Skip to main content

Moving AI Agents from Experiments to Enterprise Production

Summary

  • ARM and Celonis demonstrate that process intelligence—derived from process mining and event-sequencing of real operational data—is the missing layer between raw data and autonomous action that keeps AI agents reliable in production.
  • ARM's circle of agents operational framework evolved from three isolated point-solution use cases into an interconnected system of transactional, monitoring, and planning agents deployed in production on the Databricks Data and AI platform.
  • Trace-based observability closes the feedback loop so agents can optimize their own behavior at runtime, enabling ARM to scale from individual process fixes to enterprise-wide AI-native operations including real-time autonomous finance planning.

Moving AI Agents from Experiments to Enterprise Production

Watch: Moving AI Agents from Experiments to Enterprise Production
Most enterprises struggle to move AI agents from isolated pilots to production-grade deployments because agents lack real-world understanding of how their business actually operates. Without this operational reality, agents remain unreliable at scale and slow to deliver measurable ROI. Process intelligence provides the missing layer between raw data and autonomous action, grounding agents in actual business logic.
Watch ARM and Celonis demonstrate how process mining and event-sequencing capabilities enable agents to understand business workflows, decision points, and constraints. Learn how trace-based observability closes the feedback loop, optimizing agent behavior at runtime. Discover how combining process intelligence with Databricks data infrastructure transforms manual business processes into AI-native operations that improve autonomously.
🤝

Chapters

FAQs

What is process intelligence and why do AI agents need it?

Process intelligence is the capability to understand how a business actually operates—its workflows, decision points, and constraints—derived from process mining and event-sequencing of real operational data. This video argues that without this layer, AI agents lack the business context needed to take reliable autonomous actions and remain stuck in experimental, point-solution phases.

How did ARM build its circle of agents framework?

ARM's circle of agents is an operational framework that evolved from isolated point solutions to an interconnected system of transactional, monitoring, and planning agents. This video describes how the journey progressed from three early use cases through ARM's first transactional agent in production, expanding to include finance planning agents capable of autonomous real-time decision-making.

What role does Celonis play in ARM's AI agent deployment?

Celonis provides the process intelligence capabilities that give ARM's agents a grounding in how business processes actually unfold, rather than how they are theoretically documented. Celonis was named Databricks built-in partner of the year during the event where this video was recorded, reflecting the depth of the integration between the two companies.

How does trace-based observability improve agent performance over time?

Trace-based observability captures detailed records of how agents reason and act during runtime, creating a feedback loop that allows teams to identify failure modes and optimize agent behavior. This video explains that this observability layer is what enables agents to learn and improve at scale, rather than requiring manual intervention each time a process changes.

