Work is stalling inside private AI chats. In a Salesforce Pulse survey of 250 U.S. executives published on October 27, 59 percent said work stalls or is delayed every week because it sits in one person's AI tool, where the rest of the team cannot see it. The finding reframes the adoption question. The problem is not whether employees use AI. It is whether shared AI context exists at all.
The duplication is weekly
The stall is the visible cost. The larger cost is duplication. In the same Salesforce survey, 68 percent of executives said insights or outputs from one person's AI use do not get shared or reused by the team each week, and 66 percent said two colleagues use AI to solve the same problem separately without realizing it, at least once a week. One respondent put it plainly: "One person learns a prompt that can cut review in half and no one else knows and it stays in their AI workflow." An IT director gave the engineering version: engineers wasted time prompting AI to fix the exact same error because separate chat histories prevented them from sharing the solution.
Scale makes it worse. Among enterprise executives, at companies with 2,500 employees or more, 73 percent reported weekly stalls, compared with 37 percent at small and medium businesses with fewer than 250 employees. Salesforce notes the arithmetic: more employees means more private chat histories and more places for work to go unseen.
Integration has not closed the gap. Forty-nine percent of the organizations are deploying AI agents, but just 15 percent have embedded agents across multiple systems so they work seamlessly, and only 11 percent say their AI tools sit in one integrated ecosystem with no switching cost. Executives estimate employees lose an average of 3.6 hours a week switching between AI tools and other systems. Salesforce annualized that estimate, using the low end of each response range across 50 workweeks, at about 181 hours, or roughly 22 business days per employee. That is an executive estimate, not measured time.
Shared AI context is what executives say they want
Given a blank slate, 72 percent of executives said they would design AI to work across the team with shared context so value compounds, while 26 percent would keep AI use individual. Slack General Manager Rob Seaman stated the vendor case: "When an agent works in a shared space instead of a private chat, it picks up how your company actually operates, the projects, the debates, the decisions, all living in the conversations your team is already having."
The supply side moved the same week, on a later timetable. At SAP TechEd in Berlin on October 27, SAP said its Enterprise AI Gateway, consolidating authentication, policy, monitoring, usage tracking and auditing, and new Knowledge Graph capabilities drawing on customer specific data from SAP and non-SAP sources, are planned for the first quarter of 2027. SAP Chief Technology Officer Philipp Herzig said the company is "moving from AI learning your business to running it." The context and governance layer, in other words, is announced before it is available.
Sharing alone does not make a record true
A shared output is not yet a reliable record. Redwood Software reported on October 22, from a survey of 300 senior technology, risk and AI leaders collected by Leger Opinion, that 77 percent have experienced a negative consequence from an ungoverned AI agent, with unexpected cloud costs the most common, followed by operational disruption and compliance issues. Only 42 percent have a centralized registry of deployed agents and only 51 percent have real time monitoring of agent activity, even as 89 percent increased deployment in the past year.
Verification is also thin at the point of use. O'Reilly reported on October 27, from Censuswide research among 2,000 UK employees, that 39 percent of employees using AI generated answers have acted on a response that later proved inaccurate, incomplete or unsuitable, and only 24 percent always verify an answer before using it. Asked what would most increase confidence for important decisions, employees put human review first, at 39 percent.
Leaders and employees also disagree on whether the current stack is connected at all. Ricoh reported on October 21, from a survey of 2,017 employees and decision makers in the U.S. and Canada, that 91 percent of decision makers are confident AI investments are delivering expected value, while 53 percent of employees say AI creates friction or extra work. Ninety percent of decision makers believe their tools, systems and processes are connected, while 76 percent of employees say their jobs would be easier if those tools worked better together, and 45 percent use non-approved tools at least monthly to get work done.
The surveys converge on one distinction. Sharing a chat shares an output. Sharing a record shares something a colleague or an agent can check, cite and correct. Shared AI context compounds whatever it holds: if it holds unreviewed outputs in private histories, it compounds duplication and error, as the Salesforce duplication figures and the Redwood governance figures show from opposite ends. If it holds a reviewed record, with a source, a date, a superseded version and a person who approved it, the next answer can build on the last decision instead of rediscovering it. That is the standard Jilu is built around, and it is the test any shared context product should face, including this one: not whether context is shared, but whether a correction to it propagates to everyone who already used it.