Volkswagen Group has equipped more than 330,000 employees with Microsoft Copilot, and more than 26,000 agents were active across the group in the past month alone, Microsoft announced on 21 October. That is enterprise AI agent deployment at a scale where the hard question is no longer access. It is what all those agents remember, and whether what they remember is still true.
What the announcement actually shows
The rollout was phased. It began with Copilot Chat, which reached upwards of 300,000 monthly active users, according to the same Microsoft announcement. Employees now use Copilot with Work IQ in engineering, production, procurement and sales. Microsoft and Volkswagen also intend to expand the group's Intelligent Data Foundation, its unified data ecosystem on Microsoft Fabric, to make trusted data easier to find, understand and use.
"The real value of AI comes from making it available at scale in everyday work," said Sophie Vassarotti, Head of Group Data and AI at Volkswagen Group, in the announcement. Agnes Heftberger, Chair of the Management Board of Microsoft Germany, said Volkswagen shows how companies can move AI from a personal productivity tool into a foundation for future innovation, in the same release.
Enterprise AI agent deployment is now a record-keeping problem
One agent with a wrong fact is an error. Twenty-six thousand agents sharing wrong facts is a system. A technical analysis published on 18 October made the underlying point in small: similarity search returns the closest text, not the current truth. In that analysis, a stale statement and its replacement can look almost identical to an embedding model, two agents can post conflicting values with no flag that they disagree, and when a fact turns out wrong there is no provenance showing which statement it came from, when it was recorded, or what it replaced.
That is the failure mode behind memory poisoning: stale and conflicting entries settle into agent memory and are read back as fact. It is also the open question in our earlier post on Microsoft's team knowledge base. A curated knowledge layer makes agents faster. Who keeps it true once the deployment runs to tens of thousands of agents?
Scale multiplies each failure. More agents write more notes. More notes raise the chance that an old price, an old process, or a superseded decision is retrieved first. Without a record that tracks what is true now, which sources disagree, and why a value is believed, retrieval alone cannot settle the question. It can only repeat it faster.
What the numbers do not prove
The figures are vendor reported, and the announcement gives no productivity baseline, no error rate, and no detail on how the 26,000 agents are governed. That absence matters. A separate analysis of EY's Copilot rollout, also published this week, warned readers to separate availability from adoption, to define the workflow and its accountable owner before counting gains, and to count oversight and correction in the full cost. The same test applies here. Agents created and agents active are input measures. Completed work that a person checked and would stand behind is the output measure.
That is why Jilu treats the company record, not the chat window, as the product. Agents need a shared record that people can open and read, where a decision holds its date, its owner and its sources, where a conflict stays visible until a person resolves it, and where knowledge an agent drafts waits for review before it joins the record. At 26,000 agents, memory is not a feature of the model. It is the governance of the company, written down.