IGHS68 - Navigating the Future of AI Governance with Inna Tokarev Sela

Jim Merrifield (00:00.941)
Well, hello and welcome to the InfoGov Hot Seat. I'm your host Jim Merrifield and with me today is Inna Tokarev Sela at illumex.ai Welcome, Inna!

Inna Tokarev Sela (00:10.536)
Thank you, Jim. Happy to be here.

Jim Merrifield (00:12.759)
Yeah, it's great to have you as a guest on the hot seat get to know you a little better. And so let's do that. Would you please provide a brief introduction of yourself and your current role?

Inna Tokarev Sela (00:23.838)
Sure, Jim. So, Inna Tokarev Sela and founder of illumex.ai illumex is four years old and prior to starting company, I was with Sisense, a business intelligence vendor as VP AI. Basically, bringing a natural language tag. This was the thing before Agentic to business users. And prior to that, in SAP, a huge enterprise software company.

I got lucky to be in those key positions of leading SAP HANA Cloud Platform, building lifecycle and then PLL, Video Analytics Unit. So basically, really working with the biggest enterprises in the world and solving their cloud and machine learning issues.

Jim Merrifield (01:06.765)
That's great. You got a great background with SAP and Sisense as the VP over there. I'm familiar with that company. It's kind of a competitor with like Power BI, right? With Microsoft, very similar.

Inna Tokarev Sela (01:18.75)
you

Inna Tokarev Sela (01:23.186)
That's a true, that's a true, but they were early in this approach of actually embedded experience. So right now we have this wave of embedded analytics where you're in Salesforce and you see like the block part of and then you might be in your customer success tool and you have different smartness. So basically they were early on to be where users are. Yeah. Thanks.

Jim Merrifield (01:47.171)
That's great. That's great. awesome. Well, let's talk about your company, your current company, illumex What inspired you to start your company and how has the vision for the business evolved over time? I know it's been only four years, but I'm sure it's changed over those four years.

Inna Tokarev Sela (01:52.572)
Yeah.

Inna Tokarev Sela (02:06.906)
I must say, know, my please keep me true, the vision hasn't changed over the years, but it was pretty interesting process to explain investors in 2025 what's automated context and reasoning. Right? So in agentic right now, generally AI and agentic and everyone understand that to capture your own...

meaning business logic of your organization and being able to have a agentic work loads around that does require to capture and document your business.

reasoning into also certain systems. So basically what Illumex does is to capture business logic from existing applications and data sources. about those huge data lakes, warehouses, databases, SAP, BI tools. We capture this logic and we automatically convert it to context and reasoning, which is a machine readable, which means LLM can understand that and interact with that. So we ground LLMs in organizational systems.

Single source of truth on one side. On the other side, we connect to whenever users are in their teams and their Slack, they can really use us within any applications they already have, even their browser, and build their agentic workflow, So just use us as data copilot.

Jim Merrifield (03:30.137)
That's awesome. So you bring the AI to where people actually work on a daily basis.

Inna Tokarev Sela (03:31.022)
and

That is true, that is true, but it's a little bit more than that. So, and in the context of this podcast, I believe that only having single source of truth and monitoring and putting guard rails around that will allow us to actually make decisions based on agentic analytics and access to agentic and overall. So it requires a

Being able to monitor quality of data and ensure the data is good and data which is going to agentic and then have governance guard rails humanly moderated. I hate when people say like, you know, all agentic is a black box just Rag right? This three letters which are more in...

entertaining that I would say useful. But anyhow, like lots of people use Rag too to basically as a black box throw in different definitions and say, okay, this is most of my definitions and I expect agentic to answer based on those definitions. How would you know?

Like it's half baked, 80 % changed. So basically it's a lot of assumptions and lots of estimations around the Agentic. And I think this is precisely what they need to take out from Agentic to be useful, especially in enterprise. So we actually developed governance guard rails. We're pulling out all Agentic definitions and clear business philosophy and bring it to domain experts and governance.

Inna Tokarev Sela (05:01.168)
stewards, right, to be able to certify and confirm all those definitions which go to agentic if they want to, right, based on priorities and so on so forth. So kind of opening up the black box for everyone to see what's in it.

Jim Merrifield (05:15.427)
That's great. That's such a great point and a nice segue into the next topic around information governance, right? There's so many opportunities for AI. I know you touched on a few of them, especially with agentic AI. But what do you see as the biggest challenges for AI and information governance in the tech industry? I won't even say the next five years. Let's just talk about the next couple of years.

Inna Tokarev Sela (05:35.678)
you

Inna Tokarev Sela (05:39.312)
Yeah, totally. So everyone thinks a lot about ethics and bias.

