Exclusive: AWS outlines agentic AI strategy for enterprises

Amazon Web Services is positioning agentic AI as the next practical step in enterprise automation, as organisations move beyond chatbots towards systems capable of handling complex, multi-step work at scale.

The company argues that recent advances in autonomous software agents, combined with custom-built silicon and more flexible deployment models, are shifting agentic AI from experimentation into production environments.

Much of this strategy was on display at AWS re:Invent 2025, held from December 1–5 in Las Vegas, where agentic AI emerged as one of the dominant themes across keynotes and technical sessions. AWS used the event to frame agents not as a future research concept, but as a deployable enterprise architecture supported by mature tooling, infrastructure and governance.

Speaking about AWS's approach to agentic AI, Rada Stanic, Chief Technologist for Australia and New Zealand, said enterprises are now able to deploy increasingly sophisticated agents across regulated and mainstream industries, with a focus on productivity, cost efficiency and improved customer experience.

What is an agent?

"Agents are like autonomous or semi-autonomous software systems that can do three things: they can plan, they can act, and they can reason and then achieve a goal," said Stanic, during a recent interview in Sydney.

"Unlike the chatbot, where you ask a question and you get a response, with agents it's not about answering questions - it's about accomplishing a goal."

Stanic described an evolution from chatbots to agents, and now to what AWS calls frontier agents.

"They are autonomous and can operate on their own for hours or days," she said. "They can deal with a great degree of complexity, break that complexity into smaller tasks, scale by doing multiple things in parallel, and learn from interactions with humans."

AWS frames frontier agents as virtual specialists.

"Imagine if you had your own security consultant, your own virtual software developer, or your own on-call engineer," said Stanic. "That's how we describe those frontier agents - the security agent, the AWS DevOps agent, the software developer agent."

Humans and machines

Stanic said the value of agentic AI lies less in replacing people and more in reshaping how work is divided.

"You want to introduce efficiency by putting agents to work on mundane, repetitive tasks, and refocus human efforts on what really distinguishes us - creativity and curiosity," she said.

Document-heavy workflows, such as home loan applications, are early beneficiaries.

"An agent can process documentation, extract key information and present that back to an automated workflow or to humans for further action," Stanic said.

In software development, agents are already changing day-to-day practice.

"Agents can write the code based on instructions, and then human developers can refocus their energy on solving complex problems," she said.

Operational use cases, including incident response, are also gaining traction.

"An agent is continuously monitoring, assessing situations across many systems," said Stanic. "It presents what it believes is the root cause and suggests remediation, and then on-call engineers decide the best course of action."

She said this has measurable impact. "What used to take seven, eight or ten hours can now be assessed in minutes. That directly improves customer experience because systems don't stay down as long."

Industry adoption

According to Stanic, regulated industries are among the fastest adopters.

"Globally, we've seen regulated industries take a lead in AI adoption," she said. "Compliance and risk, supply chain optimisation and complex workflows like mortgage applications are strong examples."

Customer experience remains a major focus, particularly in contact centres.

"You still get a human answering your call, but they're far more effective because their work is augmented by AI," said Stanic.

In software engineering, adoption is already quantifiable.

"We've seen organisations where around 40% of production-grade code is generated by AI, and developers are about 18% more productive," she said.

Silicon strategy

At re:Invent, AWS reinforced that large-scale agentic AI depends on infrastructure economics as much as software capability.

"We can't forget the importance of silicon," said Stanic. "It's because of high-performing, cost-effective chips that organisations can build viable AI applications."

She said Trainium 3 Ultra servers deliver major gains.

"They're about four times more energy efficient, with roughly four times higher compute and memory bandwidth," she said. "That translates into better price-performance for AI training and inference."

AWS is already looking ahead. "We've released Trainium 3 and we're working on Trainium 4. This pace of innovation won't stop."

Graviton 5, she added, targets general-purpose cloud workloads.

"Customers have seen up to 25% better performance," said Stanic. "The goal is always the same: faster performance, lower cost and greater energy efficiency."

Cloud, control and choice

While AWS promotes cloud-first architectures, Stanic acknowledged that regulation and data sovereignty remain constraints for some customers.

AI factories are AWS's response.

"We bundle custom chips, GPUs, networking, storage and AI services, and make that available for deployment in customers' own data centres," she said. "It brings the cloud experience on-premises while meeting compliance requirements."

With more than 100 models available through Amazon Bedrock, Stanic said choice is intentional.

"You start with a sophisticated model to prove the idea," she said. "Once it works, you optimise for latency and cost - often with a smaller, more efficient model."

Modernising legacy systems

Agentic AI is also being applied to long-standing legacy environments.

"There's a lot of legacy out there - mainframes, VMware workloads, old Java versions," said Stanic. "Agentic architectural blueprints make modernisation much faster."

AWS has already used the approach internally.

"We modernised more than 10,000 applications, upgraded Java versions and saved about 4,500 developer years," she said.

"Those same capabilities are now being applied through AWS Transform to help customers reduce technical debt."