It is often easier to count how many AI licenses a company has purchased than to figure out what exactly its employees are doing with AI. This time, Volkswagen Group provided two figures that are worth looking at separately. On October 21st, Microsoft announced that Volkswagen Group had equipped over 330,000 employees with Microsoft Copilot; during the phased rollout of Copilot Chat, the number of monthly active users once exceeded 300,000. The first figure refers to coverage, while the second begins to indicate usage patterns. Neither figure can directly prove by how much productivity has increased, but they do indicate that large companies deploying generative AI is no longer just a pilot project for a few departments.

Microsoft also announced that in the past month, there have been over 26,000 intelligent agents active within the Volkswagen Group, with their use extending from engineering and production to procurement and sales. Such a scale can easily create the illusion that "factories are now operating automatically by intelligent agents." In fact, the official statements do not disclose the definition of these intelligent agents, the frequency of their use per agent, or the number of individuals using them independently, nor do they provide comparative data on time savings, quality improvements, or financial returns. A more cautious statement would be that Volkswagen is moving Copilot from general chat to work processes, but the effectiveness still requires verifiable business indicators.

The reason why this news is important is not just due to the large number of employees at Volkswagen. Automotive groups typically have multiple brands, regions, and supply chain levels, with engineering documents, production records, procurement contracts, and sales materials scattered across different permission systems. A single department's demonstration can be completed by manually cleaning data; however, to enable hundreds of thousands of people to use these systems reliably, issues such as identity management, version control of materials, cross-departmental authorization, and accountability for errors become part of daily operations. Whether AI is useful ultimately does not depend on how well it performs in a single demonstration, but rather on whether employees can find reliable information for their specific tasks, complete their work, and know when human verification is necessary.

From chat tools to work processes, what's truly difficult is connecting permissions with responsibilities.

There is a clear sequence to the public promotion process: first, provide employees with Copilot Chat on a wide scale, and then gradually expand to Copilot that includes Work IQ as well as more complex business scenarios. This sequence is worth noting. For enterprises, the chat interface allows employees to quickly test, summarize, search, and draft; however, when it comes to in-depth procurement approval, quality records, or engineering changes, the system must understand the organization's document permissions, communication relationships, and business processes. Treating the chat usage rate directly as the process automation rate can overestimate the actual level of implementation.

For example, procurement staff can ask the system to summarize supplier quotes, but whether these quotes comply with internal procurement rules, whether the prices are still valid, and whether they correspond to the contract version must still be verified by business personnel through the official system. Engineers can use AI to search for design specifications, but the system cannot silently incorporate old parameters into new projects. The production department uses agents to compile records of abnormalities, but an unreviewed generated text cannot serve as a basis for making decisions to stop or resume production. The key for enterprises AI is not to eliminate manual processes, but to link suggestions, bases, approvals, and execution records into a chain that can be reviewed later on.

Microsoft mentioned "over 26,000 active agents" in its announcement, indicating that the focus of deployment is shifting from personal question-answering to tools for specific tasks. However, the number of agents itself may include both simple question-answering assistants and process assistants that connect multiple systems, with a significant difference in complexity. Readers need to further inquire: who created them, who approved the connection of data, how many actual business tasks have been executed, and who handles failures? Without this information on the underlying numbers and responsibilities, the total number of agents only indicates the enthusiasm for creation and testing, not the actual business benefits that have been achieved.

Volkswagen Group Data and the person in charge of AI, Sophie Vassarotti, emphasized in the announcement that they aim to provide employees with a safe and reliable AI support system to help them make decisions more efficiently. This is the description of the deployment goals by the company's management, rather than the effectiveness verified by third parties. For investors and corporate clients, more comparable subsequent evidence should include: the actual usage rate of high-frequency tasks, changes in processing times, the proportion of manual reviews, employee satisfaction, as well as incidents of errors or information leaks. If only the number of people covered is announced without any information on the quality of use, it is difficult to determine whether this is a system that has been thoroughly implemented in the workplace or merely a large-scale software distribution.

The next step in data infrastructure construction is still just an intention; it is not a completed result yet.

Microsoft and Volkswagen also stated that both parties intend to expand their cooperation and further enhance Volkswagen's Intelligent Data Foundation on Microsoft Fabric, making the credible data within the group easier to discover, understand, and utilize. The announcement uses the term "intend" rather than declaring that the entire data infrastructure has already been migrated, nor does it provide a scope or timeline for implementation. This limitation is important: whether the AI assistant can provide reliable answers depends to a large extent on whether the underlying data is complete, whether updates are timely, and whether the access boundaries are clear; data engineering is a long-term project that will not be automatically completed with a single press release.

The information of large multinational corporations is often organized by brand, country, and function. When employees ask questions within a single interface, they may still be dealing with data owned by different parties. Even if it is technically possible to gather this information, compliance regulations may not allow everyone to access all of it. A better system should inform employees about which document the answer comes from, when the document was updated, and why they have the permission to read it; when permission is not granted, the system should reject the request or provide a prompt for application, rather than relying on models to guess at the hidden content. Only if the data infrastructure is solidly established can Copilot transform from a useful text tool into a reliable business entry point.

This is also a crucial battle in Microsoft's competition with AI. Model capabilities will continue to evolve, but companies do not easily rebuild their identities, data governance, and office processes. Microsoft has office software, cloud platforms, and low-code tools, which allow it to attempt to integrate AI into the interfaces that employees are already using. The cases of large corporations provide a scale example for this approach, but it is still not a universally applicable solution in the industry. The depth of processes in manufacturing companies, the structure of factory data, and security boundaries differ from those of consulting, retail, or financial institutions; the number of people covered by one group cannot be directly extrapolated to other companies.

In the short term, this announcement verifies the step of "whether it can be deployed in large enough organizations," and the public has already given a positive answer to that. The next step is to verify whether "it can continuously improve work outcomes within auditable boundaries." If in the future we only see an increase in the number of licenses and agents, without any indicators for task completion rates and quality, then this automotive group will still remain at the stage of procurement discussions. Only if they can provide baseline data for specific scenarios that can be reviewed will their experience truly have replicable value.

Source: Announcement by Microsoft and Volkswagen Group on October 21, 2026 https :// news.microsoft.com / source /2026/10/21/ microsoft-empowers-more-than-330000-volkswagen-group-employees-with-copilot /