|
Welcome back! When AI software sellers pitch their vision of the future, it can start to feel paradoxical. On one hand, sellers say AI can meaningfully automate large swaths of white collar work, which should in theory help companies save money because they won’t need to pay as many humans. On the other hand, if companies cut jobs while embracing automation, there will be fewer white collar workers buying software subscriptions. Earlier this week, Microsoft Executive Vice President Rajesh Jha suggested a fascinating way out of that paradox: Microsoft thinks it can keep selling seat-based subscriptions for products like Office 365—to the AI agents working inside Microsoft’s customers. “I'm not seeing AI as driving down seats” in Microsoft’s Office 365 productivity suite, Jha said at a UBS conference on Tuesday. “If anything, I think AI is going to be an opportunity for us to drive seat growth. If you think about an organization in the future, it’s going to have more [AI] agents than people.” Jha, who oversees Microsoft’s productivity apps including Word, Powerpoint, Excel, and Teams, as well as the 365 Copilot chatbot Microsoft sells to businesses, said AI agents will likely need their own software subscriptions, and possibly their own computers, to work the same way humans do today. “Many of these agents are going to be embodied. They’re going to have an identity. They’re going to be in the address book. They’re going to have a mailbox. They’re going to need a computer to do its computation in a secure way,” Jha said. “It’s going to want to chat with you. It’s going to want to join a meeting. To me, all of those embodied agents are seat opportunities.” Jha’s prediction may sound counterintuitive to software buyers that are looking for agents as a way to cut down on costs associated with humans, or to replace enterprise software. (Granted, if agents can really replace people, the savings on salaries would outweigh keeping up subscriptions to 365 or other applications for the agents.) Still, there’s a certain logic to what Jha is describing—specifically, that companies haven’t yet landed on a solution for how to grant AI agents the right permissions to use different applications or how to verify that each agent is doing what it’s supposed to do. That’s why companies like Microsoft and ServiceNow are now trying to pitch their software as a way to “orchestrate” an army of agents sold by various other tech firms. It’s also worth pointing out that this discussion is mostly hypothetical. A majority of the current AI agents sold by Microsoft and other providers can’t automate entire jobs or work on computers in the same ways that humans do; most agents work within a specific application, such as by generating Powerpoint spreadsheets in response to written prompts or generating leads for salespeople. During the panel, Jha acknowledged that such agents are “immature” and probably won’t be ready until next year at the earliest. But it’s nevertheless fascinating to hear Jha’s vision for AI agents as software customers. Microsoft, OpenAI, and other AI companies are still figuring out how to sell AI software to enterprises, as evidenced by both of those firms’ recent tempering of expectations around AI agent sales. Selling software to AI agents might prove even trickier. AWS’ Fallback Plan if its AI Models and Chips Don’t Catch On Amazon Web Services has had a tough time keeping up with the breakneck pace Google and OpenAI have set in developing cutting-edge AI for cloud customers. But in the long run, AWS may not need to develop its own state-of-the-art AI to maintain its lead in the cloud-server market. It remains the industry’s top seller of cloud servers, storage and databases for developing and using applications. And based on conversations I had at AWS’ annual re:Invent conference for customers in Las Vegas, Amazon could capitalize on that lead as businesses embrace advanced AI apps known as agents. Three years into the ChatGPT era, no company has emerged as a leader in enabling AI agents that handle corporate tasks involving multiple steps, such as producing dashboards and resolving IT helpdesk tickets. “What we're seeing now is that agents are basically a wide-open market,” said Eno Reyes, chief technology officer at Factory, a startup that sells an AI coding service to large companies in regulated industries. (Factory is a customer of AWS and other major cloud providers.) AWS wants customers to develop different agents for different teams, he said. So far, agents have been buggy and unreliable, no matter who makes them, ensuring they’re a tough sell across the industry. But plenty of them are gaining ground as AWS and other providers get hands-on with customers to make the technology work. After a slow start in this field, AWS has quickly boosted a service for developing agents, now called Bedrock AgentCore. This week it released features for setting safety and security controls around the technology to minimize its ability to expose sensitive data or behave in unpredictable ways. The features show how AWS is drawing on its experience serving many of the world’s largest companies to address the worries some firms might have about rogue AI. “Agents will become the primary way companies get value from AI,” Amazon CEO Andy Jassy said in a post on X earlier this week. One factor working in AWS’ favor is it already went through the hard work of convincing many large companies to move their computing to the cloud. AWS is betting the trust it built with cloud customers will eventually carry over in the AI era. “We're tantalizingly close [to companies getting value from agents], but one of the challenges to getting these things into production is trust,” said Byron Cook, an AWS vice president and distinguished scientist who works on AI reasoning and reducing AI hallucinations. AWS is also trying to move quickly with selling ready-made agents that don’t require much legwork. This year, it launched an agent that rewrites older code so applications can run better in the cloud, as well as an AI “vibe coding” agent called Kiro. But AWS was criticized this summer for not immediately disclosing a vulnerability in another AI coding assistant, Q Developer, that could have enabled attackers to delete data from customers’ computers. (The problem was fixed.) Q Developer is part of a suite of AI tools AWS launched this summer that help workers automate work tasks like researching documents and creating visual charts and graphs. At re:Invent, AWS launched a new agent that scans and tests applications for security vulnerabilities. To be sure, AWS has its hands full as AI competition ramps up: It appeared to make a major concession this week by agreeing to make it easier for its customers to simultaneously use Google and Microsoft’s clouds. AWS only announced one customer of Trainium3, the latest version of the AI chip it designed in-house. And the jury is out on whether any developers are interested in its new AI model, Omni. Even if those new products don’t gain traction, AWS’ core strengths and the continued rise of general-purpose cloud computing should keep its performance humming for the foreseeable future.
| 
