Agentic Finance: The Power of a Unified Data Foundation

00:10 - 01:39

Hello and welcome, everyone. We're gonna give folks just another moment to get settled and logged in, and then we'll get started.

Alright. Looks like we have a great group dialed in.

Thank you so much for joining us for today's demo of Agentic Finance, the power of a unified data foundation. My name is Najla Brown, and I'll be your moderator for today's session.

Before we officially kick off, just a few quick housekeeping items. First, today's session is being recorded, so you'll receive a copy of this presentation in your inbox.

On the right hand side of your screen, you'll see the q and a tab. Please submit your questions there at any time, and they'll come straight to myself and our speakers, and we'll address them at the end.

Finally, you'll see a docs tab in that same panel. I highly encourage you to check that out for direct links to resources you can save or explore later.

And with that, let's dive right in. I'm excited to introduce our speakers today.

Tiffany Ma is a global product expert at OneStream. Joining her is Logan Shallenberger, director AI solution consultant at OneStream.

Tiffany, the mic is all yours.

01:39 - 04:54

Thanks so much, Najla. I'm really happy to be here with you all today.

Thanks so much for taking time out of your day to join us for this webinar. I'm just going to take a, the first few minutes today to talk through the current state of Magento finance, and then I'll pass it over to Logan for a demo of how that looks like in practice.

And then I'll close it out with a brief one stream overview and then get to questions. Please go ahead and add any questions or comments that you have into the chat, and we'll get to as many as we can at the end of this session.

So let's start with why this topic matters today. There's a statistic from IDC, which is that 26% of AI's productivity gains get lost just to verifying and explaining what the AI actually did.

And so that's not a small tax. That's for senior finance professional, that works out to be about thirteen hours a week.

So roughly $78,000 a year spent double checking the black box instead of trusting the answers that came out of it. And that's the real tension in Agenctic finance today.

The office of the CFO doesn't want a black box. It wants a clear transparent box where every answer is traceable back to the number, the rule, and the source that it came from.

So when we talk about agentic finance, we're not just talking about agents that can act. We're talking about agents you can actually audit, explain, and defend in real time, not after the fact.

And that's the foundation everything else you're gonna show you today is built on. So how do you get to that glass box, so to speak? Well, it starts with where the agent actually lives.

And so Sensible AI agents are built directly into OneStream and therefore are grounded in OneStream's single data model with built in consolidations, planning, and reporting computational engines. And, you know, safe to say before any agent can be trusted by finance, it needs a foundation worth trusting.

And OneStream is that trusted foundation. It's not just a database.

It is the system of record for the office of the CFO with full governance and auditability, soft controls, and financial intelligence all built in. What we've done is we've then extended that, OneStream's computational engines, the financial intelligence and governance from our foundation into our agents that we've built into OneStream.

This means a question that's been asked in natural language can be translated correctly and into an accurate answer. That's definitely.

That's guaranteed. And then lastly, we're meeting our users where they work.

We recognize organizations are buying enterprise AI tools like Copilot, like Cloud, ChatGPT, Google Gemini, and that's why they're doing a lot of their work today. And so what we've done is we've provided an MCP MCP finance agentic layer to allow users to access OneStream's govern data and agents through this agentic layer.

And with that, I'll pass it over to Logan to start the demo.

04:54 - 20:45

Thanks, Tiffany. Let me go ahead and get my screen shared here.

Okay. We should be looking at one stream, Tiffany.

If you can just give me a thumbs up, that we got it. Good to go.

Okay. Perfect.

Thank you. Now I wanna walk through a little bit of what Tiffany was talking about in practice.

Right? How do we actually ground AI in the financial metadata that ultimately drives success for AI? Right? Within OneStream, we built our agents to understand the platform. They understand the cube.

They understand dimensions and relationships that might exist amongst them. They understand financial statements.

Right? Everything that has been core to the Longstreet system since our inception, we built that understanding into our AI. Now this little button on the top right hand side of my screen is where our agents live behind.

Clicking that button allows you to start interacting with the system in natural language. All you have to do is fire away a question.

Right? Questions come through in natural language and ultimately are translated back into your unique one student application. In this case, I grab one of my suggested questions.

The question is, show me my detailed income statement broken up by region. I like this particular question because it starts to highlight exactly how our agents are grounded in the financial truth within OneStream.

In this case, I asked for an income statement broken up by region. Income statements within OneStream can be defined in a variety of different ways.

