Learn how IT leaders are scaling AI agent deployment faster — with real-world strategies, practical frameworks, and a live demo of Tray’s Merlin Agent Builder.
Scaling AI agents across the enterprise takes more than experimentation—it demands the right foundation. In this Megacast, Tray’s CEO and CTO team up with Zuora’s CIO to share real-world strategies building scalable, production-ready agents while maintaining governance and control. See IT leaders are evolving their deployment approaches — and get a live walkthrough of Merlin Agent Builder.
Four deployment approaches for building AI agents—and how to avoid common pitfalls
Best practices for creating scalable, production-ready agents without adding complexity
Lessons from Zuora’s CIO on supporting AI adoption and enterprise-wide growth
How to maintain governance, flexibility, and control as your AI footprint expands
A live demo of Merlin Agent Builder: faster agent creation, easier management
Choosing the right path: Four ways to deploy AI agents
Scaling AI at Zuora: Lessons from the CIO
Creating and managing agents faster (live demo)
Welcome to today's Megacast, deploying high value AI agents with Merlin Agent Builder. At the conclusion of today's event, please stick around for our q and a session where we will answer the questions you submit throughout the event. Without further ado, please welcome Tray co-founder and CEO, Rich Waldron.
Hello, and welcome to the Megacast.
As AI agents redefine what is possible for enterprise automation, Tray is leading the way with a solution that makes deploying agents faster and more effective.
Today, we're not just gonna be talking about what's possible. We're gonna be showing you. My co-founder, Ali, is gonna be stepping in and doing a demo of the new Merlin Agent Builder, which we're delighted to show you today. And I'll be spending some time with Karthik, the Visionary CIO of Zuora, who's had great success deploying agents out in the wild.
So to get us started, let's look at kind of where we're at from an agent perspective.
As we make our way through 2025, we're all excited about the potential of what agents bring. We've seen the productivity gains. We can see their ability for ourselves to transform the way our organizations work. But the flip side that comes with this amazing technology is how we actually deploy it and harness it within our organizations.
Now there are show stopping production limitations with getting these things live. There's a whole new kinda agent sprawl consideration that we have to deal with, and there are business and IT risks that come with deploying such an amazing piece of technology in an entirely kinda new environment.
So when we look at where the demand comes from to get agents live, it's unique in that it's across the entire organization.
Every department is now going to their IT team, to their CIO, and they're asking to get agents stood up, getting them deployed, and they wanna get them running as operationally efficiently as possible.
There have been a few sort of standout areas where agents have already made their way. You know, we've seen a lot of ITSM agents in that kind of support function, reusing knowledge and actually being responsive. But across sales and marketing and finance, the demand really is everywhere.
So some of the challenges that come with getting agents stood up in this manner is that the promise is significant. Right? We've kinda seen this this iceberg analogy historically where the tipping point of it is the bit where you can kind of get a sense of what's possible, get a sense of what's available. And we've seen a lot of great kind of messaging and marketing around how to go and get these things running.
But the flip side of it is it's a pretty heavy deployment effort for an IT department. You know, how do you actually get these things, organized? How do you put the right governance around them? How do you manage kind of the different vendors and all of their various offerings?
How do you think about security and not sort of sending the wrong data at the wrong time into the wrong LLM? There's a whole host of headaches that come with actually successfully deploying and engaging these solutions and then maintaining them on an ongoing basis. Much like how we've had to do the same with software deployments, over previous generations, Agents kind of take that to an entirely new level.
So when we think about the options that we've got today to actually get agents live, from my perspective, there are kind of four key areas.
You could make the decision to completely build your agent solution from scratch, you know, use your engineering team, build out all of the right frameworks, the testing, and then on from an ongoing basis, manage all of the kind of scalability and change requests that come with doing so. Obviously, that's a very kind of expensive and challenging way to go about managing an agent strategy.
The second one is you could decide to go and buy a number of kind of point solutions or vertically aligned agents that work out the box. Now that could be a specific agent for marketing, a specific agent for sales, a specific agent for for IT, or it could involve going to all of the vendors that you already have. Maybe there's several hundred applications that exist in your stack and turning on the agent for every single one. That defacto kinda makes you the AI referee and the person that has to figure out, you know, which agent is doing what, what data is being sent where, and it kinda brings a whole new challenge to how you become the kind of conductor of AI across your organization.
The third option is you go and use one of the kind of domain specific or eco specific, agent builders. Now these typically exist across a slightly larger ecosystem, and they can provide a great way to get an agent stood up within that ecosystem. But you still have a lot of the challenges with connecting the rest of the stack, providing the kind of support and observability that's required to do so, and it still creates a separate place for you to have to go and maintain and manage, all of the governance and throughput that comes from maintaining and managing different agents.
And then lastly, the Tray approach, the one that we're here to introduce you to or talk you through today, which is the AI-ready iPaaS.
Now in in this solution, you're in control. We're agnostic to the various vendors or various providers that you have under the covers. You're able to go and deploy your agent solutions quickly using our entirely new piece of functionality, the Merlin agent builder, which we get a demo for you today. But, also, because it's an iPaaS under the hood, you know, the thing we're great at is integration.
We're already connected into all of your stack. We already know how to work with your different data providers. We already are able to support, manage, and change, data and get it into the right order. And then on top of that, that gives you full control over which LLM you use, which AI providers you work with.