And of course, it's super important because we need to basically moderate what's going into automated algorithms to be able to understand if they're biased, they skew to some extent. there are tools which already allows to do that on the data side, basically understand that the data sets are balanced. Additional concern is the IP protection. So nothing is leaking out to, you know,

models, foundational models, all kinds of providers and so on. So first for that, definitely lots of things could be done. So for example, monitoring what data is switched and using access controls to basically limit that. But I think what's most overlooked part of governance is actually if the business logic is matching to what user actually means. So users are supposed to use a free text language to

to interact with the Agentic. But do we govern how exactly those definitions, terminology basically that they use, match to organizational definitions. So for example, I would like to ask about how...

how many returned customers did we have last month? And I guarantee that it will be at least two systems in organization which might have slight deviation in how those returned users are calculated. Maybe your product usage system and your CRM system and so on and so forth. So at the moment if you just use agentic out of the box, your question is randomly goes to one of those definitions and gives you an answer.

Inna Tokarev Sela (07:26.622)
This is the biggest governance fault at the moment, not to be able to govern that the answers that you are getting actually match the terminology that you used for the right definition. It's totally black box at the moment. And I think this miscommunication or lack of signal source of truth and alignment is going to bite us very hard in the Agentic. And this is where lots of companies still struggling.

I think the biggest part here is putting agentic practice as a silo, right? the companies usually already have governance practices established, know, this frameworks and policies and mechanism and guard rails and monitoring established for data analytics practices. But somehow agentic becomes a new discipline, which is doesn't really adhere to the same protocols. So to the same reporting line and the governance is not embedded part of it. This is why we're actually losing everything that's

we accomplished in 20 years of data management analytics. For those practices, we actually lose a big time in agentic. Second part of governance, which companies are really struggling with, is financial governance. And I would stress that it's part of the governance. So being able to actually predict the spend on agentic costs and even predict the ROI.

for all specific business case of equations to be calculated with agentic is pretty nascent at this point. A majority of companies overspend and have really hard time to align their future spend to the actual business value.

Jim Merrifield (09:05.049)
Yeah, 100%. I think everybody's just, you know, buy, buy, buy and buy mode. You know, let's just buy everything, test everything. And I was actually asked that question, you know, what makes the ROI for an AI investment, you know, at a firm, at a corporation? And to me, it's really the bottom line, right? Is the addition of this GenAI tool or application

Inna Tokarev Sela (09:06.398)
you

Yeah.

you

Jim Merrifield (09:34.295)
going to drive more revenue to the company and also empower us or enable us to provide a better or more efficient customer service to our customers. Right. I mean, because at the end of the day, if it's not generating more revenue for the client and, the organization, you know, what's the, what's the point, right? I hate to say it like that, but what's the point, right?

Inna Tokarev Sela (09:56.392)
other pie.

Always. So yeah, I would say efficiency use cases are simpler to calculate. So for example, in customer support center, if you have a shorter waiting time and higher resolution rates, it's easy to calculate the monetary value for the added revenue. It's something that sometimes you can really calculate. So for example, we have customer which is using Agentic for the third party brokers for self quotation And till then, they had two weeks process

to prove the quotation to the end customers. And conversion rates were about 50%. With immediate self-service quotations, achieved 70 % conversion rates and higher satisfaction. So it's kind of a very measurable impact. But sometimes when you do not really improve additional, like you do not improve existing workflow, you actually introduce new one, it's kind of hard to calculate really.

Jim Merrifield (10:56.471)
Wow, yeah, that's amazing. I'll tell you, I'm learning so much on this podcast already. So let's talk a little bit about your team. So I know, you know, every question, right, with with organizations, I'm sure when you're dealing with new customers, they're like, how safe is this solution? You know, how do you develop such innovative solutions while also making sure that it's safe for the customer and the company as a whole?

Inna Tokarev Sela (11:14.238)
Yeah.

Inna Tokarev Sela (11:23.902)
Yeah, very good question. think for every company which deals with agentic and data, it's very important to actually have the high standards from day one almost like SOC 2, ISO, GDPR and all of that. Like it's a good practice to actually have those guard rails in place. But to us, it was hard decision to say, what, we understand that agentic is...

You know, dangerous to some point, right? It has some risks in that. And then for this purpose, we want to actually separate data and metadata. We want to build our agentic context and reasoning based on metadata only. We want to separate business logic and, you know, business workflows from the actual data values. So when customer actually invokes our platform and applies it to the data landscape,

they do not really exchange any data with us, only metadata. And they only see result and answers in their own environments, which means they do not touch their data or results at any point of time. So it's kind of being very mindful about what's important to a bank which uses agentic on their financial data and be able ahead of time to calculate those risks and meet them where they are. I think it's super important.

