AI Agent Interoperability Sets the Stage for the Next Wave of Digital Workflows with Joe Davis

Hello and welcome to the latest edition of the Techstrong AI Leadership Insights video series. I'm your host, Mike Va. Today we're with Joe Davis's, executive Vice President for platform engineering and AI for ServiceNow.

And we're talking about AI agents because, well, who isn't these days? Joe, welcome to show. Thanks for having me.

Glad to be here. I feel like we are moving through some eras pretty rapidly, but at this time last year, we were all trying to figure out what to do with copilots, and then we discovered that well, maybe AI agents are a better approach. So we're moved into that phase, and now we're rapidly moving into this conversation about AI agent orchestration and how do we manage all these agents because, well, we might be on the cusp of too much of a good thing.

Joe, what's your perspective on how do we kind of approach all of this stuff? 'cause I think a lot of people are out there kind of wondering right now, they're like, well, I might have 10 AI agents per employee, and how do I manage all that? Yeah, I think it's a great, it's a great question, and I think you're, you're spot on.

In terms of what we're seeing in the, in the industry really, 2025 was the, uh, the breakthrough was models could start thinking, they could start reasoning, and with that, we saw really elevated, uh, improvements on accuracy. And so I think we really saw customers start to deploy AI faster than ever before as a result of the improvements in the models. And the natural kind of progression that you see is, okay, well, I've got all these different vendors I'm dealing with, could be Microsoft ServiceNow, Google OpenAI, and I'm using some amount of AI from many of these vendors.

How do I just keep the keep track of it all? I mean, it's a huge, it's a huge problem. And I, I would say it's not unique to ai.

I mean, we, we've had, you know, you've had the classic problem of you have a lot of software and a lot of vendors, and, uh, you had the classic questions of, well, what am I doing with it all? What value is it providing the organization? Who's using it?

Do I have too many licenses for this particular software product? I think those same questions that customers or companies have always had around their, their software, it applies to AI as well. Um, the way that you, um, interact with these AI though changed.

So the classic tools about counting users and classic applications, we, we kind of had to rebuild how we tracked and monitored and figured out, uh, value for ai. Um, the observability of AI is fundamentally a little bit different than the observability of classic software systems, if that makes sense. It does.

Um, how will this all play out then? Because if we all have agents and we're all assigning them different functions, but then the company's probably gonna have agents that are functioning kind of like digital employees that are responsible for some task, and I need to orchestrate what they're doing, and then I need to coordinate that with all my downstream partners and vendors and customers. So this AI agent landscape gets pretty complicated in a hurry.

So how do you think we're gonna be able to apply some, uh, reasoning to the actual workflows themselves that the AI agents are being assigned? This is one of the, uh, primary questions that our vendors and our customers were asking us, really with the advent of, of agent of AI earlier this year, which is, we need you to work with other vendors. We need your AI solutions to integrate and work with other software products and other AI systems that's super consistent.

If you talk to a company there, there's an implication or there's kind of an assumption that they're gonna have multiple AI vendors, they're gonna use AI for multiple systems, and these things have to work together. So we were really excited about MCP, um, which is this, you know, integration protocol is the way to think about it. It makes it easier to integrate end systems.

Um, I think in the, you know, if we rewind time a few years ago, every system you wanted to connect to in AI was a unique project or was a unique integration. And so there's some amount of overhead that was just a constant tax every time you wanted to add a system that really has been eliminated with, uh, MCP. Um, and so it's a, it's a unified protocol that basically if you can get the other system to speak it, it's almost free to add, uh, you know, n number of systems and plus one number of systems to your AI because you don't have to change the source system or the, or the core reasoning, uh, loop of the, of the system.

So I think MCP was a, it was a breakthrough. Um, and I think we are seeing a lot of adoption throughout 2025. Um, A two A was another protocol that, um, we were really excited about and it solves an additional problem.

So MCP solves the problem of what if you have AI on one end and a software system on the other, and you want them to, to talk to each other? A two A was a, is a protocol that's, um, been open sourced, uh, it's owned by the Linux Foundation. We were one of the founding members of this.

And it says, what if you have AI on each end? So it goes a step further. It's not just AI talking to a system, it's two ais communicating with each other and what are the protocols for that.

And very much like how we're having a discussion and it's very unstructured and we just talk to each other. It's kind of the thesis behind A two A AI is very, the protocol for A two A is very unstructured and it allows for multimodal, um, back and forth between these two AI systems. You can pass videos and images and text, but it's very much like how humans speak.

Um, you don't have to define the variables that you want to communicate with upfront. Um, so we're really excited by that. I think there's still more to do on the adoption side with a A two A, because again, it assumes that you have AI deployed and active on both ends.