And if you wanna change any of them, that becomes extremely easy as well. And then lastly, all of your governance is contained in a single place. It's one set of executions. It's one platform for building and maintaining all of your agentic solutions.
And so with that, I wanna, spend some time, walking through a example of somebody that's been super successful at deploying this strategy within their organization.
So today, I'm delighted to be joined by the visionary CIO of Zuora, Karthik Chakkarapani.
Karthik, who I've been delighted to have spent some time with in the past, has a unique perspective on how to get AI deployed within an organization. He's been a fast mover and has a way in which he's looked at delivering a whole new experience to the employee base at Zuora. So, Karthik, please join me on the stage. I'd love to spend some time talking through your agent approach.
Hey. Thank you, Rich. Great to meet all of you. Thanks for the opportunity to share our journey at Zuora.
Thank you, Karthik.
So to kick us off, I wanted to spend a moment on a topic that we've discussed heavily in the past.
The thing that's always kind of excited me about the way that you think about adopting technology within Zuora starts with the experience it's gonna create for the company as a whole and that becoming kind of a a differentiator and a competitive advantage when it comes to, how the company as a whole performs.
Perhaps you could, introduce that approach a little bit and talk about how you how you've used AI and agents to really go far and beyond from an experience perspective.
No. Definitely, Rich.
Thanks for all the partnership over the years, and, we are using Tray.ai at every part of our landscape.
We had an overarching vision for Zuora, enabling a true digital workplace experience.
What that means is enabling our employees to be more efficient and effective in terms of finding information, getting help, automatically fulfilling their, some of the incidents and service requests. So we started with this with this key outcomes. And then when we looked at the, our prior current state, what are some pain points and inefficiencies that we had across? So we did see some fragmentation in terms of processes, data, technology.
And that was at the same time, a year ago, the whole concept of generative AI, the agentic AI was slowly shaping up and, driving the technology landscape. So we took and we are big on experimentation.
So we immediately partnered with, folks like yourself and others, in the industry and see how we can, look at this holistically and see how we can complement our existing processes and make that experience more seamless and holistic. And that is where, we started on our, AI agent, journey. Because historically, like any other organizations, we do have, many bots and workflows automation. But the thing is those were all, like, one of solutions which was focused on a particular, domain or application.
But as an employee looking in from the from the experience of an employee, they would like to go to one place and look for a smart assistant, to to get the job, done in a very efficient and effective manner. And we live in brief on Slack, so we naturally want to see how we can use Slack as an interface to drive, many of the actions that they would like to take. So we have rolled out, an AI agent, internally for all of our employees. And, obviously, it was not a technology first approach.
It was more of an experience driven approach that addresses the pain points, challenges, and also, meet their, meet their ongoing, needs as well. So we already talked to all of all of the employees.
They get help on from from IT, HR, finance, procurement, and we also experimented some vertical use cases within, within sales and marketing. It's an exciting journey ahead. We have some very promising, results, in our the for the things that we have implemented till date.
Yeah. As you say, it it's been amazing to watch from, my perspective because you had a really clear vision from the outset. And I think many of the other CIOs I speak to, that they sort of fit into a few different camps. There's this urgency to take advantage of agents.
There's this kind of urgency to, get something into production quickly. Yep. But that idea that your experience first means that kinda under the covers, it doesn't matter so much what the various sort of tooling solutions are or the platforms are. It's allowed you to be able to sort of switch those around and almost in a way, abstract those UIs into the sort of Gen AI experience that you've created.
In going about that, approach, you know, obviously, we're delighted to partner with you on it, and and you selected Trey as, as as the vendor that could kind of bring that that agent experience to life. Could you talk through that approach a little bit and and, you know, why Tray kinda helped you in in that journey?
Yep. Definitely. But I want also want to emphasize for others who are getting, getting to start on this journey, it is like to make sure this is no longer a hype. It is real.
Right? So picking the right agent approach is not just about the technology. It's about aligning with the key business goals and outcomes and employees needs. At Zuora, we focused on more of a practical AI that just delivers the right impact and value, not just about hype.
And, yes, obviously, we did have faced some multiple obstacles, and every approach has some chaos. It's more about speed versus control or flexibility versus complexity.
The key is finding the right balance to drive adoption and scale. So at Zuora, as I said, like, we are partnering with you guys. We also, partnered with one more vendor, AtomicWork.
So between AtomicWork and Tray.ai, we were able to enable that, AI agent, on Slack that enables our employees, to get all the systems they need so that they can be more productive in their day to day work. Again, as I said, this is just a tip of the iceberg that got you began. As we started implementing, Rich, what we found out is there are so much potential in this that we have not looked before. We were only able to discover that as we started implementing and using it. For example, I'm really looking forward to see how I can use more and more of the, Tray.ai's Merlin Agent Builder and the iPaaS solution because our goal is not just an employee interacts with an agent. We also want to avoid an employee interacting with multiple agents in the ecosystem. We have, like, almost two net plus SaaS apps in our landscape.
It is going to be a fragmented experience and disjointed experience if we, if an employee has to go to each of the SaaS application and interact with their AI agent. It's almost like, someone walking into an hotel, and they see, like, multiple concierge desks, to meet their needs during the straight. Imagine, imagine that experience versus going to a single concierge desk in the hotel and get all your needs addressed. That is the same approach, we we want to take at Zuora. If you want to have something get from a travel and expense solution or from a CRM solution or from a customer success solution, we want our core AI agent to connect to those, the second tier or the third tier or second party and third party applications and interact with their AI agents and provide that single seamless interface. And that is the promise and that I'm looking for in leveraging more and more of this iPaaS solution that you guys have come with.
