get that, I'm only gonna get access to that.
00:16:57.039 --> 00:17:02.720
And so for us, like we have a ton of financial data, invoice data somewhere, and it's all scoped basically on the user.
00:17:02.799 --> 00:17:05.680
So if a user can do it, they can do it with NCP.
00:17:05.759 --> 00:17:08.160
If the user can't do it, they can't do it with NCP.
00:17:08.240 --> 00:17:14.400
If there's something scoped where they can only see certain sets of data, that's gonna be scoped to to how they their agent.
00:17:14.559 --> 00:17:24.319
So basically they're inheriting this, that their agent is inheriting the same security rules, policies, governance that is bestowed upon that actual user.
00:17:24.559 --> 00:17:26.079
Yeah, no, that that makes a lot of sense.
00:17:26.160 --> 00:17:28.720
Bobby, any anything else you want to you want to add there?
00:17:28.960 --> 00:17:42.799
Like AI governance in the world of MCP is certainly like a such a such a a thing for every organization to think about, you know, any other concerns that you know you could advise, you know, our listeners on, you know, that you flipped into.
00:17:43.039 --> 00:17:48.720
Yeah, I think we are being fairly progressive at VIXO in terms of like just nothing built beats speed at the moment.
00:17:48.880 --> 00:17:53.039
Let's try and build and just kind of push things and push people to try and use these tools.
00:17:53.279 --> 00:17:58.640
And especially in terms of how we're trying to build our own internal MCPs that connect to our homegrown systems.
00:17:58.880 --> 00:18:01.039
We're essentially building as we go.
00:18:01.200 --> 00:18:02.720
We kind of push an MCP out there.
00:18:02.880 --> 00:18:07.839
The users try and hammer it with what they're trying to do, saying, hey, I can't do XYZ action.
00:18:07.920 --> 00:18:08.240
Great.
00:18:08.400 --> 00:18:09.200
Let's talk about that.
00:18:09.279 --> 00:18:16.400
Is that a quick add to our MCP to kind of continue the development and kind of get out of your way to allow you to continue to build and learn?
00:18:16.559 --> 00:18:22.480
Um, so we are definitely very much on the let's just kind of see what happens and build fast.
00:18:22.960 --> 00:18:26.559
But I don't know if that's a good fit for everyone, but it's what we're uh adopting right now.
00:18:26.799 --> 00:18:28.319
That I mean, that's music aure.
00:18:28.640 --> 00:18:40.960
I mean, that's the same way we're thinking about things at times of poly AI because the ability to use this technology to enable organizations is so is so exciting that that kind of makes sense from an approach.
00:18:41.200 --> 00:18:45.119
You touched a little bit on some of the things you're building internally.
00:18:45.359 --> 00:18:54.079
Where do you see the role of like REST APIs going in the world of more and more agent use of MCPs?
00:18:54.160 --> 00:19:03.359
I mean, we talked a little bit about humans using their agents, but I mean, the beautiful thing about an MCP server is agents can use it that you've deployed like autonomous agents as well as humans.
00:19:03.440 --> 00:19:15.039
So where do you see sort of the role of REST APIs now in an organization from a development perspective with things like MCPs and agent development kits playing a role in continued updates and changes?
00:19:15.359 --> 00:19:19.119
I think for me, REST APIs are still the foundation, right?
00:19:19.279 --> 00:19:23.839
Like at the at the end of the day, the MCP still has to talk to something to interface.
00:19:24.000 --> 00:19:27.680
If you're going to do CLI, it has to talk to something to interface.
00:19:27.920 --> 00:19:36.079
I think REST has proven to be a pretty stable paraby for interacting with objects, right?
00:19:36.160 --> 00:19:43.039
So I think kind of like the the this is a noun, and you can create, read, update, or delete that particular noun.
00:19:43.200 --> 00:19:45.519
And the way that it's nice and clean, like that.
00:19:45.680 --> 00:19:54.240
I think when people have a nice clean REST API to extend from, you can tell that they have a nice clean MCP.
00:19:54.400 --> 00:20:05.920
They have a nice clean CLI because things are very segmented and very easy to deal with, opposed to more of a graph or a SOAP type of mentality that's highly customized.
00:20:06.079 --> 00:20:08.559
That's a lot more difficult to deal with, right?
00:20:08.640 --> 00:20:11.440
And it it really comes down to discoverability, right?
