In the enterprise, a confidently wrong AI agent answer can come from something more basic than the model: the agent doesn't know what the business means by its own terms, like what counts as revenue or which customer record is current. 在企业环境中，AI 智能体给出“自信但错误”的回答，其原因可能比模型本身更基础：智能体并不理解业务术语的含义，例如什么是“收入”，或者哪个客户记录才是当前的最新版本。
The VB Pulse tracker has followed how enterprises are responding to that across three separate waves since June. Across all three waves, most respondents report running, piloting or building a governed semantic layer: 58% in June, 63% in July and 67% in August. But far fewer name that layer as their agents’ primary source of business context: 21%, 19% and 13%, respectively.VB Pulse 追踪器自 6 月以来，分三波调研了企业对此的应对情况。在所有三波调研中，大多数受访者表示正在运行、试点或构建受治理的语义层：6 月为 58%，7 月为 63%，8 月为 67%。但将该层作为智能体获取业务上下文的主要来源的企业比例却低得多，分别为 21%、19% 和 13%。
Context failures remain widespread. Sixty-four percent of August respondents traced a confidently wrong agent answer to missing or inconsistent business context in the past six months. July was 68%, and June was 57%. 上下文缺失的问题依然普遍存在。8 月份有 64% 的受访者表示，过去六个月中，智能体给出“自信但错误”的回答是由于业务上下文缺失或不一致所致。7 月份这一比例为 68%，6 月份为 57%。
Agents get context somewhere else智能体从其他地方获取上下文
Respondents named several different primary sources of business context for their agents. Retrieval over documents was named by 32%. Direct queries to live systems, where an agent runs SQL or calls an API or an MCP server and reads whatever comes back, were named by 21%, up from 11% in July, a change large enough to call. The gap between retrieval and direct queries is too close to call. Long-context loading, pasting a large set of documents or records straight into the model's context window, was named by 19%. The governed semantic layer was named by 13%.受访者列举了智能体获取业务上下文的几种不同主要来源。32% 的受访者提到了文档检索。21% 的受访者提到了对实时系统的直接查询（即智能体运行 SQL、调用 API 或 MCP 服务器并读取返回结果），这一比例高于 7 月份的 11%，增幅显著。检索与直接查询之间的差距目前尚难定论。19% 的受访者提到了长上下文加载，即将大量文档或记录直接粘贴到模型的上下文窗口中。仅有 13% 的受访者提到了受治理的语义层。
Neither direct queries nor long-context loading carries a governed business definition by default. A direct query can return whatever sits in the table without telling the agent what the column means. Long-context loading hands the model raw material without necessarily giving it an agreed definition of what the business means by its own terms. Either approach can be paired with a governed layer; the survey does not tell us whether respondents are doing that.直接查询和长上下文加载在默认情况下都不具备受治理的业务定义。直接查询可能返回表中的任何内容，而不会告知智能体这些列的含义。长上下文加载则是将原始数据交给模型，而不一定会提供关于业务术语的统一标准定义。这两种方法都可以与受治理层配合使用，但调查并未显示受访者是否采取了这种做法。
Enterprises adopting a governed layer report more context failures采用受治理层的企业报告了更多的上下文错误
Enterprises running, piloting or building a semantic layer report a confidently wrong agent answer far more often than enterprises still evaluating one or with no plans at all, 78% against 37%. A July wave of a separate 101 respondents found the same pattern, 89% against 35%. The gap holds even after setting aside enterprises that don't run agents on their own data or don't track root cause: 79% against 41% in August, 90% against 46% in July.与仍在评估或完全没有计划的企业相比，正在运行、试点或构建语义层的企业报告智能体给出“自信但错误”回答的频率要高得多，比例为 78% 对 37%。7 月份对另外 101 名受访者的调查也发现了同样的模式，比例为 89% 对 35%。即使排除了那些不使用自身数据运行智能体或不追踪根本原因的企业，这种差距依然存在：8 月份为 79% 对 41%，7 月份为 90% 对 46%。
VB Pulse did not ask respondents why the two groups differ. The August report offers two possible explanations. One is that a semantic layer gives a team a correct definition to check a wrong answer against, so these enterprises catch more of their own failures rather than producing more of them. The other is reverse causation: enterprises that already had context failures may be the ones that went on to build a semantic layer, rather than the layer causing the gap. The survey cannot tell which explanation is driving the gap — or whether the layer itself changes the underlying failure rate.VB Pulse 并未询问受访者这两组企业为何存在差异。8 月份的报告提供了两种可能的解释。一种是语义层为团队提供了一个正确的定义，以便核对错误的回答，因此这些企业能发现更多自身的问题，而不是产生了更多的错误。另一种是反向因果关系：可能正是因为原本就存在上下文错误，企业才去构建语义层，而不是语义层导致了这种差距。调查无法判断哪种解释是导致差距的原因，也无法判断语义层本身是否改变了潜在的错误率。