You made a couple of, really interesting points there that I'd like to sort of pick up on. The first one, the experimentation piece. Mhmm. I think the the way that we've enjoyed working with your team is it's sort of prototype first to get the proof point and then get the iteration to go from there.
And I think as you talked about, one of the sort of successes of the process so far is you you had a really good idea of what experience you wanted to create, and you aligned the teams around that. But you weren't afraid to move quickly to get something into, you know, an early production and really get that feedback loop going so that you could go and make some tweaks to it. You know, maybe there's a different type of data that you needed to bring in or there's a tool that you needed to change. And that's something that I think has really led to the success of the of the overall deployment downstream.
Is that something that you did sort of deliberately and and and sort of from from using an iPaaS's perspective that plays to the strengths of being able to quite quickly iterate on a solution that's already been stood up? Absolutely.
With this fast evolving landscape and changes changes consistent. Right?
So we need to as an IT organization, we need to be more adaptable and keeping in pace with the trends and because, we need to be we need to have a very, very good, system.
When I say system, the people process technology and experience end to end, whether it is in whether it is in lead to cash or hire to retire or it can be anything. So we want to make sure that we have a good robust system. In order to do that, we, we diligently follow the experience design thinking and agile methodologies, in our in our organization.
So, yes, we do have a vision. As I said, we do have, we do we do have a vision for many of our programs. We want to know what the North Star is, but we don't wanna be doing analysis and paralysis in terms of, okay, let's do this for the next three, six months. So we pick the top use cases, in terms of effort and value and look at, okay, what are some low hangings, quick wins, and what are some transformation use cases?
We kind of do a mix of all those things and do some one or two iterations and test it out and prove it out. And we're also looking for the right partners. Right? So we don't want to be, we want almost when I look for partners such as you, it almost like an extension of our team so that we can co-design, co-develop, co-experiment, and provide the value quickly.
And that is approach, we have been taking, with you guys and other partners that we work with. And that is how we are able to, to, to implement our own AI agent in a very rapid, manner. The because the results are promising, for example, we used to have, like, a couple of thousands tickets or requests coming in every month. Now we have, reduced it almost by almost three fourth.
And now this gives time back to the employees. This gives times back to the to our help desk support l one and l two so that they can focus more on more higher, advanced and strategic use cases and needs. And we want to extend this experience to all of our, second party and third party applications as well so that it makes it the whole organization can be more efficient, and this will help to drive top line and and improve our bottom line as well.
Yeah. That sort of ROI piece is fascinating because as somebody that has been in the automation space for for over a decade now, it often wasn't until you got out until the line of business teams that you are able to have a a a very clear sort of measurable ROI. Right? Because the business impact would usually be tied to maybe there's a lead lifestyle called process, and therefore, you're changing the way that that leads a, a process or or there's a downstream revenue number that gets impacted.
What's interesting in the the insight that you've just provided is in the first month, there's a measurable impact from the agent. And the way that I see it, and and, obviously, I I know from having spent some time with you, That sort of unlocks the next layer of what we go and do next because you've kinda got that ring under your belt. You can demonstrate, hey. Look.
We've reduced the number of tickets. That's changed how we're performing as a business. And then that kind of helps accelerate those next few use cases because you've got momentum as a as an organization. Is that do you is that kind of been the case?
Absolutely. Absolutely. This one is true in any, digital transformation programs. Address your core fundamental pain points and challenges.
Right? If I go enable an AI agent, let's say, for some, a use case within a domain, hey. You know what? All my basic problems first.
And that is the approach we took in place. We want to experiment in our own IT domain, because we did see some pain points and challenges where we were not able to respond quickly to our employees and, we had some disjointed processes and systems. So we mapped out the entire, employee, support, system, And that is how we were able to do it. And now we have a good, matrix of use cases, more like an horizontal organization wide use cases where every employee will need to do x, y, and z.
And lately, we have mapped out all the vertical use cases by domain like sales, marketing, and we're also working with the product and technology team. There are some very good use cases now that all of our employees know what an AI agent can do, for our basic needs. And we are getting more ideas and and use cases from our cross functional business teams. For example, we're working on an exciting use case, and that is where we experimented with Tray.ai iPaaS is is how do we completely automate our content generation for our marketing folks.
Right? For example, because as any organization or marketing team has to send campaigns, has to do some blogs and post, so our goal is at a at a click of a button entering a few prompts, okay, I need to send and campaign or do a LinkedIn blog post for this particular product, for this particular industry, for this particular persona in a particular narrative style, and we have already trained the model on what the content is, and we are and it generated a beautiful output.
Now that opens up lot of possibilities. Okay. How do we take that content generation and automate that entire marketing campaign life cycle? And that gives back so much time back to the marketing team where they can focus on focus on, okay, is the content helping?
Are we getting the leads? Are we getting, high quality, outcomes and including the whole campaign effectiveness changes the whole game of campaign effectiveness changes over here. Likewise, we are we are also experimenting a few use cases on sales and customer success. Everybody has to meet with their, with customers.
We are in a SaaS world, so we help always have to, ensure that they're getting the right value and managing so that we can we can retain them and and grow more with their existing customers. Can we do at a click of a button? Okay. I need I'm meeting with this customer in a in a week.