00:20:11.519 --> 00:20:26.319
If you have good REST nomenclature, you can get away, even with an agent, to just point to it and it will just try things because it knows the structure of a good REST URL and API, and it can do a lot without even knowing that there's documentation under the hood.
00:20:26.400 --> 00:20:38.559
So when you expose kind of that same kind of principles and thinking and you kind of think of things as primitives, and you start to then expose your primitives as resources and you start to to to do that and and cascade out.
00:20:38.720 --> 00:20:51.599
I think a good, a good bare bones REST API, this well thought out, makes it super fast for engineers to build MCPs and build CLIs to then take advantage of that, right?
00:20:51.680 --> 00:20:55.440
So I mean, I I think it's just, you know, I don't think they necessarily go away.
00:20:55.599 --> 00:20:57.680
I I think they still stay as a building block.
00:20:57.759 --> 00:21:02.160
I just think you're not going to see as many people consume them directly.
00:21:02.480 --> 00:21:03.119
It's a good point.
00:21:03.200 --> 00:21:20.160
I I think the auditability of it, what we've leaned into, at least as we've developed out an MCP server that we were, you know, working and and and partnering with you guys on, was the thought of auditability around APIs and what APIs the MCP server was using, at least as a starting point.
00:21:20.559 --> 00:21:26.880
Is that something that's really been important, kind of Bobby, as you guys have thought internally about internal tools?
00:21:27.119 --> 00:21:30.880
Because it's kind of a new technology and REST is not a new technology.
00:21:30.960 --> 00:21:37.920
So you kind of get to bridge that gap of like still maintaining sort of the API auditability behind an MCP.
00:21:38.079 --> 00:21:39.519
Is that something to think about?
00:21:39.839 --> 00:21:45.119
Yeah, I think what's what Derek kind of shared earlier, like MCPs maybe aren't aren't the most efficient here?
00:21:45.279 --> 00:21:52.319
They're having to load a bunch of context in, they're doing a lot of discoverability, what's there, but they allow the end users to kind of do things on their own.
00:21:52.400 --> 00:21:56.079
They don't need to physically know everything that's in that API or what it can do.
00:21:56.400 --> 00:21:59.519
They can use natural language to explore and try and build what they want.
00:22:00.000 --> 00:22:03.519
Don't think REST APIs kind of go away, that they're still a great foundational block.
00:22:03.759 --> 00:22:14.559
If you're trying to do anything at scale or anything that's really reproducible, REST APIs are definitely something that play a part absolutely, especially you want to start refining these processes and bringing token costs and all that kind of stuff out of it.
00:22:14.799 --> 00:22:18.559
It's just another way of honing a process.
00:22:18.720 --> 00:22:21.839
Once you kind of have the business, say, hey, this is what I want, this is what good looks like.
00:22:22.000 --> 00:22:22.160
Great.
00:22:22.319 --> 00:22:28.000
How do we take what good looks like and make it really, really proficient and efficient to run and cheap?
00:22:28.319 --> 00:22:43.039
So as you've kind of like really detailed some of the digital transformation that you've gone through at VISO, enabling your workforce via their agents, you know, giving access control to agents that human have humans had, building things internally.
00:22:43.200 --> 00:23:00.400
What's some stories that you have on some of the efficiency gains you've seen or some of the some anecdotes around it, or just data that you've been tracking to show how much more efficient or much more in touch your workforce is, you know, as you've enabled them via agents connected to MCP servers?
00:23:00.720 --> 00:23:04.640
Yeah, I think I think just a lot of anecdotes of kind of what's possible, right?
00:23:04.720 --> 00:23:14.240
So I mean, I think one of the things that people get hung up on probably wrongfully so, is looking at AI simply as efficiency, right?
00:23:14.400 --> 00:23:24.880
And and what I mean by that is, you know, we we saw this in spades that the we we used to even tease around, you know, AI means I forget we automate the infrastructure, right?
00:23:25.039 --> 00:23:37.759
Because, you know, when you look at when you look at so many of the processes that are out there, people's first inclination is I do A, I do B, I do C, I do D, and I run them in order, and I'm having to do a manual touch point in each one of these.
00:23:37.839 --> 00:23:40.799
And now I have AI and I'm gonna automate that, right?
00:23:40.880 --> 00:23:42.160
And let's use AI automate.