Jim Merrifield (12:44.449)
Yeah, excellent. I think that's so important. So, you know, thanks for the overview of the company and, you know, all the innovation, all the compliance, the risk, and obviously talking about how agentic AI is really changing the game, so to speak. And that brings us to my next question. I know you'll be keynoting the ARMA Infonext conference in Savannah, Georgia at the end of April. And I think the topic is around agentic AI.

Inna Tokarev Sela (12:49.374)
Thank

Inna Tokarev Sela (13:14.086)
Yeah.

Jim Merrifield (13:14.099)
So let's talk about that a bit. What are you planning to share with the audience? Obviously we don't have time to talk for 60 or 90 minutes about the topic, but can you give us a sneak peek of what is going to be talked about?

Inna Tokarev Sela (13:24.807)
Yeah

Inna Tokarev Sela (13:29.566)
Yeah, totally. Well, Jim, you have to show up for starters. I'm not giving you a summary over here. Please do show up. But anyhow, it's going to be lots of emphasis on how governance could be built into agentic practice.

how agentic practice actually starts with data, right? Everyone's speaking about the AI of the data, what it actually means and from the governance perspective as well. And then how governance could be actually intervened in any agentic interaction. So this is more like on technology side. And then I will touch widely this topic of what's the role of modern governance person, right? So governance practitioner, is it the new CISO?

the GRC going to be revamped. So what is the organizational structure even to support agentic practices in organization? Because if you speak about agentic, if it's new revenue stream, is it CEO, is it chief revenue officer? If you speak about, again, reorganization of security practices coming from IT to the business side and having more like a leadership position.

what's going to be. So it's a lot of chance. There are a lot of chance in this area and there are lots of new position names. So I'm going to review a few of those trends. Anyhow, without governance there is no agentic.

Right now, I think it's kind of picking up this trend and it's going to be prevalent in three to five years. would say that agentic is going to be commoditized. So foundational models are going to be highly non-differentiated. But the way that you apply it to your organizational business and data is going to be super important and the governance is going to be the biggest part of it.

Jim Merrifield (15:29.783)
I love it. can't wait. You got me excited. I can't wait to hear that keynote, to meet you in person. I know we've been having some prep calls over the last several weeks. again, looking forward to seeing you in person in Savannah at the end of April.

Inna Tokarev Sela (15:30.758)
Yeah.

Inna Tokarev Sela (15:47.12)
Likewise, likewise. It will be interesting to really hear what practitioners have to say. Everyone coming from different industries. My main outlook is coming from financial services, insurance, pharma, retail and software. We also have some footprint in Oil & Gas and utilities, so it will be really interesting to understand how different verticals advance with their controls and governance practices and agentic approach.

Jim Merrifield (16:16.621)
Yeah, you'll have all those players, all those players there to talk to, so, and to listen and learn from your expertise. So that's what it's all about. So I know, and we've talked about a lot here. Before I let you go, is there anything else you'd like to share with the audience as a final thought?

Inna Tokarev Sela (16:18.106)
you

Inna Tokarev Sela (16:33.488)
Yeah, I think we should be excited about what agentic has to offer.

And for starters, try to learn at least one tool every week to get exposed to it because I do not expect my people to start using agentic if I do not have this inspiration from my side. And basically, I also taught my kid, he's nine year old, what ChatGPT is. And I do not regret for a single second because what's future is about, about us humans being moderators. So we would rather understand what our tools are.

to be able to moderate them. It's about the future of augmentation and application free. Right now we are learning tools, but in the future it's not going to be tools, it's going to be an ask and results. So we would rather understand what the art of possible, because who really understand what the art of possible could be, navigating this landscape professionally and personally with much higher satisfaction.

Jim Merrifield (17:35.359)
Excellent. Well, pretty sure pretty soon your your nine year old will be teaching all of us about agentic AI and whatever the next, you know, subset of AI is out there. So that's.

Inna Tokarev Sela (17:46.302)
Let's see, let's see. He's all about soccer now, so let's see how it goes. Yeah, that's true.

Jim Merrifield (17:51.661)
Yeah, we're all about the sports when we're young. Yeah, for sure. That's awesome. So great. Thanks so much Inna. And again, for joining me today on the hot seat and to our audience, if you'd like to be a guest on the InfoGov Hot Seat, just like Inna here, all you have to do is submit your information through our website, infogovhotseat.com And thank you so much and enjoy the rest of your day.

Inna Tokarev Sela (18:16.52)
Thank you, Jim. Thank you for having me over.