Can you generate a brief? Right? An executive brief that shows what is the latest engagement, how are they how are they using it, what is our option, what are some potential opportunities, what are some support cases, are there any projects? Today, we have to go to multiple systems to get that information.
That is what I'm truly excited, like, using, using Tray.ai iPaaS and other capabilities.
Can we enable that simple experience, where a customer success person or a salesperson can go with and generate that information in a moment's notice and so that they can spend more time in inferring, the insights and having the right productive discussions with the customers. So these are some common examples that we are looking into enabling.
Yeah. I I think there's a great takeaway for, the the folks that are joining us today, which is by getting that first win established and by creating an experience that, employees resonated with, it meant that as new experiences or new agents were introduced, you're starting from a strong place. People are already recognizing, actually, this made my life better in some way, or it helped me do a task I didn't really enjoy doing. And because that win has been established, exactly as you say, it's then, hey.
We're doing this now, Mark. And it it sort of trains people that that's the best way to react, and it's in many ways, it it people start to see the agent as something that is supporting them and making them better. And I think that the the Zuora team's done a phenomenal job with that because it's a truly kinda holistically aligned approach where everybody's bought in on it. There's a lot of internal messaging, which I think you do a really good good job of.
And as a result, there's excitement when these new things kinda get created. I know our team love working together with the Azure team because it's it's sort of a bit of a mind meld, and we start working on new ideas together, and they can get stood up very quickly because you've already got a lot of the data in the right place.
That's right. Yeah. We are truly excited of the potential. At least a year ago, the AI landscape was evolving every month.
It's almost like evolving every twenty four hours. And we do have some challenges. Okay. Which LLMs to go after?
And Yeah. We're working with our partners. Okay. Can we just ignore all the LLMs behind the scenes?
And can we have an interface layer that decides based upon your employees' inputs, which LLMs to pick from the back end? Right? So that is something we are experimenting right now.
Yeah. And I guess one last area that I'd like to cover, which is, you know, the the it's a critical topic, but the kinda the governance angle to this, you know, is is somebody that's looking across the whole stack that is driven innovation in such a way.
Perhaps you could just give us some insight into, how you thought about approaching the governance angle.
Obviously, from my perspective, having all of the executions occur in places that make it easy to track and to be able to put some guardrails around it is critical. And you already gave a couple of examples of things that you've sort of stood up where you can, you know, rely on downstream systems, but ultimately get one type of orchestration. So perhaps you could just kinda unpack that for us a little bit.
Got it. No. Great question. This comes up a lot in our own internal team meetings, and we talk, with our partners as well. I would add, on top of governance, it's also about security and compliance as well. Right?
So AI is powerful, but without the right governance, security, and compliance, it will result into chaos. Right?
And, we have taken a structured approach. We have, starting to establish some clear PR data controls. We are following the we are continuing to follow the zero trust framework.
And with some human oversight and more of a scalable framework, for for deploying and implementing AI responsibly.
A good example, is, if an employee has a certain access to a certain application, the same level of access should be there when they when they engage with an AI agent.
It should not be above or below, and that is the expectation that we want to set as well. And we want to make sure we are not opening new doors and windows that doesn't comply with our, with our governance and security. And, also, this adds a new layer of, what I call as a systems of contact stripe. Because all this time, we were talking about systems of record, systems of engagement, and all those things.
Now it is more about adding one more layer on the top called systems of context, and we want to make sure because the employees are going to be engaging in a new way. It's more about the context setting. In order for the context setting to be working, we have to make sure the right the data, the controls, the same, governance and securities all, is all addressed. And that is what we are working every day to make sure that this is consistent with our existing policies.
Yeah.
The from my perspective, the the key to this is, controlling the box that the LLM or the agent itself operates within. So from a kind of guardrails perspective and exactly as you point out, you know, make doing that check on a user's permissioning level, but actually doing it in a way that that is effectively a Boolean. Right? It's determining, does this person have access or not? It can't be tricked by a prompt injection or anything else behind the behind the scenes. It it very clearly defines what somebody can do and what they can't do.
And I think the flip side of that is the compliance side, having, you know, a detailed way of outputting everything the agent has done, a, makes your life a lot easier from a SOCs perspective, but, b, it allows you to start to recognize what are the ways in which the employees are engaging and how do we make sure not only do we maintain the right security posture, but it also starts to help show the things that people are asking that you want the agent to do next.
So you sort of get a dual benefit by having the right kind of visibility in in the agent platforms that that you work with.
The one thing yes. Absolutely. Great point. The one thing is, like, how when we have when we access a CRM application, we are only able to see, what we have access and whether review view it or update it.
If if a different person comes to a CRM application, they can only do what they're eligible, but LLM changes again. Right? Let's say if ten people have access to the same LLM, when they upload some information, others will be also be able to see it because now you're training that, LLM, whether it is within your security firewall or outside in the world. So those are some things that I think we are trying to figure out.
How do we how do we continuously, enable that zero trust, policies and security, not only at an agent, AI agent user experience, but also carrying it all the way to the integrations, carrying all the way to the LLM. Because we don't want to be in a situation where others can post a simple prompt and get their information from the LLM. The LLM should be smart enough. Hey.
You're not you're not authorized to query this information. That is something we're working towards, Rich. I think that is our big, big, priority this year. Right?
Yeah. And that that that governance approach of sort of limiting the place where all the executions happen so that it's not occurring in every single separate application where you're maintaining AI control over every different thing. Like, that's there's a lot of shiny objects out there, and I think you've been very diligent in in the approach that you've taken.
