AWS wraps five quality layers around the LLM assistant used by over 4,000 of its leaders

On accuracy and speed, AWS's post puts it plainly: "a capable large language model (LLM) alone can't guarantee either" ^[6]. Throttling is the clearest case. It is a quota problem. NarrateAI treats inference capacity as a set of independent model-account quota spaces, and its failover spreads requests across them to reduce the throttling users see ^[7]. Anyone who has waited on a quota increase ticket will understand why the answer was more accounts. Adopting it has an operational cost. Someone has to run those accounts. An answer may also come from a different model than the previous one, so every downstream evaluator has to hold each model to the same bar.

The router exists because the data behind a question varies widely while the deadline stays fixed. A focused question draws on a handful of document sections, a full regional analysis on hundreds, and users expect the same response time for both ^[9]. Concatenating everything into one prompt breaks once the sections exceed the context window, typically 200K tokens, and the forced truncation costs quality ^[10]. Always splitting into several passes avoids truncation but pays for multiple LLM calls on every query ^[11]. NarrateAI routes by data volume, so most queries finish in one pass and the large ones fan out in parallel ^[8].

Validation latency is handled by overlap. Each paragraph is checked as soon as it is produced, while generation continues ^[12]. Several independent evaluators run against that paragraph in parallel ^[13]. The unit of checking is the paragraph. A check that needs the whole answer, such as whether a figure in the opening agrees with one near the end, cannot run until the last paragraph exists.

The number check is the most careful piece of engineering in the post. Exact matching runs first because it is cheap, and semantic verification runs "only when needed," in the post's words ^[14]. A figure copied straight from a retrieved section can be settled in the cheap stage. A computed figure, such as a gap to target or a change between quarters, is less likely to appear verbatim in any source section, so it has to go through the semantic stage. Whether AWS's result carries over to another team depends on that team's mix of quoted and computed figures.

AWS reports approximately 99 percent numerical accuracy with the five techniques running together ^[4]. By its own figure, the remaining error rate is about 1 percent ^[16]. The post says a wrong number in front of leadership "carries immediate professional consequences" ^[15]. The passage that introduces the 99 percent does not give a measurement method, a query count, or a breakdown by layer.