00:23:42.319 --> 00:23:50.960
And we look at it and go, that's like the worst use case of AI ever because it's super expensive to pull in context, like you're not doing anything special with it.
00:23:51.279 --> 00:23:59.279
You're you're literally paying a really expensive tool to pick stuff up from A, move it to B, pick it up from B, move it to C.
00:23:59.519 --> 00:24:04.480
You're almost just hitting an AI usage leaderboard at that point and not actually solving like the problem.
00:24:05.039 --> 00:24:06.799
And what it's moving is the same stuff.
00:24:06.880 --> 00:24:09.920
It's like, okay, the subject line always goes from here to here.
00:24:10.079 --> 00:24:12.720
It's like there's no there's no logic, there's no nuance.
00:24:12.799 --> 00:24:15.200
It literally is like, I want to map this data, right?
00:24:15.279 --> 00:24:17.759
And so we started to say, like, whoa, whoa, timeout.
00:24:17.839 --> 00:24:22.880
Like we need to start to separate what's automation versus what's AI, right?
00:24:22.960 --> 00:24:30.640
And and so what we've started to kind of say is AI is really
fantastic to do automation that you don't have to wait for engineering on to prove a point, right?
00:24:30.720 --> 00:24:40.799
So if you want to go A, B, C, D, E and automate that and not have to go get an engineering's backlog or have you know big product to get you to the top of the list or whatever, you can go do that, right?
00:24:40.880 --> 00:24:44.559
And you can start to show, like, hey, look, I saved a bunch of time by automating.
00:24:44.640 --> 00:24:46.559
So we'll give you a great example from us, right?
00:24:46.640 --> 00:24:48.160
Is you know, invoices, right?
00:24:48.240 --> 00:25:07.920
So invoices come into an email box, somebody from accounting was kind of picking those up, looking at those in a folder, dragging them from one folder to another, opening them up individually, copy and pasting them into an Excel spreadsheet, then loading that spreadsheet into another tool that then put them into invoice processing and then verifying that they processed out into the accounting system, right?
00:25:08.000 --> 00:25:15.839
So, like five touch points, a lot of like dragging, but no intelligence involved in literally mindless moving stuff from A to B.
00:25:15.920 --> 00:25:16.160
Right.
00:25:16.240 --> 00:25:16.720
Hey, great.
00:25:16.799 --> 00:25:18.160
Use AI to automate that.
00:25:18.240 --> 00:25:19.599
Now you're out of that business.
00:25:19.680 --> 00:25:21.279
You're you're not having to do that anymore.
00:25:21.359 --> 00:25:22.559
You got efficiency, right?
00:25:22.640 --> 00:25:31.519
What you used to have 300 invoices a day in the backlog or 3,000 invoices a day in the backlog, and you struggled to keep your head above water between four people.
00:25:31.680 --> 00:25:34.400
Now that's fully automated, cradle to grave.
00:25:34.480 --> 00:25:38.000
Nobody's spending any time on that other than the final reconciliation piece.
00:25:38.160 --> 00:25:39.359
That's fantastic, right?
00:25:39.599 --> 00:25:41.599
But it's not really the fantastic use of AI.
00:25:41.759 --> 00:25:45.200
And so we've started to say is do use AI to do those type automations.
00:25:45.359 --> 00:25:47.119
Let us know is where it's working going good.
00:25:47.200 --> 00:25:50.559
And what we can do as engineering now come back and say, okay, great.
00:25:50.640 --> 00:25:54.079
What does it look like us to take the token spend out of that, right?
00:25:54.240 --> 00:25:56.319
And start to then make it efficient, right?
00:25:56.400 --> 00:26:06.799
So so we gave you some productivity gains, but we didn't really get a whole lot of efficiency because the reality is your time and the time you were spending that might be pretty close to what the token spend is, right?
00:26:06.880 --> 00:26:16.880
So at the end of the day, it's like we just we we gave you more capacity, but we exchange that for a token spin that then may have been similar to the cost of you involved.
00:26:17.039 --> 00:26:23.599
But what it does is it allows us now as engineering to come out and say, like, hey, could we just do this all through AP back to REST APIs?
00:26:23.759 --> 00:26:30.559
Can we just do this all via REST APIs and some lightweight glue and basically pull that out of the system and get the token spin back?
00:26:30.720 --> 00:26:32.079
Absolutely we can, right?