That's right.
I guess one last question before we before we wrap up today. You gave a a couple of examples of, some of the things that are coming around the corner. What's perhaps the one that you're most excited about or one that's kinda nearest term that that you're looking forward to to getting stuck into or is or is already underway?
Yeah. I think, maybe I'll put it this way. We cannot solve the problem internally. So I'm looking for the right partnerships such as you, Atomic Work, and others in our landscape to how do we solve and enable this, agent of agent experience.
Right? Yeah. Because every SaaS app comes with their own AI agent, and, whatever is easy for organizations such as, for us, we don't wanna be going and building that agent of agents. We want a simple framework.
Hey. For this workflow or for this experience, this is what we want to do and start configuring it or even orchestrate it much better so that all the magic happens behind the scenes. It's almost like I'll give you one more analogy. This is something I've shared with the teams as well.
It's almost like when you drive a Tesla car, I don't care what is happening behind the hood. Right? Yeah. There are so many AI work that is done behind the scenes.
But as a as an user or a driver, I don't care how it has been done as long as I'm able to go from point a to point b and leveraging all the insights and experiences from the click of a button in the dashboards and getting the work done. And that is that is the type of experience we are we are looking internally at Zuora as well. And, we want to make that, agent of agent orchestration experience much easier. That is one of our big priorities this year.
Yeah. We the way that I look at the iPaaS space as a whole is that it's perfectly suited to where I believe the challenge is. And if you think of sort of the history of iPaaS and everything it's had to adapt to over time, very clearly the next place is that kind of horizontal agent, that agent of agent deployment model because that's effectively what was happening in the software stack world. So we're we're super excited and proud to, partner with you and the team at Zuora and can't wait to see some more of those, creations come to life. So thank you so much for joining me today and and unpacking some of that for the audience.
Well, thank you, Rich. Thanks for having me. Thank you.
Thank you.
A huge thank you to Karthik for joining me today. Some incredible insights about how Zuora have been able to move so quickly and yet so effectively in capturing the power of agents.
So with that, I wanna come and introduce you to how you can get started with your own agent strategy and how you can accelerate it using the Merlin Agent Builder.
So to begin, I wanna talk a little bit about how the Tray platform is set up to go and support your agent journey. Now we've spent some time now innovating our foundation for development and deployment from the ground up. Right? We've had to rethink how do you go and help people deploy these amazing experiences or create these agents in a seamless manner and solve what I think is one of the biggest underlying problems, which is the ability to take action across the entire stack. And so in doing so, the key component to our approach is that we're not just an agent builder. Under the hood, we have the capability to handle the orchestrations, the constructions of microservices.
We spent a long time helping our customers build AI driven workflows, thanks to the fact that we've found our history in iPaaS. And so by adding this agent capability and sitting it seamlessly within one platform, it allows us to provide a kind of composable approach to deploying AI and deploying integration.
So that sets up really what the big announcement is today and what we're excited to demo for you very shortly, which is the Merlin Agent Builder. This is a completely new experience on Tray that allows you to construct very powerful agents that get all the benefits of iPaaS, but done seamlessly with a whole new UI, with a whole new set of guardrails, and an easy and effective way for you to go to from prototype to production just as we heard Karthik did earlier.
You see, the Merlin Agent Builder sits right in the heart of our platform. You know, we're already known for our unique and powerful enterprise core, approach to elasticity in the way that we can handle executions unlike any other vendor.
We have the full scope when it comes to iPaaS. So we have a history of building out orchestrations, connecting into third party solutions, working with databases, cleaning up data, and moving it to different places. And now with this Merlin layer, not only are we helping you construct your solutions using AI, allowing you to deploy and harness AI using our Merlin AI palette, but now actually go and create agents that run end to end just as we heard from Zuora earlier.
So let's dig into the Agent Builder itself.
This is a whole new experience within In tray that allows you to visually construct an agent. You can start from a prebuilt accelerator that comes packed with all of the tools that you need to get going straight out of the box. We have a whole host of accelerators that exist as a way for you to kick start your agent journey.
The core premise to our agent builder is that it comes jam packed with tools, and these tools can be thought of as skills that you're equipping the agent with.
Because Tray connects to over seven hundred services, out the box and obviously has a rich history in connecting to any third party solution and or database, it means that us adding tools is an extremely easy thing to do. In fact, our customers can go and add their own custom tools instantaneously and package them up within the agent. So when you go through the experience of constructing your agent, you can pick an accelerator, You can get through and look at the tools that are available to that agent, and you can mix and match them to ensure that they match your stack. If they if they don't quite fit, you can easily jump into the tool and extend it in any way you like. You're no longer restricted by the capability that the agent has.
So with that, I'm delighted to hand over to Alistair Russell, my cofounder and CTO, who's gonna give us a demo of the Merlin Agent Builder out in the wild. Ali, over to you.
Thanks, Rich. Yes. Today, I wanna show you how easy it is to create powerful agents using Merlin Agent Builder. I'm gonna start by showing you exactly what's possible with an agent build using Merlin Agent Builder. In this case, an ITSM agent. Then I'll show you how Merlin Agent Builder makes this possible.
