Deepak Bharadwaj, Workiva | Workiva Amplify 2026

Welcome back to Amplify 2026. We're coming to you live from Las Vegas. I'm Allison Casic alongside Christa Casease and we're going to delve into the fact that AI agents seem to be everywhere right now, but in all of these other important areas, um being fast and capable isn't necessarily the enough. It there needs to be right. It needs to be correct and trustworthy. >> Absolutely, Alison. And we need that trust for sure. >> Yeah. Let's dig into it with our next guest, Deeppac Baratch. He's the EVP of and chief product officer here at Book Wara. It's a mouthful. Welcome to the cube. >> Thank you, Ellison. Hello, Chris. Good to meet you all. >> Great to have you on the cube. So, Deepek, it seems like every every enterprise um software company is announcing AI agents right now. What's different about building AI for financial reporting, audit, and compliance? >> Great question, Alison. This is something that um I get asked every day and it's really about the risk profile of our buyers and users and because of who we serve in financial reporting and GRC and uh sustainability what they want is the work that they do can be trusted uh because ultimately in many of these cases somebody is signing their name to it and so really the opportunity for us uh and the work that we are doing in our products and solutions is making sure that not only are agents operating on what we call trusted data because once the data comes into our platform it is certified it is trusted it is governed and so on and so forth but the AI that is working on the trusted data is also producing output that can be trusted so can you trace the output of AI uh is the output of AI or the work of AI defensible in terms of who did what and when and then uh do you have audit trails because we want everything that happens within our platform to be audit ready as well. So that's really the difference in terms of how we are looking at AI agents, agentic applications um and that's where you know our customers expect us to go as well. Yeah, thank you for that context um Deepo and I had the opportunity to listen to your your keynote this morning um alongside your your CEO and really that was kind of one of my big takeaways was really looking at work knowledge in particular and kind of the ability of work to potentially provide that grounding that data lineage those controls and auditability and also the human escalation to make sure that there are those guardrails for you know human execution. you know, almost kind of this cognitive surface um to be able to provide that kind of control layer. So, I guess when you think about your product development, you know, does that ring true for you in terms of where you think we're heading? >> Yeah, absolutely. And there's a a few different pieces to that, right? One is you mentioned guardrails, right? So, that's absolutely a big piece of uh what our customers want and expect. So how can we make sure that the AI or the agents don't go off and do things that they're not supposed to do? Uh then there's this notion of context. Knowledge is one of those ways in which you can uh create context and think of knowledge as um kind of a a set of instructions that you can provide to the agents so that as they are doing their work uh not only not only are they um respecting the guardrails but they're also following some of these instructions. so that the work work or output of AI is according to the company's policies or standards and things like that and then you have all of the data like I said that exists within our platform which is trusted data and so when AI operates on that trusted data with the context that is provided by knowledge with the guardrails that we then impose and then we have all these tools that we allow humans to get in and supervise and uh verify the output of AI whenever and wherever they want to. Right? So, you take all of these mechanisms and that's what really puts together a platform for uh trusted data and trusted AI. >> Well, I'm curious where should an AI agent, if at all, ever be allowed to act on its own and it probably can. And where where should it stop and ask for human approval based on what you're saying? I I think a good way to think about it is something that is fairly repeatable and it's known in terms of how a task needs to be completed is a good place for an agent to to do its own thing. And usually these tasks are not what I call rules based, right? And so there are instructions that you and I can describe in plain language. An agent is great for that and it can follow instructions and execute. But if you have tasks that require business judgment or someone needs to look at it and approve and sign off, that's where humans have to come in uh and make sure that what the AI is producing is correct. Right? So for example, in our case, if I'm rolling forward, which means I'm taking previous year's filings like a 10K or a 10Q and then using that as a starting point for my next quarter's filings, that feels automated and you can just let an agent rip and complete that. But if it's producing a disclosure, right? So you got these sections within the 10K that uh someone has to actually think about whether that is an accurate representation of what you're trying to disclose or convey. You don't want an agent to produce that and it show up in a 10K and that's where the business judgment is required and oftentimes that's where multiple people will argue about one word in that disclosure >> right and that's just something that at this stage to your point Dupac we just can't trust AI with but for those actions that we are trusting AI with you know I have been really glad to hear today um the themes around kind of the provability of the AI agent or the explanability So from your perspective, how do we more tactically go about proving why an AI agent took an action? >> You