AI at work: agents replace chat windows
The enterprise AI story is shifting from flashy chatbots to embedded agents and orchestration — vendors are wiring AI into existing workflows so tools act inside apps employees already use. Atlassian added visual AI and third-party agents to Confluence, while reports say OpenAI and others are moving toward cybersecurity and managed-agent products that prioritise governance and operational plumbing. That trend matters because the winners will be the companies that solve integration, observability and control, not just the firms with the fanciest model demos. (techcrunch.com) (axios.com) (a16z.com)
The office chatbot is getting demoted from a destination to a utility. On April 8, Atlassian added Remix to Confluence so a page can turn into a chart, infographic, diagram, or app without leaving the document where the work already lives. (techcrunch.com) (atlassian.com) Atlassian also added three partner agents inside Confluence: one for Lovable to make prototypes, one for Replit to turn technical docs into starter apps, and one for Gamma to build presentations. The point is that the employee stays in Confluence while the software does the handoff that used to mean copy, paste, and reformat. (techcrunch.com) (atlassian.com) That is a different product idea from the 2023 and 2024 wave of chat windows. A chat window asks a worker to stop what they are doing, open a separate box, and translate a real task into a prompt, while an embedded agent starts from the file, ticket, or page the worker already has open. (techcrunch.com) (openai.com) Big companies are buying that second model faster than the first. Andreessen Horowitz said on January 30 that 78% of surveyed Global 2000 information chiefs were using OpenAI models in production, but the survey was about deployment inside large-company systems, not about employees casually chatting with bots on the side. (a16z.com) The same survey covered 100 Global 2000 companies with at least $500 million in annual revenue, and more than half had over 10,000 employees. In companies that size, the hard part is rarely generating text; the hard part is wiring a model into approval chains, internal data, and software people already use every day. (a16z.com) That is why the plumbing is suddenly the product. OpenAI said on February 23 that its Frontier Alliances with Accenture, Boston Consulting Group, Capgemini, and McKinsey are meant to help customers move from pilots to production with system integration, workflow redesign, and secure agent deployments. (openai.com) (cnbc.com) The technical shift underneath this is tool calling, which means the model can use other software instead of only writing an answer. OpenAI’s Responses application programming interface added support for remote Model Context Protocol servers in May 2025, which lets agents reach outside the model and act on external tools and data sources. (openai.com) (openai.github.io) Once an agent can take actions, companies stop arguing only about model quality and start arguing about control. Atlassian says admins can enable partner agents in administration settings, and OpenAI’s cyber program says access to stronger capabilities is being gated through identity and trust checks rather than thrown open to everyone. (atlassian.com) (openai.com) You can see that caution in cybersecurity first because the risks are easy to picture. Axios reported on April 7 that Anthropic is limiting Mythos Preview to a handpicked group of tech and security companies because the model is strong at finding and exploiting flaws, and OpenAI has separately said it expects upcoming frontier systems to pose high cyber risk. (axios.com 1) (axios.com 2) So the new enterprise contest is less like “whose chatbot sounds smartest” and more like “whose agent can be trusted with the keys.” The vendors that win will be the ones that can show where the agent got its data, what tool it touched, which admin approved it, and how to shut it off when it goes wrong. (openai.com 1) (openai.com 2)
