On a stage in New York this week, an Amazon Web Services executive asked an AI agent to resolve a customer issue, and the agent did the work: pulling records from a corporate database, checking the company’s policies, drafting a response and handing it to a human for approval. The demonstration, part of AWS Summit New York, was designed to show what the company means by enterprise-grade AI agents.
The tools announced at the summit address the objections businesses raised during the first wave of AI chatbots. The updates include deeper integration with enterprise data security, so that agents operate within the same permissions and access controls as employees; orchestration that lets multiple agents divide a task among themselves; and more precise tool calling, so that an agent summons the right internal system rather than guessing.
Amazon’s official blog laid out the details, and the subtext was clear. The company is positioning its agent platform against Microsoft’s Copilot ecosystem, which has an early lead inside many corporations because of its reach through Windows and Office. AWS is betting that security, control and integration with a company’s own cloud data will matter more than the convenience of having a copilot already installed.
The stakes are the next phase of cloud computing. Analysts have argued that agents, not chatbots, will drive the next round of spending on AI, because they automate work rather than answer questions. The cloud providers that win the agent platform battle would be positioned to capture the associated compute, storage and data fees, an opportunity measured in tens of billions of dollars a year.
The summit’s theme was that agents should do real work rather than act as toys. AWS executives pointed to customer-service resolution, procurement and claims processing as the tasks they expect agents to handle first, and they argued that the value of AI in the enterprise will be measured by completed work, not by chat sessions. That framing is also a rebuke to the first wave of consumer chatbots, which generated enthusiasm but few invoices.
Enterprise adoption so far has been cautious. Many companies piloted chatbots in 2024 and 2025, then discovered that the tools made mistakes with company data and could not act on the systems they were supposed to manage. The new generation of agents is designed to close that gap, but it faces the same trust problem, and vendors are discovering that permissioning and audit trails sell better than demos.
AWS’s approach leans on its existing advantages. Companies that run their operations on AWS can give agents access to their data without moving it, and the security controls that govern employees can be extended to the agents themselves. That is a different pitch from Microsoft’s, which rests on the assistant being embedded in the tools people already use.
The multi-agent orchestration piece is the most technically ambitious. Coordinating several agents, each with its own task and tools, introduces failure modes that single agents do not have, and the industry has not settled on standards for how agents should hand work to one another. AWS’s announcement puts a stake in the ground, but customers will judge it by whether the orchestration works in production.
Security is the make-or-break issue. Corporate data has become the most sensitive asset in the AI economy, and businesses have shown they will not connect agents to their systems without granular controls. AWS’s updates put access controls at the center of the platform, with agents inheriting the permissions of the employees whose work they share, and every action logged for review. Executives at the summit said those features were the ones that persuaded early customers to move from pilots to production.
Pricing will shape adoption. AWS has said agents will be billed per task, a model that aligns cost with value but also means customers cannot predict their bills. Enterprise buyers have pressed vendors for fixed pricing, and analysts expect the major cloud providers to converge on subscription models over time, as they did with the first generation of AI features.
The competitive field is crowded. Microsoft has its own orchestration tools, Google has been adding agent features across its cloud, and a new generation of startups sells agent platforms for specific industries. AWS’s scale and its grip on enterprise infrastructure give it a strong hand, but the market is young, and the early lead belongs to whoever ships reliable, boring software.
For businesses, the message from New York was that agents are moving from experiments to infrastructure. The tools announced this week are not research demos; they are products aimed at the people who run payroll, logistics and customer service. Whether they work as advertised will be tested in the unglamorous settings where the first wave of AI already stumbled.