How vendors are approaching the gap differently供应商如何以不同方式应对这一差距
Some vendors moving toward direct queries are putting context around the query path in different ways. OpenAI's Data agent, launched inside ChatGPT Work in September, connects directly to warehouses including Snowflake, Databricks and BigQuery and pulls context from Slack, BI dashboards and file storage in the same pass. Asked whether the agent builds a persistent context layer across a customer's tools, Arpan Shah, general manager of the enterprise technology vertical at OpenAI, told VentureBeat it does not.一些转向直接查询的供应商正在以不同方式围绕查询路径构建上下文。 OpenAI 的 Data agent 于 9 月在 ChatGPT Work 中推出，它能直接连接 Snowflake、Databricks 和 BigQuery 等数据仓库，并在同一过程中从 Slack、BI 仪表板和文件存储中提取上下文。当被问及该智能体是否会在客户的工具之间构建持久的上下文层时，OpenAI 企业技术垂直领域总经理 Arpan Shah 向 VentureBeat 表示，它并没有这样做。
"It combines all the various contexts in their state, as opposed to generating a net new context layer itself," Shah said.“它结合了各种状态下的上下文，而不是生成一个全新的上下文层，”Shah 说道。
Keewano blurs that distinction further. Its KeewanoDB, also launched in September, lets agents read raw event sequences directly through MCP, the kind of direct query the survey is measuring. But Mark Kardashov, co-founder and CEO of Keewano, described a semantic layer built during ingestion, using large language models, that attaches additional context to each event as it lands. A direct query against KeewanoDB is not necessarily a query with no semantic context behind it.Keewano 进一步模糊了这种界限。其同样于 9 月推出的 KeewanoDB 允许智能体通过 MCP 直接读取原始事件序列，这正是调查中所衡量的直接查询方式。但 Keewano 的联合创始人兼首席执行官 Mark Kardashov 描述了一种在摄取过程中利用大语言模型构建的语义层，它在每个事件落地时为其附加额外的上下文。因此，对 KeewanoDB 的直接查询并不一定意味着查询背后没有任何语义上下文。
Snowflake is approaching the same gap from the routing side. The company's Cortex AI Gateway, which auto-routes a task to a smaller model first and escalates only when needed, depends on the task arriving with context already prepared. Baris Gultekin, Snowflake's vice president of AI, explained in August that without good context, a model has to do the exploratory work itself, writing and testing SQL, searching through data and retrying when something doesn't work, a process that's expensive and typically needs a more capable model. Packaging the context in advance removes that step. Snowflake says its Horizon Context and Cortex Sense tools provide context capabilities for that preparation before the model has to explore the data itself, while Snowflake's existing access controls govern the data and models the agent can use. Snowflake 正从路由层面应对这一差距。该公司的 Cortex AI Gateway 会自动将任务先路由至较小的模型，仅在必要时才升级，这取决于任务在到达时是否已经准备好了上下文。Snowflake AI 副总裁 Baris Gultekin 在 8 月解释称，如果没有良好的上下文，模型就必须自行进行探索性工作，包括编写和测试 SQL、搜索数据并在失败时重试，这一过程成本高昂，且通常需要更强大的模型。提前封装上下文可以省去这一步骤。Snowflake 表示，其 Horizon Context 和 Cortex Sense 工具可以在模型探索数据之前提供上下文准备能力，同时 Snowflake 现有的访问控制机制可以管理智能体可使用的数据和模型。
"The interesting part is what it says about where differentiation has moved," Sanjeev Mohan, principal at SanjMo, told VentureBeat in August. "Snowflake isn't really selling routing, it's selling routing that never leaves the governed data boundary, with access controls, tagging and cost attribution already attached."“有趣的是，这说明了差异化竞争已经转移到了哪里，”SanjMo 负责人 Sanjeev Mohan 在 8 月告诉 VentureBeat。“Snowflake 卖的其实不是路由本身，而是卖一种永远不会离开受治理数据边界的路由，它附带了访问控制、标记和成本归因。”
What this means for enterprises这对企业意味着什么
Building a semantic layer does not decide what an agent reads. Two-thirds of August respondents said their organizations run, pilot or build one. Even among that group, only 20% name the semantic layer as their agents’ primary source of business context.构建语义层并不能决定智能体读取什么内容。三分之二的 8 月份受访者表示其组织正在运行、试点或构建语义层。即便在这一群体中，也只有 20% 的人将语义层作为智能体获取业务上下文的主要来源。
Whether a direct query carries business context behind it depends on what sits behind it. Keewano says it attaches context at ingestion, Snowflake pairs context preparation with its existing access controls, and OpenAI's Data agent assembles context on the fly without building a persistent layer.直接查询背后是否携带业务上下文，取决于其底层架构。Keewano 表示在摄取时附加上下文；Snowflake 将上下文准备与其现有的访问控制相结合；而 OpenAI 的 Data agent 则是在动态组装上下文，而不构建持久层。