Forget tokenomics – agents are a personnel cost

If an organization is surprised to see IT costs explode when it starts running agents at scale, the problem might not be in the IT budget – it might be a category error. At least in the opinion of HPE CEO Antonio Neri.

“I don’t think of [AI agents] as an IT cost,” Neri tells ITPro at HPE Discover 2026. “I think about the cost of the workforce because, to me, an agent is no different than any other employee I have to hire.”

“So if I hire [a person], there is a cost to onboard him and make sure that he's productive by giving him the tools and everything else,” he says. “When I train an agent, it's going to cost me a number of tokens to train it to drive the best productivity. If I'm going to spend a million dollars to train an agent, it has to be way more productive than [the person]. Otherwise, why am I doing that?”

This turns on its head much of the conversation around agents, tokenmaxxing, and costs as agentic AI starts to eclipse generative in organizational use.

AI sticker shock

ITPro wrote in May about “tokenomics” and so-called sticker-shock when the AI bill finally came due for many businesses. Famously, Uber used up its entire annual AI budget in four months and has now capped the amount each employee can spend on agents per month.

One answer that’s been given to the problem is a shift to on-premises IT rather than ‘cloud-based’ AI like Anthropic’s Claude Code or OpenAI’s Codex, and moving agent costs from OpEx back to CapEx. For an enterprise hardware company like HPE, surely this should be an opportunity to increase hardware sales?

“Sure,” says Neri. “But we think about it as a total cost of [the] workforce, not as an IT expense.”

And is this something that he’s educating his customers about, or do they see it another way?

“Yes, I am,” he says. “I’m spending more than 50% of my time with our customers, and when I bring that concept to life for them, they say ‘ok, ok!’”

“I’d rather give a little bit more budget to IT to reduce my entire cost of the workforce, which is – depending on what you do – probably one of the biggest cost levers,” he continues. “I have only two large cost drivers: people and materials. The materials, in networking, we have a lot of that under control because we have the IP, but a lot of servers, obviously, CPUs, and GPUs I don’t control.

“[The people] I have 65,000 employees. It’s not about reducing employees, which in some cases will happen, but it’s about making those 65,000 employees way more productive with the aid of AI by freeing up capacity for the humans to do more value-added [work] while the agentic models do more work at much bigger scale. I think about that, the total cost of [the] workforce”.

Human / AI interaction

This is an idea that’s been seeded here and there throughout the conference. In Rami Rahim’s keynote, which took place just a couple of hours after Neri’s, one of the demonstrations floated the idea of taking humans out of the loop once a process has been approved.

Katrina Pickett, senior product manager, Juniper Networks, showed how HPE Marvis might help detect and remediate a problem with a sluggish network. The conversation between person and machine was in natural language, but the fixing was agentic. At the point that the fix was applied, Marvis asked if Pickett wanted to have an agent apply the same fix automatically in the future, without human intervention, which would present a time saving.

Human-in-the-loop (HTL), however, has been a key mantra of not just AI but automation more generally. Is it time to take another look at this attitude?

“Well, once you’re confident about [what’s happening], then you can remove the human from the loop and just put on the loop,” says Neri. “I think there needs to be some sort of oversight, but not doing the actual work.”

“In the case of networking, a lot of things were done manually through [the] command line interface, by logging in the console, and putting hands on a keyboard,” Neri continues.

“The scale is now so big that humans cannot possibly manage that scale. Therefore, there is no other choice than to automate things and use artificial intelligence to manage massive amounts of data.”

Naturally, there are caveats. The story of the company that had its entire codebase and all backups erased by a rogue agent is quickly passing from fact into legend; a cautionary tale for any business thinking of letting agents roam free in their network.

“I always say, you don’t get what you expect, only what you inspect [so] I think there needs to be some sort of oversight,” Neri says. “Human beings make mistakes, agents may make the same mistake, right? But as you’re doing it at a bigger scale, it may have a huge impact.”

Mistakes happening are “inevitable”, says Neri. Ultimately, though, there has to be an element of trust that the technology will work as intended. Guardrails can help with this, and, from Neri’s perspective, what HPE is doing with GreenLake Intelligence and Marvis Actions means customers can deploy agents with greater confidence to whatever extent they wish.

HPE still makes everything it always did – it has storage appliances, and it has servers a-plenty, and they’re popular too. According to the company’s most recent earnings call, traditional server orders are booming, growing triple digits year-on-year, while Alletra Storage MP orders also increased triple digits.

The conversation, however, will always revert to networking when it comes to visions of and for the future.

“There was a question in a previous session [that was posed] to Rami (Rahim), which said how far are we to really being fully autonomous, and we believe it’s within the next 12 months,” says Neri.

“It’ s not going to be everywhere across the network, but there are many aspects of the network just running itself.”

“Our opportunities are in the three domains: networking, cloud, and AI. Networking is a huge opportunity; we see it is. And it’s driven both at the edge to the campus and branch, the on ramp into the AI. You need that connectivity.”