Full transcript

[00:07] My name is Dale Williamson. I am the AMEA CTO at Databricks. So, it's really cool to be here today and to moderate this third time we have been together. So, for some of you in the room, you don't understand that this is act three.
[00:23] Um Harsha did the first moderation in November. I then did a second round moderation in I think it was end of January. And we're doing sort of act three. We got act four set up for December in
[00:38] Munich. So, um Let's just do do a quick round of introductions. So, James, do you want to introduce yourself? Yeah, hi everyone. So, my name is James French. I the director of process intelligence at
[00:53] technology company called Arm based in Cambridge in the UK. Arm is a Well, I used to say it was a chip designer, but that's recently changed. We're both a chip designer and and soon a chip manufacturer. And just so you all know, you all use
[01:10] Arm all the time. Anyone in this room who thinks that you can run AI without Arm, good luck with that. Um Harsha, do you want to Yeah, absolutely. Good afternoon, everyone. My name is Harsha Dronamraju. I lead our AI product at Celonis. So, if you're not familiar
[01:26] with Celonis, we're a process intelligence company. We help you understand how your business actually operates and how it ought to operate. We're built-in partner of Databricks and also work with Arm and James. So, looking forward to it. And what's super cool is Harsha is being quite humble.
[01:43] Celonis got built-in partner of the year yesterday at our partner awards. So, it's big big deal. Two kinds. Awesome. Right, so so so we're going to set the scene a little bit. This is kind
[01:58] of like a journey that we've been on together. It's it's sort of a There are three parties involved. We're helping James to sort of realize this process optimization, process intelligence, and like one of the fundamental things that
[02:13] people do not understand is how complicated this is and how fast it changes. So so when we did act one, it was a very theoretical conversation. When we did act two, it was off the back of three use cases that had been done
[02:30] where we're trying to figure out very point solutions. About a week ago, I went up to Cambridge, the one in the UK, the famous one, the original one. Sorry, can't help myself. Um and we we
[02:45] had we had a demo of the latest and greatest that James has uh and his team have produced. Um they call it the circle of agents. And I want James to kind of take us through what what it does and what they've learned from point
[03:03] solution to kind of an end-to-end flow. Yeah, so just to pick up on the kind of the history of that. I think if we look back in November when we first started talking about it, we were talking about theory. We were talking about
[03:18] experimenting. We had in our minds this idea of um well, Arm finds itself a good intersection of being a Databricks partner and being a Databricks customer and and sitting alongside Solon. So, we've got in our domain, we've got both insight
[03:36] into process and a fantastic well-governed data domain as well. And so, we thought this was ripe for the world of agentic AI that was coming. And so, we started to really sort of think about well, how do we bring those two worlds together? And what does that mean in
[03:52] terms of how we can optimize how agents run, how we can improve the efficiency and performance of those agents, but also then think about well, what do we do to monitor and assess those? Um and so back in November we had this idea. We thought this is going to be really,
[04:08] really important. Like the context that Celonis gives us in terms of process is going to be the rocket fuel that we need as we go into this kind of world of agentic AI. Um but it was a thought. And as you said, back in November we said,
[04:24] well, you know, let's go and experiment and and take a few ideas. And we picked three use cases. One was based around um how we might interact with ServiceNow for request management. One was around how we were thinking around uh knowledge management in our organization. And then
[04:41] the third one was a uh one that was based around our commercial flow, lead to royalty, and how we might use agents to help support our opportunity management, order management side of things. And that's where we started in November. Rolling forwards into to January when we
[04:58] first got our heads back again after that uh kind of initial meeting, I guess where we got to was those early foothills of experimentation and understanding where we were finding some of the challenges. And I guess that's where we are now
[05:14] progressing now beyond uh uh January's sort of time frame. When looking where we are last week, we've taken the circle of agents, which is our our Salesforce-based agent that's looking at order management, and we are now in a position
[05:29] where we're bringing those two worlds of data and process together. So we're we're close to productionizing our first true transactional agent in into the organization. One that will operate fundamentally changing the way that our
[05:45] sales folk interact. No longer going into Salesforce to manage opportunities, generate quotes, but actually taking them, you know, into a layer where they're able to chat using, you know, familiar interface and create all of those assets that are in Salesforce. And
[06:01] alongside that, both using the process insights that we've got from our long history of opportunity and order management data that we've got in Celonis to provide the context necessary for that agent to be optimized run in a more performant way. And then, the great