00:26:32.400 --> 00:26:41.119
And so where we're starting to see that the real like power anecdotes of things are the things that really didn't scale well as humans, right?
00:26:41.200 --> 00:26:48.960
There's some nuance involved in what's happening to where a typical automation couldn't occur.
00:26:49.200 --> 00:26:54.160
I'll give you an example is you know, we we basically didn't get a service request in.
00:26:54.240 --> 00:26:56.240
That's that's a work order, whatever you want to call it.
00:26:56.319 --> 00:26:58.319
And my ice machine is dead, right?
00:26:58.559 --> 00:26:59.519
Ice not dispensing.
00:26:59.599 --> 00:27:00.640
I'm circle K.
00:27:00.799 --> 00:27:09.279
I have uh the drink cooler, it's you know, 108 outside, and nobody can get ice out of the ice machine and my water fountain's down, I need help.
00:27:09.440 --> 00:27:11.359
Okay, and we're gonna dispatch somebody.
00:27:11.680 --> 00:27:19.920
So we've got some automations already built in to go determine who the best person dispatches for you know, the best provider to dispatch for that, everything about that.
00:27:20.400 --> 00:27:22.160
But the reality is there's so much more to that.
00:27:22.240 --> 00:27:23.759
There's the facility that you're in, right?
00:27:23.839 --> 00:27:28.799
So I'm in the circle K, I'm at this particular location, and I've got all sorts of data about that.
00:27:29.039 --> 00:27:34.400
Every service call that I've ever done in that circle K, I have information about that.
00:27:34.640 --> 00:27:52.799
Every asset, so the air conditioner, the coffee maker, the ice machine, the beverage machine, the walk-in freezer, the POS system, all of those are assets that I have a bunch of information on from either a telemetry perspective or from a service perspective of what's happened in the past about those machines.
00:27:53.279 --> 00:27:56.400
Right now, an SR comes in, I'm just servicing that SR.
00:27:56.480 --> 00:28:02.559
I'm looking at that asset, that work request in that moment in time, and everything else is blind to me.
00:28:02.720 --> 00:28:02.880
Right.
00:28:02.960 --> 00:28:17.200
So we're able to turn on that the service request comes in, and I'm able to look at all of the history, let's say in the last 90 days around that facility, around all the assets in that facility, and around all of the service providers that have touched something in that facility.
00:28:17.440 --> 00:28:22.480
And I'm able to create an intelligence packet about how that should impact dispatch.
00:28:22.880 --> 00:28:25.599
Give you a great example with an app that ice maker.
00:28:25.680 --> 00:28:27.519
It's broken, it's not a
beverage machine.
00:28:27.839 --> 00:28:32.000
We auto route and it says route out somebody to fix the beverage machine.
00:28:32.240 --> 00:28:32.400
Okay.
00:28:32.720 --> 00:28:34.000
That's what happens by default.
00:28:34.079 --> 00:28:35.519
That's what our system does by default.
00:28:35.759 --> 00:28:38.079
This agent picks it up and says, whoa, whoa, time out.
00:28:38.240 --> 00:28:41.839
You've had two toilets clogged in this facility in the last 60 days.
00:28:42.000 --> 00:28:45.519
This is the second time you've gone out with a beverage to a beverage machine.
00:28:45.759 --> 00:28:48.160
I don't think that this is actually a beverage machine problem.
00:28:48.319 --> 00:28:51.200
We think you've got a clogged somewhere in plumbing down the line.
00:28:51.440 --> 00:28:53.519
You need to make sure you don't send out a beverage person.
00:28:53.599 --> 00:28:55.279
Instead, dispatch out a plumber.
00:28:55.440 --> 00:28:58.960
Make sure that that plumber has a camera on their truck.
00:28:59.119 --> 00:29:02.079
They've got a water jet on their truck, and there's two people on the truck.
00:29:02.160 --> 00:29:07.119
So if they have to water jet it, they have two people to do the water jet and they don't have to come back for a second call.
00:29:07.359 --> 00:29:08.319
Turns out they show up.
00:29:08.480 --> 00:29:10.559
Sure enough, 20 feet down the line.
00:29:10.799 --> 00:29:13.279
There's a problem, they have to water jet it.
00:29:13.440 --> 00:29:17.039
In the past, that would have been we sent out a beverage guy, he clears it, says it's all good.
00:29:17.200 --> 00:29:19.519
Five days later, we get another call, but it's stuck again.