So IT support is a common pain point, for IT teams. It's something we see across our customer base, and IT teams don't often scale at the same rate as the rest of the business. So they find themselves underwater with a lot of manual tasks and support triage to do. Building an ITSM agent is a perfect way to offload a lot of this work, and there are three key requirements for a successful agent. Firstly, knowledge. So you need to ground your agent in your company's data to ensure that it doesn't hallucinate.
Secondly, guardrails.
You know, they help keep your agent on track. These are basically instructions telling the agent what it can and can't do, how it should use its knowledge and the tools that it has available.
And speaking of tools, thirdly, tools empower your agent to, take action on your behalf. If an agent doesn't have any tools, then it's basically just an intelligent chatbot.
So I've granted my agent and my company's knowledge by connecting it to two data sources, a Google Drive with documents such as our company handbook and our company policies, and also an IT support Slack channel. Because quite often, IT teams see the same request or issues come up again and again, and having that knowledge exist that will exist in the Slack channel is actually quite critical.
I've given it some guardrails, so that it can be focused on being an ITSM agent and supporting our employees according to our company policies.
And the key is that it won't deviate off this track. And finally, I've given it, my agent a number of tools so that it can resolve support issues, provision access to systems, request approvals, reset multi factor authentication, etcetera.
Now my agent has been deployed into Slack. Let me share my screen.
But it could also be plugged into any number of other channels, such as Microsoft Teams, Zendesk, Jira, or it could even be used by developers directly via API.
Right. I'm in our, IT support Slack channel, and I'm gonna enter a simple message. Do I start with a simple support issue?
The agent is listening to messages in this channel, and where it can, it will trans provide support, to the to the user.
So you can see that it's responded now, and it's, given some pretty detailed instructions on how to solve the issue, including links to a GitHub issue. Now this resolution actually came from within the support, Slack channel itself. So you can see there's a conversation back up here, that includes the same sort of error message, and there's a a back and forth between the, you know, IT support and the, the user that's raising the issue, including some links. And, actually, the, the initial kind of, instructions to fix it didn't work for the user.
So they actually said, you know, there's they tried some other things, and, actually, this is the one that this is the method that actually, sir, worked for them. And the agent has understood this context and actually provided, the, the this sort of new response with both solutions. It's a simple approach. And if if solution one doesn't approach, then try that one.
So all conversations in this channel, are actually ingested into the native vector storage that your agent comes with, and is made available to the agent. So it can semantically search for similar support issues.
So the agents are grounded, you know, in, existing support knowledge, but it's also grounded in other company knowledge. So let's ask a different question. So let's say let's close that.
Say, how much how much holiday do I get?
So the agent is gonna use the knowledge that, it's been grounded in to answer this question. So you can see here it's responded with, you know, I've got twenty eight days holiday. I could roll over up to five days. You can actually see the response includes a link to the source documentation. So in this case, this is the company handbook for the, for the company.
It's used to get this information.
But it's also actually, giving me a button here to book holiday. It's it's basically suggest that it can help me actually book the holiday itself. So it's already showing it has capabilities beyond just knowledge. So this is the ability to come and take action on your behalf, but more on that in a minute.
Now I'm gonna try something slightly different.
So I'm gonna drag a screenshot in here.
Just got help.
So I'm having an issue with, Okta, and this is the issue. This should be the screenshot of the issue that I'm having logging in.
And I'm not gonna give it any other context as the agent is gonna use our Merlin intelligent document processing feature to extract, the the context of the image of the the the actual image itself, and so it can understand the full context of the request. Again, this is a really common occurrence. You know, users often just provide screenshots of the issues they're facing, and, you know, being able to automatically triage these is incredibly powerful. So, I said the images of screen drops are indicating I'm not assigned the application to GitHub.
And it knows, you can see from its response that it knows again, it's granted that company knowledge. It knows who I am based on our HR platform as I'm a member of the engineering team. And it also knows based on our company policies that I can actually get access to GitHub without any, any prior approval, because I'm an engineer. So I can click yes, meaning provision access to me.
And you can see here I've actually got, Okta here. This is my Okta login, and, I don't have any apps at the moment. And it will go and actually provision that access.
By providing our agent, with, you know, existing company knowledge, a few simple instructions, and a couple of tools the agent has been given the agency to support this user on its own, taking the burden off the IT team. It's even recognized that there's an, there's a GitHub desktop app as well, which I could install. So I'm gonna click that. And that that uses our device management software to, to go and, you know, provision that for my, my actual, machine, my hardware. It knows what hardware I've got. Obviously, the flexibility is you could easily swap out Okta for another SSO provider, you know, another, device management software, you know, use Microsoft Teams instead of Slack, or even plug it into your own internal API systems or databases.
So if I go and refresh this, you can see I should have access to GitHub Enterprise now, and on my laptop, in fifteen up to can take up to fifteen minutes to use Jamf to install that software as well.
So now I wanna ask what other applications I have access to. So, what other apps can I access?
And based on the company policies, that we've defined and and, again, the guardrails that the agent has access to, it should give me a list of of applications that I can access.
So if you click through here, you can see that, you know, I can access GitHub Cloud, GitHub Desktop, Docker Notion, etcetera. So and there are actually other applications as well that, that I can access such as DocuSign, which need manager approval. So I'm gonna ask act I'm gonna sort of ask if it I can have access to DocuSign. Okay.
I get access to DocuSign.
