[entrustai]

The intermediary approach shifts trust rather than eliminates it — you're trading provider visibility for intermediary visibility. The more meaningful architectural question is what the intermediary retains after the request completes.

Zero-retention at the inference layer — processing the prompt purely in memory, logging only metadata (latency, verdict, rule hits), and discarding the payload immediately — reduces the exposure surface considerably. It doesn't solve identity linkage at the payment layer, but it means there's nothing to subpoena, breach, or misuse at the content layer.

For regulated industries this distinction matters a great deal. "We never stored it" is a much stronger compliance position than "we stored it but encrypted it."