00:29:19.680 --> 00:29:21.119
We then send out a plumber.
00:29:21.279 --> 00:29:22.319
Plumber's gonna camera it.
00:29:22.400 --> 00:29:27.119
They say, Yeah, there's something pointing down there, but I don't have a water jet on me, or I have a water jet on me, but I don't have another guy.
00:29:27.359 --> 00:29:29.519
We're gonna have to come back in two days, right?
00:29:29.680 --> 00:29:37.279
The the meantime, the actual moneymaker for for this Circle K is a beverage machine that's no longer dispensing ice, right?
00:29:37.359 --> 00:29:42.640
So then that's kind of a real world, like nuanced way that we can look at that.
00:29:42.720 --> 00:29:57.039
And underneath the hood, there's all sorts of MCPs involved within that, looking at all of our systems, looking at our invoice system, looking at our asset system, looking at our you know, prior work order, looking at our facility, looking at the customer, looking at the service provider, and kind of gluing that all together.
00:29:57.200 --> 00:30:02.319
That whole system was not written by an engineer, that was written by an operations person with skills.
00:30:02.559 --> 00:30:45.839
That story is an amazing story in a sense that like that use case is amazing because it really touches on, I mean, the value of like a solution like a
pol, like what we offer with Polyai, where you're using an agent, a voice agent, chat agent, like a customer service focused agent that you can now connect to a number of systems to think through intelligently in a way that just a human being answering a phone would not necessarily be able to connect all these dots as quickly as agents can as to what's truly the issue at a client's site, not, you know, just a beverage machine, but you have a plumbing issue that's actually going to cause three additional service calls as opposed to being hit, you know, in a single, in a single service call.
00:30:46.079 --> 00:31:02.400
Something that I wanted to ask both of you as we kind of put a bow on this conversation was, you know, you're you're in an industry that to the average person might be perceived as not like, you know, as a blue-collar service industry, but you're thinking about things in such a technically advanced way.
00:31:02.640 --> 00:31:09.839
We're very excited that you're a client of ours and excited to continue to work with you on things like what we've done with our polyi MCP server.
00:31:10.160 --> 00:31:28.880
What should CX leaders elsewhere in other industries be thinking about over the next 12 to 18 months as they think about either, you know, their industry, if they're in a service industry in another, because in my eyes, you guys are at the forefront of tech of thinking about the right way to implement AI technology.
00:31:29.119 --> 00:31:32.799
You know, Bobby first and then yeah, Derek, just would love to hear your thoughts on that.
00:31:33.039 --> 00:31:38.319
Yeah, for kind of calls back to what I stated earlier, like MCP and CLI availability is now the litmus test.
00:31:38.480 --> 00:31:53.440
If there are key platforms that we have at this business that don't have those connectivities, that is probably the step one of me either looking to solution that, trying to build our own MCP, or two, how do I pivot to either a custom solution or a new solution that has those capabilities?
00:31:53.759 --> 00:32:05.279
I cannot condone or tolerate having those types of really key systems that don't have connectivity with our agents because it's just going to put blockers or friction points for trying to roll this out in the way that we want to.
00:32:05.519 --> 00:32:08.960
So look at your systems, look if they have MCP access.
00:32:09.039 --> 00:32:14.640
If you don't, how do you engage with your vendors to see if that it's on their roadmap the same way that we engage with Poly AI?
00:32:14.799 --> 00:32:16.319
Said, hey, do you guys have an MCP?
00:32:16.480 --> 00:32:17.440
At the time you did not.
00:32:17.519 --> 00:32:22.480
We offered to kind of write our own, but we were wanted to use a first class one from you guys, and you guys built one.
00:32:22.640 --> 00:32:24.000
Like that's a great partnership to have.
00:32:24.079 --> 00:32:30.079
So I would, if you don't currently have MCP access for those key systems, reach out to
that vendor to see if if that's on the roadmap.
00:32:30.160 --> 00:32:32.559
And if not, maybe adjust your thinking from there.
00:32:32.880 --> 00:32:33.599
Derek, yeah.
00:32:33.920 --> 00:32:39.440
I think for me, you know, speak direct to the the CX leaders out there, right?
00:32:39.519 --> 00:32:44.480
And it's like no matter how fast you think you're going, you're not going fast enough.
