Building Trust First: Operationalizing AI Agents with a Solid Framework

Hello and welcome to the latest edition of the digital CXL Leadership Insights video series. I'm your host, Mike Beard. Today we're with Kush PE by who is the Senior Vice President for Platform Fundamentals for ServiceNow.

And we're talking about, well, what does it mean to be responsible when it comes to ai? Kush, welcome the show. Thank you, Mike.

Thanks for having me today, and excited to talk about this very important topic with you today. All right. It seems everybody would agree that we should be responsible for ai, but nobody knows exactly how to go about implementing that.

And what does it mean beyond the fact that we all agree that it's a nice idea. So what exactly should people be doing here and what is the process that we need to think about as we apply it to our workflows? Well, Amazing question.

I think last two years have, uh, taught us that AI is super, super powerful but with great power and comes great responsibility too. And that's why ServiceNow has embedded responsible AI into each and every layer of the platform and the products. The way we do it is we have embedded a four key principles into our responsible AI thinking.

One is human centricity, like in each and every layer of the platform, we are adding ai, but we are making sure that it's persona driven and that the human is in the loop when decision making needs to happen. The human is in the loop when it needs to know that this part of the product is done by AI and why it's done by ai. So the human can control if it needs to go autonomous or the human needs to make decisions.

The second one is transparency. Again, in each and every layer you can understand what model is being used, why is it used, what data it's trained on so that they are not no biases. This makes sure that like the accountability is right there into the platform.

The third principle is inclusion. At ServiceNow, we strongly believe that we wanna build amazing products which create value for all our customers, global customers. That's why inclusion is very, very important.

Whenever we are building any AI feature, whenever we are fine tuning any large language model, we are very, very cognitive about what data sets we are using. So that bias or any other things like that don't get introduced into the algorithms or the software that's make sure that like we are building an inclusive software for everyone. And last one is accountability.

How do we share the metrics? How do we share clear documentation? How do we share all the things which customer care for to the customer through the right mediums so that accountability value is in place?

So these are a couple of, uh, principles which we adhere to. And we are also using these principles which are based on the NIST AI Act on top of it. ServiceNow is also member of AI committee in partnership with Meta and IBM where we further these things in the industry.

Like you said, everyone wants to do it, how to do it. So forming these consortiums helps bring responsible AI thinking in the industry and we learn from each other. Mm-hmm.

It seems to me at least that trust evolves over time and it's one of those things that's easily lost and hard to regain. But will I have some sort of ability to dial up the level of trust that I assign to given AI agents? 'cause it may change over time and I may gain more confidence in the AI agents' ability to execute something autonomously.

And this seems like it's gonna be a, a little bit of a dynamic relationship for a while. A hundred percent. I think one of, one of my favorite quotes from our CEO is that like you earn trust in drops, but you lose that in buckets.

And this has been always our principle when we are always thinking about like how we build products, how we deploy them, how we create value for our customers. And your question is really, really amazing. One of the products which we use internally and we have also given back to our customers is AI control tower.

Now we understand that customers will use our AI and third party AI in their environment, but we, what we want customers to have is full visibility on what that AI is doing. Metrics through performance hallucination, and all the other metrics value which it's creating. When you have all that information in the central command center and you can manage your entire AI footprint from ServiceNow to other third party vendors, then because of transparency value, your trust may go up because you are seeing the AI in action.

When you are seeing it in action, you can, you can essentially dial the autonomous city of that AI based on the trust you have. And in some cases, when you see that the AI is not creating value through AI control tower, you understand why it's not happening. Maybe the data set is not correct, maybe the environment is not correct, whatever is the reason you can fix that problem so that you can dial it up or down based on the scenario.

Mm-hmm. How easy will it be to swap out the underlying large language models used to drive some of this stuff? Because I may at some point, uh, lose faith in a particular LLM, it may be updated in a way that, uh, the database or the data underlying it has been tweaked or poisoned, or there just may be advances in one LLM that's faster or better than the next.

And I wanna take advantage of that. So, um, I guess how disposable are l LMS gonna be Another fantastic question. Uh, another uh, thing which we have learned in the last two years is like the advancement in these technologies have been unprecedented.

Every three months we are seeing new LLMs, new models, new modalities coming in, solving problems, which were harder a couple of years ago. That essentially means that like, hey, as we adopt new LLMs, as we adopt new versions of LLMs, we need to be transparent to our users, to our customers. And we do that through model cards in ServiceNow.

We have publicly available model cards in which we tell you what version of a model we are using, how did we train that model, what are the metrics of that model? What could be unintended, uh, biases in that, if any, and all the other information which, uh, is uh, required for someone to make a decision on. Another thing which we do is, like in our SDLC, we have embedded AI governance in each and every layer.

When we are taking any general purpose model, we are first doing research on it to figure out what is the problem statement we are trying to solve. And for that problem statement, which generic model is best to use. Once we have that, we fine tune it with our specific data so that it's giving accurate answers.

Because once you fine tune it to the problem statement, you have in hand with the data, you can essentially control the accuracy of that model. Once we have that, we have super extensive eval, uh, training so that we can validate that, okay, whatever we are trying to do, is it happening or not. Then based on that we deploy it and understand like if there is a, uh, the same quality metrics sticking to each and every geo, each and every scenario, and then deploy it, check the metrics.

