Enterprise AI Governance Moves to Runtime as Token Costs Reshape Deployment

Shiv Agarwal, co-founder and chief executive officer of Singulr AI, said enterprises need to rationalize what they intend an agent to do against what it actually does at runtime. “Someone needs to really do some rationalization between the enterprise intent and the actual runtime state,” he said. “Governance is becoming very important, but it’s becoming hard as well in this world, just the pace at which the AI is getting adopted.” Agarwal said defining an agent’s allowed actions is only part of the challenge; companies also need to assess behavior during execution. Singulr has extended its agent governance to local AI tools on employee endpoints, he said.

Agarwal described the kind of activity that can be difficult to track. “[It’s] as simple as an action where it reaches out to the internet to get some data or sends out an email to a private account or something,” he said. “That could be deeply hidden inside an agentic action.”

Sri Viswanath, founder and chief executive officer of Sycamore Labs, said some chief information officers are responding by narrowing where agents can operate. They want agents and large language model calls kept within their virtual private clouds, according to Viswanath. “They’re also trying to architect in a way that all of the data stays there,” he said. “They don’t trust sending it outside because then we don’t know where it goes.”

Agarwal called for greater standardization across AI governance frameworks, including the European Union’s AI Act. Victor Jakubiuk, co-founder of MisaLabs, emphasized the need to keep pace with rapidly evolving AI capabilities. “AI has no limit to the speed of development, speed of progress,” Jakubiuk said. “As an enterprise, as that AI lead, you want to be able to take the latest and greatest and stay on top.”

The same symposium surfaced a parallel concern about cost. Helen O’Sullivan, AI business development manager and solutions specialist at Dell Technologies, said some customers she supports in the Bay Area are running through token budgets in weeks that had been allotted for months. “You still use those frontier models, but you use the open source and infrastructure on-prem to at least eliminate some of the token stress and token maxing going on right now,” she said. That pressure is reshaping enterprise AI deployment strategy, with on-premises infrastructure and open-weight models returning to discussions alongside frontier models in the cloud.

Tim Wood, sales director for the Northwest region at Intel, said enterprises can evaluate deployment options before committing to production. Testing open-weight models in sandbox environments can help companies assess what their workloads require, he explained. “A lot of enterprises already have CPUs in their data center. There may be existing resources and headroom in the infrastructure they already have that’s capable of doing some of these next-generation apps and workloads,” Wood said.

O’Sullivan said agentic AI is changing the balance of processor demand, with heavier utilization of central processing units for agent workloads, even as many buyers remain focused on graphics processing units. “We saw the rebalance and the heavier utilization on the CPU when it comes to agents. Customers are getting there,” she said. “They’re not quite there yet from the scope I’m seeing.” She added that enterprise AI scalability requires assessing the whole stack, and that customers often ask how quickly they can get a graphics processing unit when the more important question is about the data. Dell and Intel are taking a consultative approach, weighing existing infrastructure and capital expenses against operating expense models before recommending a plan, according to the executives.