Before a customer-support rep's question about a shipment ever gets an answer, Amazon Bedrock's Intelligent Prompt Routing has already inspected the request, checked it against a cost and latency budget, and sent it to one of several models in the pool — each with its own training data and blind spots. AWS launched this feature specifically so enterprises would stop hand-picking a single model for every task. The rep sees one seamless answer. What happened first was a dispatch decision with real consequences for what that answer contains.
This is not exotic infrastructure anymore. Startups like Not Diamond and OpenRouter exist entirely to do this routing for companies that do not want to build it themselves, sending a legal question to a more conservative model and a casual product query to something cheaper and faster. The routing rules encode a judgment about which model is 'appropriate' for which kind of question — and appropriate is doing a lot of work in that sentence.
Newsrooms have always made a comparable choice about which reporter or desk covers which story, shaped by expertise and house style. The difference with a Bedrock routing policy is scale and visibility: it applies to every instance of a question type, instantly, and the engineer who wrote the routing rule is invisible to the person receiving the answer.
The people writing these rules are platform engineers optimizing for cost, latency, and liability exposure — a reasonable set of priorities for infrastructure work, but not one that includes the word editorial, or criteria about which answer would actually serve the person asking.
The stakes rise as routing gets more adaptive. A static rule — legal questions always go to the more conservative model — is at least auditable after the fact. A system that learns which model produces fewer complaints for a given query cluster starts to resemble the engagement-optimization dynamics that shaped recommendation feeds: reinforcing whatever scores better, regardless of whether the score tracks accuracy.
Regulators asked for visibility into recommendation algorithms years after those systems had already shaped public discourse. Bedrock, Not Diamond, and OpenRouter's routing policies are still young and simple enough for someone outside the infrastructure team to actually read one end to end. That window will not stay open long, and it is the only point at which outside scrutiny is cheap.