mean proving the output of AI agents? >> Yes. >> Yeah. Um, you know, that's where things like verification start to come in, right? So you can as a human go in and you can verify the output of AI. It takes you back to where the answer actually came from. It shows its work. And so you can then as a human validate that it did the right thing. Now what is likely to happen is as we get more and more comfortable with what the AI is doing that level of control will change. It will it may reduce in some cases you might decide that you need an extra pair of eyes uh for certain things that are higher stakes. But really from a platform standpoint, what we want to do is provide the toolkit that anyone can use to either uh run our agents that we ship out of the box or build their own agents with uh agent studio which is also something that we announced today. >> Yes. And on on that point about launching um agent studio um here at Amplify that what that you announced, why is it important do you think to let users build their own AI agents rather than relying only on agents created by software companies? That's that's the uh best part about AI, right? Um if you think about the way we have solved for use cases that our customers have had, we've we've made platforms extensible traditionally, but in order to do anything useful, you needed technical skills. So we have platforms with APIs and the ability to do some scripting and Python and and whatnot, right? Highly technical. Our business users, they really understand their business and with AI now they can just describe what their business process is and then let our agent studio which in itself is an AI application create AI agents that will do the work for them. And so then what you get to is our customers can run their business the way they want to run it, their processes the way they want to run it. uh it's it's going to be impossible for us to create software for every user at every customer in every industry vertical in every part of the world. And what this lets our users do is not be restricted or confined by what our software can provide but really leverage our our extensibility toolkits whether it is agent studio or MCP and then uh do what they want to do on top of our platform. So they get trust and they get extensibilities. >> Yeah. Yeah. And sort of is almost balancing between um providing that governance layer and also kind of facilitating that intelligence because as you mentioned you know through those kind of MCP extensions. >> Yeah. And we always make sure that actions that are happening within the platform are are permission enforced. So AI is never doing things on its own. It's doing things on behalf of whoever is supervising it. And you know ultimately someone is going to be accountable and like I like to say you cannot send an agent to prison. >> It's true there has to be that human accountable at the end of the day. Um and along that vein Deepo you know you were kind of talking this morning in the keynote around the multi- aent orchestration and I'd love to double click on that with you because I think that's going to be a big problem for enterprises to solve moving forward. um because we have these potential agent swarms or these chain of agents that are you know either collaborating together or triggering action from other agents. So can you talk to I guess you know from your perspective you know where stands now I guess in terms of really solving that for customers and what we might think about looking out for over the next 12 to 18 months. you kind of answered part of that question yourself, right? So guardrails is a is a very important piece to that. And so we have the guardrails. So even though you have multiple agents doing uh whatever they're trying to get done, those guardrails come in place, uh we've got the ability for humans to engage at the critical decision point. So this is not just giving them a goal and letting them rip end to end. Humans can come in and they can supervise and make sure that these agents are doing things under their control. uh but it's it's really about ultimately what what are we letting agents do within the platform right so they don't really have this unfettered access to go off and do anything so there's a specific set of what we call capabilities that I demoed within agent studio and that's the list of things that they are actually able to do within our platform and so that's the way we we kind of manage what these agents are able to do um so yeah we want swarms of these agents and them orchestrating but they really have this umbrella of governance that is going to uh control what the output is or what the work is. >> Absolutely. And as you were kind of mentioning earlier, that's kind of what differentiates an agent from answering a question or taking a single action to, you know, like you're alluding to being being able to, you know, complete a process like, you know, filing a 10Q or a 10K. >> Yeah. To totally and an end to-end workflow is is what we ultimately want to make sure we can orchestrate. uh and whether that is by automating pieces of it or by making sure that humans can come in and do their work, collaborate with their teams and get assistance from the agents. >> Do you think AI could eventually allow audit teams to get greater uh risk coverage without simply adding more people? >> Um yeah, I mean that's that is the that is the ultimate goal, right? And so when we start to automate things like um I spoke about uh we spoke about automated uh uh testing as an example one of the pieces that we want to automate is take the what is called evidence right so think about purchase orders invoices these come in as PDFs and spreadsheets and it's all unstructured and so if you're doing this manually what typically happens is you cannot look at every transaction you you pick like 10 15 100 transactions. Then you look at it