You would be hard pressed to find a Fortune 500 executive who will publicly admit that their company is struggling to scale AI. Privately, under Chatham House rules, that turns out to be roughly where most of them are.

I recently spoke to Ed Nelson , co-founder and strategy director of the AI Infra Summit , which, prior to their flagship event this month, convened a closed gathering of Fortune 500 executives. The participants were at executive management committee level or one level below. The purpose of the gathering was to discuss what the money being spent on AI was actually buying and how to scale AI implementation.

What struck Nelson was that which was not of interest to this group. Almost nobody wanted to argue about which model was best, or whether the technology could do what was being asked of it. The conversations were about absorption. The real constraint, as he puts it, is “organizational metabolism.”

Implementing AI at scale remains an obstacle.

A number of enterprises had successful AI projects. For example, Nelson describes an enterprise software company that replaced around 250 sales dashboards, used by 4,000 sales people, with a single agent now handling up to 30,000 prompts a week. One dashboard it retired had been costing $3 million a year. Another company implemented an agent to optimise tax payments, this created savings of hundreds of millions of dollars.

A third, in the health sector, used generative AI to summarize tens of millions of customer service calls and identify which messages actually persuaded people to move to a lower-cost treatment. According to Nelson, that produced a margin improvement of $170 million in one part of the business, alongside $200 million in savings passed to customers. In my own company, we recently put agents to work on understanding data compliance across every country in the world, it was only afterwards we realised the time savings and capability uplift.

In discussing this with Julio Martinez, CEO of Abacum , which provides a financial planning platform for businesses, he told me that “the biggest mistake is assuming that automating finance work creates a more strategic finance function. It does not. AI can make existing processes dramatically faster, but the real opportunity is deciding which of those processes should disappear altogether and where finance should spend its time instead.”

When the Fortune 500 executives were asked to describe their own organizations, the most common answer was that whilst they had a few live cases where AI delivered measurable value, for the most part, AI was not embedded in their core products and strategy.

So why is AI not being used more in live operating models?

Companies deliberately wasting money

Some are trying to buy their way through the problem. Nelson describes an enterprise software company that built approximately 18,000 agents in a single quarter, expecting only a few hundred to prove genuinely useful. It ran well past its cloud budget and had planned the overrun in, on the theory that mass experimentation teaches an organization faster than any carefully governed pilot will.

Whether that is discipline or indulgence is a fair question, and Nelson notes that spending has since shown signs of being reined in at some large enterprises.

The harder problem is to change mentalities. Nelson describes a healthcare company with an operational process running to roughly 200 individual steps. Managers proposed automating three of them. The executive argued that with AI, most of those steps did not need to exist at all. Automating inefficient processes does not transform companies.

Everyone becomes a manager of agents

Agents can be multiplied in seconds. People cannot. Technology improves week by week but the pace of change is limited by the behavior, confidence and skills of humans.

The responses Nelson observed were mostly attempts to overcome human inertia and limitations. One company now requires an AI-related objective in every employee’s performance review. A CTO described asking software engineers to set aside the craft they had spent 10 years acquiring and orchestrate a set of coding agents instead. Another paired interns who had grown up using AI tools with experienced engineers who understood the business, hoping something useful would come out of the cooperation.

What none of them regarded as sufficient was a technology function acting alone. This echoes last month’s Forbes article and the need for change being led from the top. Nelson recalls one executive’s comments to IT leaders who turn up having been charged with getting AI into the company by themselves: if you are the only person on that journey you might as well start getting your resume together. Nelson describes the shift as “less of a CIO implementation agenda” and more a question about the operating model. This switches the conversation from a technology to a strategic (and therefore CEO) conversation.

The advice being offered to people entering the workforce points the same way. Nelson says the consistent message was that, as narrow AI will handle much of the domain expertise, generalists will prevail; architects, lateral thinkers, designers, evaluators.

Work is moving from performing every step of a process to supervising, validating and redesigning processes that agents execute. Many employees will become managers of agents. So do these Fortune 500 leaders know howto turn their employees to supervise AI? Rather worryingly, Nelson, who ran a panel dedicated to that question, was blunt about the answer he got: “I didn’t hear anything very satisfying on the upskilling side.”

For years I have argued that the limiting factor on AI is not the technology but the organization around it. Building 18,000 agents is one way to discover that. It is not a cheap one.