So what it's gonna do now is it's gonna go and look at the the company policies, and it's gonna actually, suggest that I need to get manager approval for DocuSign because that's part of the policies. And, also, it's gonna prompt me for a business justification reason. So, as part of the guardrails, you can see here requesting access, requires matter approval. Could you please provide a brief, business justification for what I need? So in those based on the guardrails that, you know, to get access, you're gonna have to pass that business justification reason over to the manager. I want to take front end designs and create the front end application code.
So I've given it a bit of a strange reason for wanting access to DocuSign.
And, again, giving, you know, the, you know, the sort of simple guardrails around, you know, DocuSign, what type of application it is, you know, what sort of department I'm in, the fact that I need badger approval and business justification reason, it will determine that actually my business justification reason isn't really appropriate for DocuSign.
It's, you know, I'm gonna have to it's gonna prompt me to give me a better one And actually suggest that maybe I'm talking about something else. So there you go. So unless your business justification seems to be about front end development work, DocuSign is actually signature platform. So maybe you want Figma, but I'm gonna say no. I do want, that's gonna give it a better reason. I want to sign vendor contracts.
So the agent should now then ping my manager and also create a Jira ticket to track this request going forward.
Okay. So it's just created a Jira ticket for me that I can track the request in, which should include my business justification reason, and that would have been sent across to my manager for a for approval as well.
So now I'm gonna show you another example of the agent providing support. This time, I'm actually gonna be, acting with a human in the loop approval. So human in the loop approval. So I've got access to GitHub, but I'm actually having issues logged getting logged in. So for demo purposes, I'm just gonna show you, that I do have multi factor authentication set up in my Okta account. So you can see here I've got a phone set up.
And, I'm gonna go into a a direct message with the IT agent here. I'm gonna say, I'm having issues logging into Okta. I have a new phone.
So I've explained what the this this situation is. And the agent knows that it can use a tool to reset my, ultra multifactor factor authentication call. But, also, it knows that the tool requires verification from my manager in the form of a video. And this is actually how we do things here at Trey.
A lot of companies receive this. You know, if you wanna do something like this, reset your multifactor authentication, you have to get some sort of verification from IT or your manager in the form of a video. So the agent's actually gonna ask me to, to basically provide a video, sort of for verification. So here you go.
So, you know, record a short video message in this thread, including my name, my work email address, confirm that you're requesting the multifactor authentication. So I'm gonna record a video, just get to here.
There you go. I'm just gonna sort of look at the camera on there.
Hi. This is, Barry, and barry at tray dot dev. And, yeah, I need a multi lateral resetting. Thank you.
Let me, send that through.
So now I've recorded the video.
The agent will then send this across to my manager for verification.
So I should actually see on another Slack, instance, which I've got just hey. You can see that the, Square agent is requesting an MFA reset.
There you go. So I've got a message coming through here. So if we go to this new chat, right, you should see oh, there we go. So it's sent my manager a this is a a different instance.
So it sends the sent the manager a request in Slack. I'm gonna verify that. You know, that's the video that I just recorded a second ago. You can see that.
So, it should then also reverify in here. And, also, if we go back to the original thread, it should update me in this original thread that the manager has actually, you know, done that verification.
Wait. Just a second.
There you go.
The, the, you know, the managers verify the request. And if I go into Okta and refresh this, you should see that my multifactor authentication is now being removed, and I can reset it up if I want to.
Basically, this shows a multistep process with human in the loop. You know, the agent has, the agency to act in the middle of two people, orchestrating the process, automatically.
Very, very powerful. So now let's take a look at how that actually works.
So I'm gonna start in my, trade workspace that I'm using to build my ITSM agent. Tray workspace is isolated isolated spaces where you can group or organize use cases and users together, with full role based access control. Now, the Tray Composable AI integration platform, is an AI first iPaaS that allows you to create automations, data integrations, APIs, and any type of AI infused process across your your tech stack. So you actually start with a sort of a blank automation, or you can, you know, use one of our many prebuilt templates that and I just kinda integrate and automate and infuse AI into those processes.
And now you can create agents as well. So when creating a an agent using managing, you can choose to build an agent from scratch, or you can choose to start with one of our prebuilt agents. So just like the ITSM agent I just showed you, you know, you can start with one of those which comes with some prebuilt tools. So let's take a look at the one I've actually been, building first.
There are a few things that I want to highlight in here. So so firstly, you can choose the the large language model that the agent uses. So this is the intelligence that the, the agent has itself, from a selection of Tray-hosted models. You can also bring your own, models from various different providers as well.
Maybe you've got a custom sort of model that you train fine tuned or trained to to sort of, you know, respond to specific types of tasks. You can customize the agent instructions and guardrails to suit your company's needs. So you can give it sort of a mission, sort of personality. You can give sort of, you know, the the the tools specific instructions.
In this case, I've actually, added a a data table which has some of our application access policies, so the the policies that we used earlier. So you can see here DocuSign, the request policy is either team or manager, which means you either have to be in one of these two teams, so finance or executive team, or you need to get manager approval. So, these are some some basic sort of policies that are given to the agent so it can, ground it in the the sort of, you know, the the guardrails of the company knowledge that you have.
Then, obviously, you give your agent tools for it to complete tasks on its own. So tools are a tray low code workflows, and we have a growing library of prebuilt tools, available. So, you can sort of choose on the ones we've already built. Our ITSM agent comes with a number of common tools, you know, that you can configure such as creating tickets in Jira or, you know, sort of, you know, doing multi factor authentication in Okta as we've just shown, or you can sort of, you know, choose your own or build your own.