[06:18] thing off the back of that is we can using all of that trace data that we're getting from runtime in in Databricks and passing that back into Celonis to help visualize and understand how that agent performs in the context of the overall process. And I think that for us
[06:34] has been the sort of the the unlocking of the potential we see here. A way of us monitoring how the agents performing in in in uh production and in the context of the process as well, which I think is really important for us. There have been some significant challenges along the
[06:50] way. We do recognize, and I think we're all suffering, the fact that this is a really dynamic environment, right? No one day do you feel as though you've got a stable position, and you're always thinking, "Well, what's next, and how are we going to develop that?" Um and of course, for us, it is our first true
[07:07] production release. So, we're going through all of the things that we need to think about in terms of the organization, about how we release, how we manage, how we govern, what are the right levels of control that we need to understand how agents are going to perform and how we manage those in in a production environment. So, some of it
[07:23] is pipe cleaning for us as an organization. It will build upon, you know, the first one going in, but some of it is also recognizing that every single day something's changing, and how do we manage that? It's like an amazing journey, cuz
[07:38] those of you who've kind of followed, maybe, listening to us, It's kind of theory to experiment, to end-to-end, and this is just one process. Um I want to kind of run into Well, first, what process intelligence,
[07:54] what it is, what it does, Celonis's superpower, but we heard a lot today from Ali about context. And and I I'd love for you, Harsha, to just tell us one, the whole mapping sort of
[08:10] process intelligence, process mining side, and sort of where where you sort of see this context piece play out. Yeah. So, the fun thing I was reflecting about this conversation, and, you know, a lot has changed, obviously, but some things
[08:25] have stayed the same. One is, you know, back in January, we're like, yeah, like, there's this missing layer that sits above your data between your agents, and how do we um how do we make the most of it? And I think we've learned a lot since then. So, you know, when when I look at all the great stuff that we you all launched today, you know, I think
[08:42] we're just getting much more discerning as builders in terms of how do you deliver context, in what format, and also what type of context matters in what use cases, cuz it's not all the same. So, you know, in terms of how we've delivered context, you know, back when we started, we're
[08:58] talking about prompt engineering and kind of filling up the prompt window with a bunch of interesting stuff. Then we discovered, wait a second, there's these notion of tools, MCP servers, context engineering, it's evolving into harnesses, it's evolving into kind of a a much more robust way to think about it. So,
[09:13] the how is evolving, and then to connect again to this notion of context is different depending on what use case you have and what you're trying to solve. You know, when I think about what Celonis offers, we have a very process-oriented view of the world, right? So, we've been around since uh 15
[09:29] plus years. Everything we do is uh based off of a world view that an organization is a series of intersecting events, and the order matters, and depending on how an order of the order of events happened, that actually tells you a lot about what's
[09:44] actually going on. And so, when we think about context, we take a very processed view towards it, and I think it comes in three different flavors. The first is, you know, when you're looking at an event, when you're looking at a sales team that's looking at something, you want to look at the
[09:59] history, right? So, it's not only what am I looking at, but how did it get here? What are the three or four steps that happened in this business workflow that led to this landing on an agent's desk or a human's desk? How does it typically happen in the past? What are
[10:14] the different variants that it could have taken? All of these are kind of traditional process mining capabilities that we built into the platform. Now, as we kind of grow on that, we're building on that with things like situational awareness. So, not only where has it been, but what am I looking
[10:30] at right now? What are the objectives, the goals, the process it abides by, the permissions? All of that matters because it gives you the agent a clear opportunity set of what it what action it should take. And then finally, as we think about where we'd like to take our platform,
[10:45] we're we're spending a lot of time exploring kind of the future both what might happen, foresight, and anticipation. So, where might this thing I'm looking at and go? What are the other directions it might take? And so, from our perspective, we always look
[11:01] at context as this 360° view of the flow of events, which I think is a really nice compliment to kind of having this robust data set estate that Databricks is providing. See, I love that cuz, you know, we we often joke about this. We say sort of
[11:17] they have a context layer, we have a context layer. And actually, the the irony is that they're contextually different. Um and and I think this is the really and then we can make it funny. Yeah, I've just I'm just going to keep saying it cuz eventually people will catch on that it's funny. Um