Um, so what we're seeing typically is a customer may start with NCP and then evolve or graduate into A two A as they deploy more systems. And I think the other thing that we're seeing is there's more open source, uh, happening in the AI space than ever before. And, uh, companies and protocols are starting to, um, emerge dealing with a ui.

So I think in the past, AI was all text-based. You know, you go to, uh, an AI system and it's going to respond to you in a textual manner. And fundamentally, there's a lot of text on the internet, so there's a reason why there's, you know, it's really, really good at text and, and language understanding.

But now we're seeing patterns emerge where you want to have rich widgets, let's call them calendars and weather and, and maps, um, that the AI emitting or producing. And so there's an emerging standards also that go beyond just kind of the integration between systems, but also the experience itself, uh, in terms of like a two UI would be an example of a emerging protocol that tries to have an open source standard for, um, the experience of the AI itself. So how will the user experience evolve?

And I'm asking the question is, 'cause today we have a UI that we use to engage with an application, and often we find ourselves getting a little small case of whiplash as we move from one UI to the other along some workflow. Um, will there be kind of be one consistent UI that is some sort of agent that I am now invoking a lot of backend headless services through? I mean, how do you envision all this coming together?

It, it's obviously early days, but that seems to be, uh, the trend is that there's more standards emerging around the experience. The more systems that implement these, then it's, you can, you can disintermediate the system on the backend from the experience and you have a more consistent experience across any system that you're talking to on the backend. So I think that's gonna continue.

Um, and you're gonna see these experiences go well beyond the classic, there's a text or there's a table and you can have really rich interactive UIs that the AI is dynamically generating. So it's thinking and it's using the context and the problem is trying to solve with the user to dynamically generate a really rich contextual experience. I think that's gonna be a really big, uh, thing, uh, as we, uh, go forward.

Um, the other thing I'd say is we can't forget that, and this is a big investment that ServiceNow is making. AI is also able to do long running kind of more complicated tasks without human intervention. And so we're, we're really focused on the autonomous use cases as well.

And so there really kind of is no UI for that or we're, we're not sure what the UI is. Maybe it's a dashboard that you can use to monitor all these autonomous AI agents. Um, but there isn't a, a, a standard, I don't think that we've seen yet for how do we monitor or govern kind of all this autonomous AI that will be expanding, I think 2026 and beyond.

Mm-hmm. As those UIs are dynamically generated, um, how tailored will they be to the individual user preference versus the AI's ideal version of what that's gonna be? How much influence will the human have over the way those things are rendered versus how the AI renders it or will the AI learn over time?

My preferences, I think that's a great question. I mean, it's ultimately, it's up to the system how much control and variability you give to the user. Um, I think the area of learning and memory is still, um, in its nascent days.

Uh, I think, uh, the ai, you know, what it learns about the user, what it remembers and how that affects the experience, it's still very, very early days. I don't think there's many systems that are awesome at this yet. I think the, the way that many systems are built today is they have system prompts and they use user, uh, context and the experiences very consistent from user to user.

Um, so I think you're right. That's an emerging thing that we may see more of. Uh, it's very hard to test if everybody has a different experience.

How do you, how do you determine, you know, what the right answer is? So there's some telemetry that needs to, I think an observability that I think needs to, um, we're still in very early days on from an AI standpoint to, to make sure that the experience is working well for all the individual users as you learn about them and they have their own preferences. Or maybe this is where some old fashioned personal responsibility kicks in and I just gotta figure it out for myself and make it work for me.

Right? That's right. I mean, we can't forget the, the fact that there is, um, you can still favorite things.

You can still, uh, save things and recall them and use them in your own working style. I mean, that's not necessarily learning. It's the user saying, I want this thing and I want this thing to be here and the next time I come in, uh, and those I think are still gonna be around and people have a, we develop, uh, habits and so, and for efficiency.

And so I think that's gonna be something that you have to do no matter what. Basically. Do you think CP servers are gonna be everywhere then, or we kind of gonna have, you know, hundreds and thousands of these things drew and across the enterprise?

I mean, how will that play out? What we are seeing is customers, uh, that are kind of early in the adoption cycle, assume that there's an MCP server for ServiceNow. So for ServiceNow, we provided an MCP server that they can integrate with.

So if, if that plays out for every vendor that they talk to, then there would be an MCP server or many MCP servers in front of every company, basically. Mm-hmm. Uh, so there could be, there, there should be thousands, I guess.

Um, there are companies that are building an MCP gateway, like an like a one to rule them all when we'll see kind of, if that works out. I can see that being a, a pattern that might make it simpler for customers. But, um, really the only unique thing across vendors is the authentication protocols, um, and the personalization, I guess.