In my case, my income statement accounts are defined through both my hierarchies in my account dimension and in a standardized report within my application. So what my agent is ultimately going to do is it's going to read into all of those different components to understand what does an income statement even look like.

Right? In this case here, it says, I'm told of my North American income statement, and I'm checking to see whether there is a regional detail report that fits. So it's going and building a report based on the prompt that it's been given.

In my case, what it's doing is it's gonna build a custom report on the fly. It's gonna use my income statement account structure from my standardized report.

It's gonna build a new report off of that structure and incorporate the region dimension as I've asked it to. So that's what I mean about being grounded in the OneStream application and with all of our metadata.

Right? It can read all of these different hierarchies and understand what I actually mean when I ask for an income statement. One of the things that you'll notice with OneStream's AI is a heavy emphasis on auditability and transparency into where those numbers come from.

That's why in the transparency into where those numbers come from. That's why in the conversation itself, every single number is highlighted in blue, and that's because it's ultimately a hyperlink back into the report that the agent surfaced that number from.

That report is listed at the end of this conversation down below. You can see the structure of that report.

I'm pulling my account tree for my income statement accounts, and I'm pulling these three regions into my columns. It also highlights to me the point of view that was taken.

The point of view is an important attribute for this conversation. My agent went and selected a number of different dimensional components to get to this level of detail.

It selected what queue we're looking at. It selected what entity we're looking at, which consolidation member, which scenario, time period, view, so on and so forth.

Right? Important attributes to understanding the data that has been driven into this conversation by my agent. Now, obviously, I can open up this report.

This report can be surfaced, and I can drill down on any of the numbers that are coming through in this report. It's one of the capabilities that is unilateral across the OneStream system.

Right? When you see a data point, you can always drill down on any given data point. There's one in particular that I want to go into.

It's this North America number, and this is really gonna highlight what we mean when we say grounded in the OneStream application. This particular number, we're looking at a $398,000,000 number that's, denominated in US dollars, and I'm gonna be drilling down on my North American equipment entity.

This entity is a parent across multiple different entities, some of which are denominated in foreign currencies. This guy here, that is my Canadian entity denominated in Canadian dollars.

One Street itself is a living and breathing application. We do a lot of things dynamically.

In this case, this foreign currency translation can be done on the fly when we request that particular number up above. So by taking this approach of using reports as a mechanism for how our agents interact with the data itself, not only are we maintaining the lineage of every single data point that the agent is using as a part of its conversations all the way back to the cube itself, but we are also tying into any of the logic that you have built into your OneStream application as well.

Things like dynamic calculations and aggregations. Right? When we ask for a parent level number and that number comes alongside with calculation, those calculations are done on the fly when that number is surfaced in the report itself.

Another important thing to call out here, our agents cannot hallucinate on any of the numbers that you see in this report because the agent itself is not interacting with the data directly. The way that we built this solution is by using the agents to interpret when I prompt it what this means in my dimensions, in my financial hierarchies.

Right? So what it ultimately did is it identified where the data lives within the OneStream application. It lives in these accounts.

These are my income statement accounts. When I said by region, it knew that I wanted to look at both the parent level and the two entities that exist underneath.

Right? A full picture of my operations. The report is then built off of that understanding, that translation back into my hierarchies.

The report is actually what pulls all of the data itself. By taking this approach, it cannot hallucinate on any one of these data points because a report is ultimately a deterministic object within one stream.

Right? It defines what intersections exist. And for instance, across this particular intersection, net income for North America, that is the only data point that can exist.

So by taking this approach, we have increased the accuracy of the results that we see. We've also built this in a way that prevents our agents from hallucinating around your data.

Right? So it's grounded in those hierarchies, which are still very, very important to driving success and accuracy with our agent conversations. I want to touch on another aspect that we have within our agents.

It's a concept that we call personas. It's this little drop down box here.

Personas are important because number one, we do not train our AI on any of your data. It's a nonstarter for just about every IT organization everywhere.

Right? Financial data is a sensitive topic. So we don't do any training of our data.

So we need to configure our agents to understand your unique OneStream application. One of the ways that we do that is through personas.

Personas allow us to build natural language rules into our agents. Right? I was operating underneath my company wide persona.

These are all of the rules, like, I have a rule here telling it to focus on a particular entity. Just to highlight the extent that this goes, in this case here, I have a tax reporting manager.

That tax reporting manager is inheriting my company wide persona and also incorporating rules that are unique to my tax team. Like, for instance, the use of an alternate reporting hierarchy when somebody from tax is asking about their financial reports.