[11:33] Um but the um I I think we're only really starting to grapple with just how much we don't know about how things work in an organization, how much sits in people's heads, how much is undocumented. Like what I love about Celonis' superpower is being able to kind of see
[11:49] what's really going on. Like there's a six-stage sales process, but there's actually a thousand things that happen. Where do those thousand things get documented and how do we figure out those edge cases and things? How do people actually do work? Cuz trust me, if we just let agents do it, it's going
[12:05] to be a nightmare. So we've talked about this kind of idea of building the map, building the kind of end-to-end, you know, being able to figure out how to inject agents in. We've had a successful bunch of use cases. We've had a successful bunch of end-to-end experiments. And there's been
[12:22] some learning. Now, normally what happens is people will start to optimize. You know? Map the process, lean thinking, improve the process. That's not what you're doing, is it, James?
[12:38] Well, no. So I think um we've taken a very sort of bold approach. And I think that's that's partly a cultural position in the organization. We we we tend to think of ourselves as um constantly pushing, trying to reinvent,
[12:54] and and think about what we could do differently. For us, there was a danger that we started to think about identifying a poor process, you know, um some sticking plasters that would get us some incremental improvements, but wouldn't
[13:09] really transform how we run as an organization and what we do. And I think the opportunity was there for us to to kind of throw that away a little bit, and to start with a blank sheet of paper. And so we're standing up a a program, which is
[13:26] like many organizations, it's titled AI native, AI native arm. Um but for us that's not about those incremental improvements. Um it's looking at in this brave new world where we've got this AI capability,
[13:41] how would we do it completely differently? Where do we need actually just throw away that existing process? I mean, we've got great understanding of how processes have run in the past, but it's been built around you know, a paradigm that is is largely based on humans involved in the process
[13:58] around interacting with systems through a, you know, UX, UI. And actually for us, the great opportunity is to say, "How do we radically transform each and every single one of those business capabilities in this brave new world
[14:13] where AI is at the core of everything we do?" And I think that is a challenge because it's uh organizationally, culturally, it's it's making people take a complete step back from their day job, from how they've done things in the past, and completely reinvents, you
[14:30] know, kind of the organization's approach to these uh delivering these business capabilities. And clearly, that's a massive step change. So, I think for us, whilst we do that, there is still space for us to deliver those incremental improvements. And so, this
[14:45] is not a complete, you know, we're not going to do anything until we can completely reinvented the company, but the aspiration is take the enterprise as a as a as a set of functions and completely rethink how AI changes how we operate. The finance
[15:01] function, the people function, IT, all of those core kind of uh back office functions. And so, the very first one that we're going to go after in uh in September, October, is to look at our joiner process. So, we as an organization going through
[15:17] a massive, you know, growth in terms of the number of people joining um the organization. And clearly that's a big strategic driver. We need to think about how do we bring people into the organization, make them effective from as early as possible, be they employees,
[15:33] be they third parties. And if we think about it in a completely radical different way, we think that's going to be transformative to that strategic objective. And so betting ourselves against what's important to the organization, but challenging
[15:49] ourselves to think completely differently and lay out AI agents as a core bedrock of how we do the business rather than as a supporting function to kind of traditional processes that we have already.
[16:04] See, I One of the things I love about that is you going up to the people function and there's almost like an onboarding of people, but there there's a second onboarding of the non-humans that work for them, whatever form they take and whatever kind of permutations that becomes and
[16:20] the management constructs that form around. So, this is going to be a really fascinating next step. Um from a product perspective, like you're thinking about where Whistler goes, you've made a recent acquisition. Mhm. Walk Walk us through that and
[16:36] effectively like what you sort of see this human plus machine process future look like. Yeah. So, um before we I'd love to go into that. I have a thought experiment. So, you know, something I do a lot is whenever I have
[16:51] a decision I'm trying to make at work and you can all try this at home, just like close your eyes and think about what are like the bits of information that you're trying to like process and go through in order to make that call, right? You're not even thinking about it, right? But like in your head, you're probably thinking about a number of
[17:07] things that are looking backwards, you know, what is the the my relationship with this person, you know, what is the the rule I have to follow? And so, you know, that kind of gives you a sense of kind of what goes into making a high-quality decision. And really the thought experience, how do you give that