Um, besides that, it's really easy to hook up, you know, two MCP servers, 3, 4, 5, it's, it's the same protocol, uh, across any vendor. So it, it makes it really easy, uh, in comparison to kind of, again, every integration in the past was unique and was a bespoke project. The other thing I'm trying to figure out is, it's clear to me that there are AI agents that will be assigned a task, go do it, complete the task, and then they'll power themselves down.

But we'll also have agents that are more, uh, running in the background that are handling tasks. But, you know, I need to not let them run forever. 'cause then the cost will be outta control and the agents don't seem to know when to stop per se.

And at the same time, I'm not quite clear like how much resources are gonna be assigned to these long running agents and how do I kinda make, find a way that they only run when absolutely necessary. Yeah. Look, I think this is a great question, and I think part of this is a labor, uh, displacement or kind of a cost question, which is, you know, if you were paying humans to do the task, like a call support agent or a, an IT support agent, and there's some amount of that work that AI can handle, is it cheaper?

It, it needs to be, obviously. Um, and we have seen cases, it's, it's, it's not super uncommon, but we have seen cases where you can code a mistake into the ai, maybe you code a recursive loop and it just runs forever. And on a consumption model that can spike costs pretty quickly.

Um, this is why I think you need observability and telemetry. I mean, ServiceNow has something we call AI control tower. Um, but there's other solutions out there.

So I don't wanna get too specific and necessarily about ServiceNow, though we're really proud of what we've done. Um, but the, what we see is customers need to have real time telemetry, observability visibility into the AI so that you can catch these problems as they happen. And also you need to be the ability to say, what value is this use case driving so that you can figure out if it's, you know, the cost versus, um, the value that you're getting if it, if that makes sense business-wise.

So we've, we've been developing, um, the ai AI control tower to help our customers kind of make that, make that decision and make that more evident when, um, things go off the rails. Maybe there's a safety issue or there's a security issue. You can see that in real time there's an alert that happens.

Maybe someone's trying to hack into your system and, and prompt inject your ai, or maybe they're trying to get a refund they're not supposed to. So we can catch some of that stuff in real time. And we have alerts and, and a whole kind of dashboard that we give customers.

But then there's the value question, which is, you know, how many hours of savings is this particular agent providing you? And do you have a dollar value associated with each of those hours? And you can kind of measure this.

We're seeing both of those needs from, from, from our customer base at least. And to what degree will customers maybe build their own agents versus relying or customizing the agents that you provide? And it's also not clear to me just how customizable are the agents gonna be that are provided by vendors?

Yeah, I mean, for ServiceNow, our value proposition is that we, we already have pre-built agents and they, they can be applied to our, uh, workflows in our applications in an agentic fashion, in a turnkey scenario. So like it's rapid acceleration of time to value. That's one of our key value propositions.

Um, there are companies out there that they're like, you know, or technologies out there that are around a framework, and so they're around like developer control, uh, and they're just a framework and developers can kind of do whatever they want and it's highly customizable. That's not really our, our, our pitch though. You can do a lot of customization as well as building just whole net new agents.

Um, but again, when we talk to customers and they try to understand what the value ServiceNow is providing versus another vendor, it's, we have pre-built agents that you can just turn on and start to, to get to value. Um, so you can do things, for example, like if you're an it, uh, service desk, one of the things you do is you send surveys out to all your users and, um, maybe you, you send a survey out to a user who had an IT issue and you ask what was their experience like, users respond to these and maybe some have good, some have, okay, some have not so good experiences. Previously people had to look at the survey responses and figure out, well, what it all meant.

Or maybe you had a dashboard, you had to figure out what it all meant. You can have AI do that for you. They can research all the survey responses.

It can categorize cluster them, and it can, um, tell you where your hotspots are. It can correlate it with, you know, this user had a problem that here was actually what happened. So it can just save an enormous amount of time just for that one use case.

And we obviously have, you know, hundreds or thousands of use cases like that, that we've developed. And then last question. Let's say I had an AI agent that was optimized for selling and you had an AI agent that's optimized for procurement and buying.

Won't our AI agents meet somewhere in the ether and kind of just battle it out to the point where they cancel each other out? It's a good question. Uh, we've seen some examples on the internet of this, uh, you know, uh, of, uh, two ais maybe call talking to each other from a customer service standpoint where you might call in for an, an issue and then the then AI's handling the customer call on the other end.

Um, I don't, I don't know what to do about this yet. Uh, it's not a pattern that we're seeing a ton of, um, but I, I agree it's a scenario that we need to think through, but I don't think we've run into this at scale yet, so I don't Great answer on that one. Something to look forward to.

Yes, Joe, thanks for sharing your insights and your knowledge as always, and we'll talk to you again soon. Thanks again for having me, And thank you all for watching the latest episode of the Textron AI Leadership Insight series. You can find this episode and others on our website.

We invite you to check all those out. Until then, we'll see you next time.