manually and and you see if the purchase order number in the invoice actually match the purchase order number that was approved for that invoice. And you're doing this manually. But if you start to take that unstructured content and structure it, then now you can run this testing over a larger sample because now you're not doing this manually. So now you have really automated coverage for your testing by taking AI to take unstructured content and then structuring it. So that's how you get broader coverage. Uh and then the the the folks that are doing the testing you know the auditors they can actually focus on how to go off and remediate things that are not working and that's where their judgment comes in. So less manual work more on judgment more on decision support. It's an important point, Deepo, because when we think about AI, I think a lot of times we think about speed, which certainly AI allows us to move faster, and I know you kind of again illustrated that in some of your demos this morning. Um, but what I'm hearing from you is that it also potentially helps us to reduce our risk, improve our compliance because we are able to offload some of those tasks to AI, move faster, and then also, like you say, really leverage that human judgment more. >> 100%. And yeah, I I like to think of it as layers of Swiss cheese, right? So no layer is 100% perfect. Like if you think about human review, it's not 100% perfect. Uh AI, it has its own gaps, but you start to layer enough of human and AI and multiple checkpoints, you get enough coverage, and that's how you improve trust within the uh system. >> What do you think will separate the companies that truly get value from AI from the ones that simply, you know, just have a lot of AI tools? Ultimately, it's about the customer, right? And the user uh and and so the companies that can figure out what are the problems the users are trying to solve. How can AI which is a means to an end it's not the end. How can AI enable users to become more productive, more efficient, produce higher quality work? Uh those are the companies that will will succeed versus if you start to think about AI for AI sake then you end up with product bloat. That's what you get. Uh so that's what we are focused on. We're making sure that our early adopters take everything that we've announced today and we have a number of programs that will uh put them through the paces. We want to help them on their transformation journey, but when they start to adopt, they'll provide feedback. We'll make it better and awesome and that's how we get to a point where users get what they expect from work. >> Absolutely. Deepac and is there anything in addition to that customer feedback? Is there anything else that kind of your team is leveraging when it thinks about um when your team thinks about you know kind of the product portfolio and how like the selective areas that you do want to integrate AI? >> Yeah. So customer feedback is a is a big piece. Clearly the technology landscape is is changing. It's changing rapidly. So we're always looking at how do we leverage the latest and greatest in technology and solve problems that we previously just couldn't solve or didn't even think about solving. And LLMs themselves are a representation of that. But each LLM model is different. It's getting better. It's trying to do more things. It's trying to do things differently. And so big part of our work is really mapping what problems we're trying to solve with the right model that will actually solve these problems. And it's not a one-sizefits-all. And that's why uh this is something that we pay attention to and make sure that users get the value that uh that they're trying to unlock >> in in the work product lineup. the unveiling and I know we we all love our children equally. What's your what are you most excited about? >> Oh my gosh. [laughter] Um >> I stumped you. >> I I would say I'm most excited about agent studio and for the reasons I mentioned earlier, right? That is how it's literally a manifestation of some of the tools that we use internally as we build our agents. We are now exposing that so that our customers and our partners can go off and build their agents. So while we'll still continue to build these fitfor-purpose agents, those agents will be you know where our deep subject matter expertise matters where it's doing something very specific and difficult within the platform but the way we'll get scale and the value unlock uh within our customer community is by letting them do things their own way. So, I'm very excited about Agent Studio and the possibilities that it uh it um brings to us. >> Absolutely. And Deepo, do you see potentially opportunities to kind of see what customers or partners are doing with Agent Studio and then perhaps even productize some of that, you know, um kind of use that as another lever for feedback in terms of some of the agents commonly across your customer base. That would be helpful. >> Absolutely. Yeah. I mean, we we always uh work closely with partners and customers. So, that's one way we get feedback. uh we have some instrumentation uh within our platforms. We get feedback. The one thing we don't want to do is see in terms of you know customer data or uh something that is uh you know that would uh uh cross the line on privacy. So we take care uh and and not get into areas where we are able to see customer activity specifically. But yeah, we have these signals that will tell us what's working, what's not working, how can we make life better for our customers and for our partners. >> All right, Decon, thank you so much. I I I know you have to get out of here. I think you have some meetings happening. Thank you so much for stopping by the Cube. Appreciate it. >> Thank you, Chris. >> And you're watching The Cube, the leader in live tech coverage and in-depth expert analysis. We'll be right back.