[17:23] to an agent, which is a really hard problem to solve, but like will be the key unlock for all of these automations. And so, to that extent, um to that point, you know, we recently acquired a company that called Ikigai Labs. They're based out of uh Cambridge, Massachusetts. Uh the the new Cambridge.
[17:40] Now you understand why I made the joke. It's only 400 years old, so we're not we have uh we have some we have some room to grow. But um uh they are uh they focus on uh what I would call um time series based decision intelligence, right? So, uh a foundation
[17:55] capability that they built are large uh graphical models, which you can almost think of a large language model, what that does for text, uh they built based off of tabular and time series data, which is really useful in an enterprise context. It can be more accurate for
[18:10] time series based uh generative AI. It can also be more computationally efficient. The thing we really like about Ikigai and what we think it fits into some of the stuff James you're talking about is it can help you not only understand, "Hey, at given moment in time,
[18:27] process mining is all about why something landed to this moment in this organization." What decision intelligence and Ikigai can help us with is help us look at, "Where might this go? Where might this go if I don't do anything? And if I slightly tweak, slightly left, slightly right, what are
[18:43] the other options that this can get to?" And when we have that 360-degree view, we're really excited about what it can unlock for customers like you. One thing I loved about what James said is you're not thinking about a tweak here when you see the spaghetti map of how crazy are all of our processes are, you're
[18:59] thinking about, "Okay, from a clean slate, what does a truly agentic first process look like?" And when we want to place, you know, Arms agents in that new agentic process, those agents need to look backwards and look forward, and they're passing down this this information, making decisions, passing
[19:16] it down to the next node of this new decision workflow. And if we can give you all the kind of the tools, the function calls to make that happen, we think A, it helps your agents talk to one another better. And then if I could take a a bigger step back, you mentioned something about kind
[19:32] of bringing the traces back through Databricks and through through Celonis, now we can almost create a agentic loop where the process you're learning from the prior traces, you're you're that's informing the directions you go cuz those have all kind of informed how the business operates, and you can start to
[19:48] get to a really exciting place, which is kind of an agentic process that's always improving and hill climbing. So we we did a little bit of that just to jump on the the kind of idea about how do we use that trace data to run time. So not only just in the sort of
[20:04] pure post-processing analytical space to understand return on investment type discussions, but also to think, well, how can we actually use that information to help optimize tool calling, right? So we can understand when agents have made
[20:21] kind of you know, a sub-optimal path through tool tool calling to answer prompts, we can help sort of steer and direct based on that history and based on using that trace data to kind of guide at run time as well, which is which is actually like the first step in I think what we're
[20:36] talking about there, which is the idea of actually that 360° loop. And if we couple it with the stuff that you know, coming through the Ikigai stuff that I've been thinking about is simulated outcome. Yep. You know, which is actually really
[20:52] interesting when you get into kind of agents performing simulations, understanding the impact before they take take action. I think that's a really interesting space that we we'd like to get into, which is actually that's that's the next step in the journey.
[21:07] That's let's have four, you know, when you meet again and maybe we can get there. Yeah, and you know, imagine you have a transactional database that can branch and merge and create 10,000 different permutations and a gateway that allows you to capture all the substrate and
[21:24] trace hint hint to everyone who watched the keynote today. it's coming together? Like um I I want to I want to dip into this because this this idea of kind of building a map, building a sort of optimization structure, almost building
[21:39] a sort of organizational twin, where you can run these simulations and figure out how to get the recursive loops back to improve agent performance, even improve the human performance, figure out where humans need to go, like the stuff that we don't have. Um
[21:57] when we did Act One, we went into theory. We've got some grounded evidence. How about we do a little bit more theory? And kind of talk about like you you want to do this cool stuff starting September. You're going to build some sort of primitives. James, where do you think
[22:14] the kind of idea of this simulation could help you to reduce some of the the blind alleys? I mean, you have a incredible engineering culture in Arm. Like, where do you sort of see the the resource going, the people that are going to be
[22:30] working on this going? Yeah, I think um it if I just if I sort of ground it in where we are now, I think one of the one of the things we're we're sort of learning is um the way that we've approached the problem space at the moment has been very linear, right? So, you know, kind