And in any case, in the future, if we are deprecating any LLM, we retire it. We are very transparent with our customers about it, and that entire information is available to them in AI control tower two. So this entire SDLC makes sure that there is always transparency and trust built in the platform from ServiceNow to our customers.

Mm-hmm. It also seems to me maybe we need to change the way we think about our relationships with applications and machines. And I asked the question because so many of our processes are deterministic.

They need to be done the same way a hundred percent of the time the LLMs are generating output that is probabilistic and that is, you know, it's, its best guess based on the data that it has available. But it may be flawed, it may be wrong and somebody needs to double check all that. But, um, I think maybe we think too much of the machines and not enough about the role of the humans in this process.

So how do organizations kind of strength the right balance there? So one of the things which, which, uh, we did when we started this generative AI agent journey is like, how can we use the 20 years of ServiceNow investments and take that as the foundation which can drive this agent revolution? So for 20 years, our customers have written automation on our platform, scripts on our platform, created knowledge articles on our platform, built lot of end-to-end, uh, business rules on our platform.

Now, all these things, like you said, are deterministic because they essentially like follow steps A, B, C, D, and they're always starting from the same thing and ending at the same thing. But when you give these tools to large language models, the reason on which tools I need to use in what order so that I can get to the outcome that u is asking me to do. So this brings more accuracy into the platform because our AI is not bolted on top of it.

It's layered in, into each and every layer of the platform. That essentially makes that AI having lot of tools at its disposal. It has access to all the knowledge articles, it has access to all the automation, it has access to all the pro code stuff, which is created on the platform.

And then it's using all of that with the reasoning to orchestrate that across the, uh, platform, east West, not sub. Mm-hmm. Um, there will be agents that come from places other than ServiceNow and we've seen the rise of things like the model context protocol to give, uh, agents access to external data.

And there's also this whole notion of the agent to agent protocol, but how will we extend trust to other agents and, and will ServiceNow kind of now rate the level of trust, not just for its own agents, but the agents that you're interacting with? Uh, another amazing question. One of the things which I called out earlier is like, we built a product called AI Control Tower, and it's a governance tool not only to manage ServiceNow ai, but any AI you have in your enterprise.

And what it does is it's essentially giving you an oversight on different agents which are created on the platform, different tools, uh, which are created for those agents like an MCP server or an MCP client. So this way any enterprise can actually manage all your AI footprint inside an enterprise. The second thing which we built was AI agent fabric.

Now, like you said, there are multiple protocols. There is a two A, there is a two C, there is agency, there is MCP. Now our AI agent Fabric understands all these protocols.

So when two agents are talking to each other, we find the best protocol to talk to that agent and essentially figure out acls so that the right securities maintain data governance so that if you're not, uh, entitled to see certain data, we don't give you access for that. All of this is built into the platform with visibility into AI control tower. That way a human, in most cases the AI governing body in an enterprise is having full visibility on what agents are accessing, what they're entitled to, what work they're doing with full logs into the system so that backwards auditability is also available.

Mm-hmm. So as more trust gets established, if I look at the way we work, we have all these different silos around sales, marketing, accounting, whatever it's gonna be. But if we have agents and ai, well at some point the workflows and the processes that we currently have start to converge more.

And it might be hard to distinguish where one begins in one ends and maybe we'll need to rethink how our entire organizations are structured. What do you think? Oh, I agree with you actually.

I think, uh, this is an amazing opportunity for each and every of our customers to think north, south, east, west. So that's why whenever I am talking to a customer, I first anchor on what problem you're trying to solve. And for that problem, how can ServiceNow help you?

Because as you said, now, if you have a super smart reasoning model, which can understand the problem statement and which can break down that problem statement so that it can delegate it to multiple agents and then orchestrate those agents like a team so that you can get to an outcome, then you're essentially running through various different departments and going north, south, east, west to solve that problem. Another thing ServiceNow has is like in October of last year, we geared a product called Workflow Data Fabric. Now, what Workflow Data Fabric does is like you can essentially tell us where your data is, and rather than the moving that data into ServiceNow, we access that data in place.

And we worked in partnership with Snowflake, Databricks, or Oracle to create these zero copy connectors. This essentially supercharges what you essentially called out, because if I have access to all the data and if I have a central orchestrator which can use all of that data, I can start solving problems in an enterprise which can go through various different departments. Mm-hmm.

So what is your best advice to organizations right now? Or conversely, what's that one thing you think we should be thinking through a little bit more than just simply rushing out to experiment with every piece of AI technology we find? Well, I think I, I, I mentioned that earlier.

I would say start with the big problem statement, which aligns with your mission and vision of your company and figure out how will you measure success for that. And then start using AI for it, because then you will have clear KPIs you can measure if you're in the right direction and you can fine tune it if you are not this way. Like you said, you move away from experimentation and POCs to AI actually deployed in production and achieving outcomes which create value for the enterprise.

So that would be my advice. Alright, folks, you heard in here, if you wanna operationalize ai, maybe we start thinking about trust first and then work our way backwards from there. Hey Kush, thanks for being on the show.

Thank you. Thank you, Mike, for having me today. All right, and thank you all for watching the latest edition of the digital CXO Leadership Insight series.

You can find this episode and others on our website. We invite you to check them all out. Until then, we'll see you next time.