Volume is the part of this disclosure that transfers to other companies. About 45,000 questions a month divides to roughly 1,500 a calendar day, or about 2,100 per working day on a 21-day month [12]. If that traffic is spread across the hundreds of channels Uber says carry on-call demand [3], and "hundreds" means somewhere between 200 and 500, each channel is seeing perhaps 90 to 225 questions a month, four to eleven per working day [14]. That is the shape of a problem that rarely gets funded on its own: no single team's queue looks bad enough to justify a platform, while the aggregate is a standing charge against engineering time [2].
Internal support pays back for reasons that have little to do with model quality. The corpus belongs to the employer, and the audience can usually tell a wrong answer from a right one. The cost of a miss is a wasted minute rather than a commitment made to a customer. Uber describes what it faced as a findability problem rather than a generation problem [3], and its preference for retrieval-augmented generation on time-to-market grounds [4] fits a body of documents that changes faster than any training cycle would.
The board-deck version of this reads "AI now handles 45,000 questions a month." The post does not claim that. It gives the channel volume, the architecture of scraped wiki and internal Stack Overflow content embedded with an OpenAI model into a vector database and retrieved into an LLM prompt [5], and three hallucination controls: relevant retrieval, verification against authoritative sources, and keeping the index current [7]. There is no deflection rate, no accuracy measure and no figure for engineer hours returned [13]. The disclosure is a denominator, and anyone building a business case from it supplies the numerator themselves.
The obvious objection is that a copilot sitting over fragmented documentation just launders weak docs into fluent answers. Uber's own controls concede the mechanism, since two of the three depend entirely on the authority and freshness of the sources behind them [7], and the security screen narrows the corpus further, because Uber says many internal data sources cannot be exposed in Slack channels at all [6]. Even so, that caps the ceiling rather than voiding the case. Retrieval quality is bounded by document quality, which turns a documentation problem that was previously ignored into one with a visible score.
The sequencing worth noting is what the retrieval choice moves onto next quarter's budget. Fine-tuning would have bought a model refresh cadence; retrieval buys an ingestion pipeline, and Uber says it runs a custom Apache Spark application to load the vector store [8]. That is a data engineering commitment with an owner and a maintenance schedule, alongside the feedback loop Uber built for users to rate responses [11]. A team approving a copilot like this in one quarter is approving the person who maintains its index in the next, and the question volume is the only part of the case that arrives with a number already attached.