[22:47] of the idea of we've got a single process, we're we're running down a path, we know that we've got sort of a a a pre a predetermined pathway we're we're kind of running down. But, the interesting space is we look we broader and look at our
[23:02] organization and how processes interact, how agents are going to interact, the interplay between those things, and giving our users the the sort of the freedom to allow or or the organization the freedom to allow agents to go and solve some of
[23:17] these problems themselves. And we had a discussion, you know, a couple of weeks ago on when you visited us in Cambridge to think about like actually some of this stuff is best suited for agents to go and work out the answers to some of this problem in the call that the bit of truth.
[23:33] Yeah, and and actually we as humans probably not always the best people to go and solve some of this stuff and think about it. So actually when we start to think about like at scale agents being able to collaborate, to work together, to use kind of capabilities to simulate outcomes, to
[23:49] think about how they might solve a problem in the context of something like an enterprise process, it's it's quite transformational. You know, I can imagine how you know, roll forward in 6 months and we're thinking about a a deployed set of agents who are
[24:04] running autonomously in the background, simulate outcomes, you know, thinking about what the the kind of possible scenarios are. That feels really exciting. One of the things that we're looking at um in the next phase as part of our program alongside the stuff that we're doing with the people team is to think about
[24:20] how our finance planning cycle was could be transformed. And that feels as though that's very very kind of core to what we're talking about here. You know, at the moment when we run on a very costly based you know, very manually driven kind of uh process around our financial planning
[24:37] cycles and budget cycles. And actually if we can get closer to more kind of real-time decision-making as an organization, the ability for us to assess the viability of some of our product development, you know, kind of streams, think about whether there's a
[24:52] you know, a market for that product, whether that's drying up, and whether we need to quickly and dynamically shift resources to a kind of an alternative path. That's when the organization suddenly transforms into something that's a bit more kind of dynamic culturally and actually that's where we
[25:09] release a huge great potential for return on investment for this kind of AI space for us. So, yeah, I feel as though if we can solve that problem, if we can understand how we can bring those more simulated outcomes into it, we can get a first set of agents deployed, and we can bring
[25:25] those two worlds together. It's going to be really transformational for us as an organization. I think that's amazing. I I want to I I I I'll open this up to the audience for questions in a minute, but I want to I want to sort of tease out one last sort of question, Harsha. Like
[25:42] what we've talked about is existing processes. How does like one of the things I am is doing is kind of setting up things that they don't have. Fabrication, supply chains. How would Celonis
[25:58] drive that kind of you've got tons of proprietary data on process. How would that sort of propagate into an ideal supply chain process, for example? Like do you have stuff like that that you're thinking about, that you're working through? Is that almost simulation
[26:14] aspect? Yeah. So, uh this is an area that I get really excited about, right? Because um you know, one of the things that customers will always ask us for is this notion of just benchmarking, right? Like it's you know, I I have this on-time payment
[26:30] rate, like how do I compare? And and and that's a really useful thing that we can offer our customers, but when we can kind of help you frame how much value could I really unlock if I were to improve my processes. I think what you're describing um what you're describing is like taking
[26:46] it to the next level, which is like I don't I've never entered this new market before, and how does that even like look and what are some of the capabilities, the the value stream extensions that I have to implement? I don't have anything to announce here today, but it's definitely something
[27:01] that we're really excited because if you think about it, I want to disconnect it back to something you said earlier, which is you know, you started with one part of the business, but slowly you expanded Celonis built on top of amazing infrastructure to kind of be this horizontal layer. And so, you know, the
[27:18] as a design principle, Celonis is built to kind of allow you to kind of almost have that fog of war kind of map if you imagine those old video games, right? Where you're kind of always trying to understand what's at the periphery and what could I expose and what are the things I need to add to
[27:34] my kind of map in order to kind of understand what's available to unlock. And I think we can do that in a way that's really exciting um for for existing customers, but also for the entire community of process improvement professionals because, you know, you know, we we can help everyone
[27:52] achieve a higher output and and more productivity, which goes back to kind of what our mission is, which is making you know, processes work for people, the planet and and the people, companies and the planet. So. So, from what he said, you have to go to
[28:08] Celosphere for whatever he's going to announce in December.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.