|
Amazon Web Services has been in the spotlight this week because it can now sell OpenAI models to its cloud-server customers for the first time, thanks to a sudden change in OpenAI’s long-running partnership with Microsoft. But AWS is still chasing its white whale: creating a hit enterprise application. With that in mind, AWS on Tuesday announced a new desktop app for its suite of AI agents that automate tasks such as creating presentations, documents or images. The new app, known as Amazon Quick, runs on your computer, but in the background, so you can do other things simultaneously. Notably, it works without an AWS account, starting at $20 per month. The Amazon agents are supposed to complete tasks involving customers’ computer files and other apps they use, such as Google Workspace, Microsoft 365, Zoom and Salesforce. In other words, Amazon has joined the very crowded race to develop superagents, or digital co-workers that can use a variety of enterprise apps just as humans do. Every major tech company is betting in part on a future in which white-collar workers no longer manually use different enterprise apps at work. Instead, they’ll oversee a suite of AI agents that connect to applications on their own, either through application programming interfaces or by taking over desktops and browsers to perform computer work. Quick is basically Amazon’s version of Claude Cowork, an AI-powered tool that takes over a computer to handle workplace tasks involving numerous applications, from Microsoft Office apps to Slack. The app connects to those apps through native connections, standard application program interfaces, and model context protocol servers, an Amazon spokesperson explained. For those wondering what’s behind the hood, Amazon said the new Quick app is powered by “a bunch of models” but wouldn’t disclose specifics. Quick previously ran on a mix of Amazon’s own Nova AI models and Anthropic Claude models, my colleague Catherine reported in January. Based on Amazon’s presentation on Tuesday, it’s not obvious why customers might flock to Quick, though AWS says BMW, 3M, Mondelēz, Southwest Airlines, and the NFL are using it. But Amazon made similar claims about an earlier app, Amazon Q, which handled AI coding and other tasks, and Q didn’t seem to become a big hit. AWS may be the cloud services king, but it has struggled for years to develop the kind of enterprise apps that have long boosted rivals Microsoft and Google, which pair their apps with cloud server rentals, to great effect. IBM Is Still Fighting For Coding Relevance In a similar vein, IBM is still trying to get into the AI coding race, where it has long been behind. While its coding tool, watsonx code assistant, has gained some traction helping companies modernize their codebases, IBM’s coding tools haven’t generated the same developer enthusiasm as those from Microsoft-owned GitHub, Cursor, and frontier AI labs. Now IBM is marketing Bob, a coding tool it developed in 2025 for its own engineers and made generally available to IBM customers on Tuesday. IBM Chief Commercial Officer Rob Thomas said he thinks Bob is “ one of the most interesting new product innovations from IBM in a long time.” Bob, like watsonx, is primarily used for a narrower goal than coding agents like Claude Code. Bob aims to help customers such as consulting firms Ernst & Young and Blue Pearl translate their databases into new coding languages so they can easily update their websites, apps, and features. To be sure, customers can also use it to develop apps on top of the data IBM stores for them, such as an app for financial modeling. Thomas said Bob can pick and choose between the most advanced (and costly) models like Claude and cheaper open-source ones to handle customers’ tasks more efficiently compared to better-known coding assistants. That reminds us somewhat of Cognition’s coding agent Devin, which also figures out which AI models to use for a given task. Bob can also use IBM’s homegrown Granite small language models. “ The biggest gate to progress becomes token consumption and token cost ... so Bob is constantly optimizing cost and performance and choosing the model that delivers the best cost performance,” Thomas said. Salesforce CEO Says His Company Will Also Charge For AI Based on ‘Outcomes’ We’d be remiss if we didn’t call out a comment from Salesforce CEO Marc Benioff in our Sunday story about AI pricing models. He said his company will join the trend of charging customers when AI tools actually complete a task, known as outcome-based pricing, rather than just based on how much the tools are used. Benioff didn’t give specifics or say when that pricing model might be used, but his comment implied it could be based on Salesforce’s Agentic Work Unit, a new metric it introduced to track how many tasks an AI agent completes. The comment comes as software firms rush to introduce new pricing models as AI threatens their traditional seat-based business model. Microsoft-owned Github became the latest enterprise app to shake up its pricing strategy, saying Monday that it will shift to usage based pricing. Salesforce currently uses a per-seat subscription model that includes a certain number of AI credits based on usage. AI agent tasks that require more tokens, or bits of data processed by an AI model, cost more credits. For example, Salesforce said on its pricing site that an AI agent that helps a company schedule field technicians might use up 120 credits for every appointment it schedules and cost $360 per month, or about 60 cents per appointment. The number of credits it takes to complete a task depends on how many actions the AI agent takes, such as identifying a customer by email, finding appointment time slots, and answering the customer in a chat to tell them their appointment was scheduled. In an outcome-based model, customers would likely be charged only when an AI agent completes a task, and credits would be measured by how many tasks are completed rather than how many actions an agent takes. Salesforce rivals Sierra, Zendesk, and Intercom already use task-based pricing models for their AI tools.
| 
