For most of the past three years, enterprise AI meant a chatbot: something you asked a question and read an answer from. That definition is quietly being retired. This month Adobe turned Acrobat, the reader that opens more than 400 billion PDFs a year, into what it now calls a "document productivity platform," with agents that can search across thousands of contracts, pull out renewal dates and reporting terms, and assemble the findings into a finished presentation. Microsoft, Box and others are racing down the same road. The shift sounds mundane, but it marks a real change in kind: AI is moving from answering questions to taking actions.
Nowhere is that leap larger than on the factory floor. Enterprise resource planning systems, the transactional backbone of manufacturing, were built decades ago to record what happened, not to do anything about it. Now vendors like Microsoft, with an "agentic" version of Dynamics 365, want AI agents that can spot a looming stockout, choose a supplier and draft the purchase order. The prize is obvious: work that once meant a person shuttling between five systems collapses into a single automated step. So is the hazard. "AI raises the bar on the data foundation rather than replacing the ERP," one Salesforce engineer warned, noting that firms with messy data will simply "get confident wrong answers faster."
That phrase captures the week's real theme. The appetite for agents is running well ahead of the ability to govern them. OneTrust's 2026 governance survey found that 87 percent of organizations now encourage the use of AI agents, yet nearly half admit to at least one incident in the past year in which an AI system took an action nobody had approved. Only 5 percent say accountability is clearly defined across the whole AI lifecycle. A third of respondents said employees had reached for unapproved tools simply because the sanctioned ones arrived too slowly.
The reliability gap is not small. A SAS and IDC study spanning 28 countries found that even state-of-the-art agents can exceed a 25 percent error rate on complex tasks, which the authors flatly called "unacceptable in high-stakes decision-making." Trust falls to match: worker confidence dropped from 76 percent for ordinary generative tools to 66 percent for autonomous agents, and in one striking figure, 97 percent of users said they override an automated recommendation in some situations, most often because the system cannot explain how it got there. Tellingly, the same study found that companies with formal governance were far likelier to see strong returns, a reminder that oversight and payoff are not opposites but partners.
The manufacturers moving carefully offer the template. Even the executives building agentic ERP systems draw a firm line: let the agent detect the problem and draft the purchase order, but keep a human hand on the send button for anything consequential. "The key is defining those guardrails upfront," said one, "agreeing on how you do and don't want the system to act." The question, as another engineer put it, is no longer whether AI can perform a task. It is who owns the outcome when the agent gets it wrong, and this week most companies still could not answer.