It's gonna automatically use the hierarchy that was built to support tax reporting as opposed to GAAP or I r. IFRS hierarchies.

These personas are important because this is how we can start to affect the behavior of our agents and ultimately increase the experience that our users will have with AI. They are also important because these affect the external AI tools that you may have plugged into OneStream as well.

So I'm gonna show you an example in just a moment of cloud being plugged into OneStream. This is also going to affect the behavior of cloud when cloud is interacting with the OneStream system.

Before we get into Claude, I want to highlight one additional capability. That capability is around automation of this process.

Right? The experience that we were just looking at is more around the ad hoc interaction of data. I have a question.

I'm gonna prompt the system. It's gonna return some results for me for me to interpret or leverage the intelligence of AI to interpret.

What about things that happen on a recurring basis? In my case here, I have a P and L variance analysis process. The way that this works is I've set this process up with AI.

I prompted my agent and I said, I want to do a detailed income statement review on a monthly basis. The agent asked me a couple of follow-up questions.

So any time our AI finds uncertainty, it will prompt you to get some specificity around this particular analysis that we're tasking it to do. In this case here, it wanted to know what time period, which scenario, and which materiality threshold to use as a part of this analysis.

I gave those three answers, and that is the extent of the setup. Now what happens is you can schedule this.

You can schedule this to run either on a recurring time basis, like this example here running every Monday at 8AM, or you can attach this to workflows within the system. So as soon as the consolidation process completes, go run an analysis of the consolidated data.

Or as soon as a budget submission completes, go run an analysis of any budget data that came through that submission and maybe look for any anomalies that might exist with that submission. The output of this analysis looks a little bit like this.

It's a p and l review. Right? Remember, keep in mind, all I've prompted my agent with was analyze my p and l.

It gives me a headline statement, that's this line item up above, that says performance for this entity came in and had a budget with a net income of 9,600,000 versus a budget of 5. 8, a favorable variance of about 3,700,000 or about 64%.

The largest driver was a lower than budget income tax provision while core operations were also modestly favorable coming in ahead of budget by about 350,000. It then starts to drill into all of the different variances that it found and ranking them based on the impact that they carry.

For each given variance, it's going to provide me with commentary and narrative that try to explain the why, the story behind the data itself. So our agents are interacting with the data and they're interpreting the results on my behalf.

Right? I'm gonna skip the bulk of this for just a moment and go right to the appendix to explain what happened behind the scenes. I asked the agent to do a review of my detailed income statement.

So the very first thing that it did was run an income statement. This is where it started.

I also gave my agent a materiality threshold. That threshold is going to trigger my agent to drill into data based on the variances that it finds and the variances that meet that material, material component.

Right? So the next report that it built was around OPEX accounts. It saw from the perspective of my p and l that it had variances in OPEX.

So it wanted to drill down on those. So the second report that it built was all around the child accounts for OPEX that were not represented at the level of the p and l.

And then it said some of these variances are interesting and have a relationship across other dimensions. So I built another report that pulled cost center into the conversation for variances that it found in my compensation and my marketing and advertising accounts.

I never instructed the agent to go into cost centers. But because it understands the OneStream application and it understands relationships that exist across dimensions, it intuitively knew that if it finds variances in these accounts to pull cost center into the conversation.

That's why when we actually look at the details of this review up above, it starts at the high level and drills into those details. Right? Talking about gross margin and cost of goods sold, getting into OPEX, getting into compensation, and then getting into some of my cost center variances.

Down at the bottom of this, it's also highlighted for me exception flags and follow-up criteria. These follow ups are intended to give our users a place to run.

Right? They're guiding our users on where we need to maybe follow-up and look into more detail behind any of the variances that we found in the OneStream application. Finally, I want to introduce the conversation around external tools.

In this case, I'm gonna be using Claude. Within Claude, in the customized section, you can connect to OneStream.

We give you an MCP address. This allows Claude to log in to the OneStream system and access any data in OneStream.

I log in with my user credentials, which means any security attributed to me and the data and provisioning that's associated to what I have access to also persists through MCP. I'm not going to be accessing data that I don't have access to in the OneStream system itself.

Once you log in, you're ready to start engaging with Claude to pull OneStream data into any conversations that you're gonna have. I have another example here of a p and l variance analysis that's also running on a scheduled basis inside of Claude.

The reason why I find this one interesting is because at the end of this conversation, it also built me a PowerPoint deck. All of the data in this deck cloud surfaced and sourced from one stream on its own.

