What Agents Actually Do

Matt Garman has a tell. Ask him whether AI is overhyped, and he laughs. Not because he thinks it's a dumb question, it's just that he's got a fast answer, and it involves the audience.

At HumanX 2026 San Francisco, AWS CEO Garman asked the Moscone Center crowd to raise their hands if they were already seeing positive ROI from their AI investments, or expected to within the next six months. About 70% of hands went up.

"That's the answer," he said. Then he noted the other 30% — and without a trace of condescension, added: "They'll get there."

That exchange set the tone for one of the more grounded, practically-minded conversations of the conference. On a stage crowded with declarations about AI's transformative potential, Garman kept landing the conversation back on the same question: where are enterprises actually getting value, and what does it take to get there?

His answer, across 30 minutes with CNBC's Kate Rooney: agents.

The Shift From Content to Action

For Garman, the first chapter of the generative AI era was about content: summarizing documents, drafting communications, generating outputs. Useful, but bounded. The second chapter, which he believes we're entering now, is about agents doing work.

"Agents are the way that most enterprises — and most companies and individuals, even today — are going to get most of their value out of AI," he said. "I've felt that for the last year, and we're starting to see it come to fruition."

The distinction Garman draws isn't just technical. It's organizational. The companies getting the most out of agents, he said, aren't the ones automating their existing workflows. They're the ones rethinking the workflows themselves.

"When we really see customers get value, they change how they're working. They change how they use these agents. And then they get stepwise function changes in efficiency, effectiveness, new solutions. It's really transformative when you see that happen."

To make it concrete: AWS's own internal deployment of Amazon QuickSuite — which gives employees access to AI across their enterprise data, from Salesforce to Workday to email — has resulted in hundreds of thousands of employees automating parts of their daily work. The metric Garman cited for software developers alone was striking: roughly 4.5x more efficiency coding with AI than without.

His point wasn't that AI cuts headcount. It was the opposite. AWS salespeople currently spend about 20% of their time with customers, he noted, with the rest consumed by pipeline management and administrative work. If agents can flip that ratio, "you can support 4X the number of customers. We would love to support 4X. We don't need to get rid of people."

On the Bubble Question

The AI-bubble question is never not on the agenda at a conference like this, and Rooney asked it directly. Garman's answer reached for the internet analogy, not as a dismissal, but as a useful historical frame.

"Anytime there's been a technology disruption, you see massive investments in VC. Last I checked, the internet still is pretty big. There are companies who won't make it. That is true for AI companies. Not every AI company will make it. The technology will be there."

He acknowledged the reality of valuation risk for individual companies while separating it entirely from the question of whether the underlying technology has legs. For enterprises in the room, the implication was clear: whether or not the AI startup you're betting on survives, the transformation is coming, and getting left behind carries its own risks.

AWS's Strategy, Revisited

A significant portion of the conversation covered a question that has dogged AWS for a couple of years: whether not having a flagship large language model meant the company was behind in the AI race.

Garman cited a Jeff Bezos maxim — "you have to be willing to be misunderstood for long periods of time" — and argued that the strategy was more coherent than it appeared. AWS's bet was never on a single model winning. It was on enterprises needing a secure, flexible platform that could run any model, with their own data, guardrails, and tooling intact.

"That wasn't necessarily a universal thought three years ago," he said. "Over the last three years, increasingly more people started realizing — actually, that is a pretty good vision."

The evidence for that bet now includes $50 billion invested in OpenAI and $8 billion in Anthropic, not as a contradiction to the multi-model thesis, but as an extension of it. AWS sees OpenAI and Anthropic as two of the leading model providers in the world, and both investments reflect the belief that the platform value lies in giving customers access to the best available models, not in picking one and locking in.

Garman also previewed a joint collaboration with Anthropic on a next-generation stateful model architecture — one specifically designed to make it easier to build and run agents that can take advantage of state in enterprise applications.

The Talent Question

One thread that ran through the conversation was what Garman called the talent dimension of AI adoption. AWS hasn't made the splashy AI acquisitions that Meta or OpenAI have. His argument was that mission and scale — not acquisition multiples — are what attract people who want to do consequential work.

At $142 billion in annual run rate, growing at 24%, the opportunity at AWS is not hard to articulate to candidates. But his more interesting point was about developer productivity. Kiro, AWS's coding agent, recently handled a customer bug request from receipt to published fix in a window he described as previously impossible. "When you're unblocked like that," he said, "you can think much more broadly about what you can accomplish for your customers."

The framing wasn't that AI is replacing engineers. It was that engineers can finally build the backlog they've always had.

What It Actually Takes

The most durable takeaway from Garman's session wasn't a prediction or a product announcement. It was a disposition toward enterprise AI adoption, one that treats agents not as a feature to bolt on, but as a reason to rethink how work is organized in the first place.

The 70% of enterprises already seeing ROI aren't necessarily using more advanced technology than the other 30%. In many cases, they've just changed what they're asking the technology to do.

That's the inflection point. And according to one of the people most responsible for the infrastructure it runs on, we're in the early stages of it.

Watch the full session on-demand. 

Register for HumanX 2026 Amsterdam (September 22–24, 2026) or HumanX 2027 Las Vegas (March 7–10, 2027).