Preparing for enterprise-scale AI with agents and tokenomics

At the September 2026 Splunk .conf26 held in Denver, Colorado, I discussed with Hao Yang, VP AI, Splunk, on enterprise-scale AI adoption, agentic AI and tokenomics – among other things:

What can help pave the way for large-scale enterprise AI adoption?

Hao Yang: At this year’s .conf, we are putting a lot of emphasis on trust and tokenomics. That’s really to pave the way for large-scale enterprise adoption of AI.

One of the things we talked about this morning is AI-powered data management. That's actually something that a lot of customers wasn't expecting us to do because everyone's talking about things like Agentic SOC. But AI-powered data management is a little bit behind the scenes.

There is no way your AI is going to work if your data pipeline wasn't set up correctly: it is not extracting the right field from the logs, and you can have AI looking at the data all day long but you're not going to actually find anything! We know that AI-powered data management is very important because data management is a lot of work. It's tedious. It's error-prone. And it's really, really important.

So that's why we have been investing in tools to actually help strong admins to set up the data management pipelines correctly, and also have the self-healing capability when something goes wrong. AI can detect that and give people some heads-up and also recommend how to fix it.

Why do AI costs become more significant as organizations scale adoption? Why is visibility into AI spending important, especially when it comes to token costs?

Hao Yang: A year ago, they might be thinking about AI as experimentation… cost is usually not a big concern. But now you see all the great potentials, and you’re going to roll that out for every employee. Some large organizations can have 100,000s of employees, and when you do all the math, that cost has become something pretty real. We want to make sure that we have visibility into the spending.

A year ago, token costs may not be a problem because if you are using a chatbot and you have some idea about how many tokens you're going to burn – because it's limited by how many times you're going to interact with the chatbot. Now you have agents, and the agents are going to be there doing all the things… All of a sudden, the agent can burn tokens pretty quickly. So, before you know it, it’s already a huge consequence.

That kind of changed the urgency a lot in terms of visibility. That's why everyone is now really looking at that. To me, having that visibility into how you're spending is important, especially when you have all those unpredictable events that in your environment.

What should an organization consider first when building a tokenomics playbook?

Hao Yang: You have to first think about value, because there is no point of talking about costs without talking about value. If there is no value, then you shouldn’t spend a penny in it. If there is a lot of value, then you should really figure out how to fund this.

Next, you need to have visibility into the spending, like I mentioned above.

And the third thing is really to have what you can call guardrails or control. When something goes wrong, you need to have the ability to stop it right away, rather than having to deal with the resulting mess.

How should organizations think about frontier and open-source models?

Hao Yang: I feel like there is a place for both kinds of models. They are shining in different scenarios, serving different purposes. Rather than competing with each other, I think the world is big enough to actually find use cases for both of them.

If I think about some of these AI agents that we build, we actually use a combination of frontier models and open-source models, because to tackle a really complex AI cyber-attack, or to handle some of the most complex IT incidents, we do need very capable models from the reasoning perspective, from the knowledge perspective, from the ability to adapt to the quickly changing environment.

I think frontier models still shine in those types of scenarios, but frontier models have their own limitations.

One is in the air gap environment, where you don't have access to these frontier models, and second is that the frontier models, especially the latest ones, can be very expensive. if you use that in all your tasks, it might actually have, as I mentioned, some tokenomic problems. So that's why we're also looking at open-source models to specialize in some deployment scenarios or for different kinds of tasks.