Right? All I asked was for it to do an analysis of my p and l, and it said we're gonna wanna communicate those results out, so let's build a deck around that. Right? And and incorporating the high level metrics that I found to be important to discuss.

One other example of this is a dashboard. This dashboard is live connected into OneStream.

I have a refresh button here. If I wanna click that, it's gonna go grab the latest and greatest data from the OneStream system.

On this dashboard, I have built a number of charts and graphs and visualization and even drill pads. Right? So if I see variances for net revenue, I can drill down on those.

I can say, okay, which region are driving those variances? Which products, which customer channels are driving the variances that we see from the perspective of net revenue? So by using our MCP and our agentic layer, you can connect your favorite AI tools directly into OneStream system and source the governed and controlled data from OneStream down into the conversations that you're having with your favorite AI tool. Now that's a high level overview of our AI capabilities.

I want to invite Tiffany back up on stage with me to go over a couple of q and a questions, before we wrap up the session for today.

20:45 - 22:22

Awesome. Thank you so much, Logan, for that demonstration.

I'm just gonna go through our OneStream overview, before I get to the questions. So everything that Logan just talked about, the foundation that our agents are built off of, and what that's all about.

So our vision at OneStream is to be the AI operating system for modern finance. And how do we do that is by having all the core performance management processes working off of a single data model that's uniquely unified, infinitely extensible, and AI powered.

It starts with things that an organization must do for things like core financial processes of financial close consolidations, FP and A, reporting and analytics, all essential functions needed to manage a business. And then once you have that foundational piece in place, then you can move into the fast moving operational data to really steer the business forward.

Think of workforce planning, sales performance management, demand planning. And then, of course, to keep up with today's volatile.

Volatile business environment, AI is embedded directly into the workflows our users already use every day and really help predict, the business forward. And so that is our one stream, vision.

So just to get to some of the questions here, one of the questions that came in to you, Logan, is can you tailor how much of the agent asks for input from a human versus automate?

22:22 - 23:17

You absolutely can. Yeah.

That's the that's the goal of personas. Right? Personas allow us to add any sort of rules that we want into the agents to control the behavior and to capture the context that matters.

Like. And what I mean by context, context is one of the most important things to make a.

AI operate correctly. Right? What do I mean when I say sales? For instance, if you have a sales team within OneStream, What they ultimately mean is revenue.

So when a sales team is sitting in OneStream, I would build a persona around the sales team that says, anytime somebody from this team asks about sales, what they really mean is these revenue accounts. Doing that translation on the fly.

So you don't always have to include in your prompt exactly what you mean. We can include a translational layer through those personas to interpret what the person means based on the context and the rules that have been assigned in those personas.

23:17 - 23:37

That makes sense. Yeah.

Having context is so important. That's where the persona settings come in.

Okay. A second one that came in that is a good one is if we have an enterprise cloud account and can access OneStream's agents through cloud, which one should I be logging into to use? OneStream or cloud?

23:37 - 24:05

Ultimately, we're completely agnostic. Right? We wanna give you the tools that you need to complete the work no matter where that is.

So whether you want to use the native agent experience inside of OneStream or connect your favorite AI tool to the system, completely, irrelevant to us. Right? We wanna make sure that no matter where you're working out of, that you have the ability to source that govern and control data from the OneScreen application into whichever processes you're trying to support.

24:05 - 24:22

Right. Right.

We built the MCP layer so that you can work wherever you want, whichever. However you choose.

It depends on your preference. Okay.

Then, I think we have time for one more. How do I show ROI, for an agent investment?

24:22 - 24:52

Well, we actually have a built in ROI calculator in some of the administrative tools that we have alongside our AI capabilities. The way that it ultimately works is we make some assumptions around things like cost and efficiency.

And based on the numbers of interactions and the complexity of those interactions that the users have had, we're able to track those interactions alongside those assumptions and start to calculate savings and efficiency gains around, the usage of your users within OneSuite system.

24:52 - 25:36

That's great. Yeah.

I think one of the biggest questions and challenges for finance is justifying that, AI or Genesica finance investment. So having that ROI calculator built in directly in the control tower is huge.

That's great to know. Alright.

Well, unfortunately, we've run out of time to answer all the questions, but the team will follow-up with you individually afterwards. If you did enjoy what you saw, please stay connected by scanning the QR code or going to 1stream.

com. Go to our website.

Check out our blogs. We post quite regularly on all topics, not just AI or Gentic AI.

And then we have events posted as well for hosted events that are happening all around the world. So, again, Logan, thank you so much for your time and your demonstration, and thank you all again for joining us.