00:32:44.720 --> 00:32:47.200
I can't express this enough to my peers.
00:32:47.759 --> 00:32:55.039
This technology is moving faster than anything we've seen ever by a factor of 10 or 15.
00:32:55.359 --> 00:32:58.000
And you can kind of look at it in one of two ways.
00:32:58.319 --> 00:33:00.720
One way to look at it is it's going too fast.
00:33:00.960 --> 00:33:02.240
I'm too afraid of it.
00:33:02.400 --> 00:33:08.960
I'm just gonna wait and until it shakes out and and there's a right answer, and I'll adopt that right answer.
00:33:09.200 --> 00:33:16.720
And and a lot of times in the past, that's been a pre-reasonable way to look at new emerging technology.
00:33:17.039 --> 00:33:25.839
Unfortunately, in this wave, I think those that are not riding with the wave aren't gonna have a place anymore.
00:33:26.000 --> 00:33:26.240
Right.
00:33:26.319 --> 00:33:34.160
I mean, that this is the the the kind of the separation that it's putting between companies is so gap-oriented.
00:33:34.319 --> 00:33:36.640
I don't know if that can be closed, right?
00:33:36.799 --> 00:33:45.200
And so I think the the biggest thing that happens in organizations that I see is it becomes a governance hell, right?
00:33:45.359 --> 00:33:49.920
Everybody is afraid around like, well, well, who do we give access to what?
00:33:50.000 --> 00:33:50.720
And how do we do this?
00:33:50.799 --> 00:33:51.680
And how do we control?
00:33:51.839 --> 00:33:58.720
How do we, and I'm just gonna say if that's dominating the conversations in the room you're in, you're not gonna make it.
00:33:58.960 --> 00:34:00.640
Um that's not to say governance is important.
00:34:00.799 --> 00:34:04.880
That's not to say that that they shouldn't be having those conversations.
00:34:05.039 --> 00:34:07.519
It's if it's dominating the conversation.
00:34:07.680 --> 00:34:16.079
Like if that is if that is the thing that is where the focus is right now instead of on learning, it's gonna be really rough.
00:34:16.239 --> 00:34:25.440
And part of that is I think for the first time, there there are executives has to understand how things work at a deep level.
00:34:25.599 --> 00:34:30.079
There is no more manager or leader, and somebody's an individual contributor.
00:34:30.239 --> 00:34:38.559
We're seeing CEOs, CTOs of the top companies in the world right now be hands-on, being individual contributors to how these things work.
00:34:38.719 --> 00:34:49.599
And I'm gonna say, as a CX leader, if you're not in getting your hands dirty with this technology, you're not gonna understand where the opportunities are for your organization and for your industry and your market.
00:34:49.840 --> 00:34:56.400
And the window to take advantage of those is not gonna be really, really strong without getting lost or getting behind.
00:34:56.639 --> 00:34:58.400
That's that's great, that's great insight.
00:34:58.559 --> 00:35:09.440
And I mean, like from our perspective, you know, we've really enjoyed the partnership with you and your team on building out, you know, building out the MCP for our agent studio platform.
00:35:09.599 --> 00:35:22.719
We don't see it as just a release or a tool release, but it's kind of a shift in the way that our customers develop, you know, both you you guys integrating and and working with the polyai agent studio platform.
00:35:22.800 --> 00:35:32.559
But even since we started building out our initial MCP, we brought more functionality, rethought a lot of things that we know need to be included there, auditability that needs to be there.
00:35:32.639 --> 00:35:40.239
And we've been engaged, you know, we've seen increased interest as well since we put out documentation and and and started to talk more and more about it.
00:35:40.320 --> 00:35:52.880
And it's just things are moving so quickly in in all industries that as you touched on, Derek, it's important for even executives to be kind of in the weeds understanding how technology is is transforming their workforce.
00:35:53.039 --> 00:35:57.280
So, you know, it's been an absolute pleasure to have both of you on the show today.
00:35:57.519 --> 00:36:04.480
You know, Derek, Bobby from Vixo Facility Solutions, Derek CTO, Bobby, VP of Information Technology.
00:36:04.719 --> 00:36:10.400
For everyone that's listening, I encourage you to like, subscribe, you know, stay connected.
00:36:10.639 --> 00:36:14.800
And this is, you know, another edition of Deep Learning with Poly AI.
00:36:14.960 --> 00:36:16.079
Thanks again.