So you can also use any of our seven hundred plus connectors and thousands of operations as tools. For example, if you want to, give the agent the ability to create a new account in Salesforce, or add a new comment to a Zendesk issue, you can use the those connectors and those operations to do that. They already exist. Of course, you can use our, low code workflow builder to build your own tools using Visualogix.
That's up to this one. So I've got, start to reset password flow. So, they can be as simple as as this one, which is, yeah, resetting the starting the Okta password reset, flow, and it only needs a couple of steps, basically. So once it loads, it goes, it just, you know, gets the user and then passes that user on based on email address and then passes that user over to the reset, Okta password flow.
Or they can be a bit more complex, such as this one, which is, one that installs, applications on a user's machine using our device management software, in this case, Jamf. So this provides actually extra guardrails, you know, such as, you know, does the user own this computer, for example.
But actually, in fact, the, the holiday one we had earlier, so if we took a look at this, this one actually isn't built out yet. It's just a it's just a sort of skeleton tool that, had available. But I could easily go and add, you know, sort of, an HR platform into here to actually go and create the, you know, so bamboo or something like that. You know? You can you can easily do this to sort of go and actually add, holiday by this sort of holiday approach in there, or if you've got, like, Workday or Aimly or one of those HR platforms, you can easily go and create that part time offer, request, via a simple work plan like this.
So when it comes to knowledge, switch to this tab. You know, Tray is the only iPaaS that, has a native vector storage built in. So you can ingest knowledge from a variety of sources directly into this vector storage so the agent can be grounded in your company knowledge. So you can see here I've got, this knowledge it's got the company handbook information in there.
It should have all those sort of Slack messages and things like that ingested as suspected so they can be semantically searched. And we have a library of prebuilt, data source templates to get you started. So you can easily connect to, you know, connect it to places like Google Drive, Jira, Notion, etcetera. You can see here I've got, a Google Drive one.
You can also easily build your own data sources using these low code workflows. In fact, all of our prebuilt tools and data sources are workflows, so you can customize them if you need to, needed to. You know, in this case, our Google Drive, source template uses Merlin IDP, our intelligent document processing connector, to extract data from images and PDFs in your Google Drive. But perhaps you have your own machine learning models trained, to extract specific data from specific types of documents.
You know, maybe you've got legal documents or specific types of order forms that you've trained the model on attracting the data, you know, very, very accurately.
So now you've seen, a working ITSM agent example, and I'll show you how Merlin agent build can empower you to easily deploy or build your own agents on this. Let's do, another quick example.
Now the same company, which is, BRY Construction Supplies run an online store, selling construction machinery. With the flexibility of Merlin Agent Builder, I can easily deploy an agent into, an Intercom support widget. So in this case, it's, it's just on the the website. You can ask any questions such as, what grills do you have, compare them with images.
And, again, this pulls in from the foundational knowledge that it has, in this case, from the store catalog itself, as well as, sort of these sort of, you know, product images and things like that, as well as also some, some files in the local file system. So here we got, we got some construct manuals for some of the things that don't exist on the the the website itself.
And this is deployed in via a Intercom support channel. So you can see here there's an Intercom trigger and, you know, a conversation is created or replied to. It'll actually invoke the agent and do it that way. So you can see here, it's responded to my message with, some, you know, some details about each one and the images. But, of course, it can also do, other things like, you know, from actions such as starting return. Want to return an item.
Or doing things like, you know, updating shipping addresses, updating my, so, you know, my details and things like that. Any action you can do via, you know, the Shopify connector and API, you can, you can expose as a tool. So it knows also, in terms of sort of the guardrails and instructions around this that only certain types of order can actually, be returned, ones that are fulfilled. So it's asking me which item I wanna return and, why am I returning that. For example, there's a there's a specific list of, codes, for return items. I can just say I want to, the laser level, it doesn't turn on.
And the agent will match up what I've said with all the all the sort of, you know, relevant, you know, codes that are required for returning the item. It'll choose the right one based on that, which will probably be, the item's faulty.
And it'll actually create the return for me. So there you go. So it's created return number one zero six r two. Reason detected because I said the device doesn't turn on.
So today, I've shown you, how you can easily create powerful agents using Merlin Agent Builder, starting with our prebuilt agents, data sources and tools, or building your own from scratch. Grounding the agent in your company's knowledge, guiding it using guardrails, and then equipping it with the, tools to complete tasks. But, really, this is just the tip of the iceberg. You know, the possibilities are literally endless. I'm really excited to actually see all the agents everyone's gonna start building with modern agent builder. So, right, back to you, Rich.
Thanks, Ali.
As we just saw in the demo, this is a great illustration of the power of the Merlin Agent Builder. It gives you everything you need to manage all of the parts of the agent deployment headache that lurk beneath the surface.
And when you do this using the foundation of the AI-ready iPaaS, it unlocks your ability to be successful with agents across the entire business. Just as we heard with my conversation with Karthik, who with a single agent was able to reduce the number of tickets that IT received by three x.
When you consider this capability being rolled out across the entire organization, working with multiple agents and multiple people, the outcomes are pretty mind blowing. And so with that, let's head over and answer some of your questions.
We want to thank you for your time today and hope you enjoyed the session. We will be sending out the recording shortly. If you'd like to explore more with Tray's Merlin Agent Builder, visit the page on your screen now. You can also find the link in the chat along with more great resources to explore. Have a great day.
CTO
CIO
CEO