Enterprise AI goes agentic

A year ago, a company could bolt a chatbot onto its website and call that “enterprise artificial intelligence.” In April 2026, the money is moving to software that can open a purchase order, check a policy, route an approval, and leave a record of every step inside the system where the work already happens. (decrypt.co, oracle.com) OpenAI said on April 8 that enterprise customers now make up more than 40% of its revenue, and it said that share is on track to reach parity with consumer revenue by the end of 2026. The same post said its application programming interfaces process more than 15 billion tokens per minute, which is the plumbing behind companies building these task-running systems. (openai.com) The shift is not from “chat” to “smarter chat.” It is from answering questions in a box to taking actions in finance, human resources, customer support, and supply chains without forcing employees to copy and paste between five tabs. (oracle.com, salesforce.com) That is why the new selling point is not raw model intelligence. The selling point is control: permissions, approval chains, audit logs, observability, and data boundaries that let a chief information officer explain exactly what the system did on Tuesday at 3:14 p.m. (learn.microsoft.com, oracle.com) Oracle made that case directly on March 24 when it launched Fusion Agentic Applications inside Oracle Fusion Cloud Applications. Oracle said these agents can securely use enterprise data, workflows, policies, approval hierarchies, permissions, and transactional context, and it drew a bright line between that and a copilot sitting outside the transaction system. (oracle.com) Microsoft is pushing the same idea from the platform side instead of the applications side. Its Hosted Agents in Foundry Agent Service are pitched as managed hosting for containerized agents with scaling and observability, so developers do not have to hand-build the web servers, memory, security hooks, and rollbacks around every agent. (learn.microsoft.com, azure.microsoft.com) Salesforce is pushing from the customer-data side. Agentforce is marketed as a platform for deploying “digital labor” on top of the workflows, data, and integrations already inside Salesforce, which tells you where the market is heading: not one giant brain, but managed workers tied to specific records and specific permissions. (salesforce.com, businesswire.com) The reason incumbents have an advantage is boring and powerful. The company that already owns the payroll rules, vendor master data, approval ladder, and customer history has the easiest path to turning an artificial intelligence model into a worker that can actually do the job. (oracle.com, salesforce.com) That also explains why “agentic” has become an enterprise word before it becomes a consumer one. A consumer assistant can be impressive if it drafts a good email, but a corporate assistant becomes valuable when it can follow a policy, stay inside a permission boundary, and hand compliance a clean trail after the fact. (decrypt.co, learn.microsoft.com) So the race is no longer just about whose model writes the best paragraph. It is about who can wrap models in the least glamorous parts of enterprise software — identity, monitoring, approvals, connectors, and logs — because that is what turns a clever demo into something a bank, insurer, hospital, or manufacturer will let touch real work. (openai.com, learn.microsoft.com, oracle.com)