The ATLAS experiment relies on real-time hadronic jetreconstruction and b-tagging to record fully hadronic eventscontaining b-jets. These algorithms require track reconstruction,which is computationally expensive and could overwhelm thehigh-level-trigger farm, even at the reduced event rate that passesthe ATLAS first stage hardware-based trigger. In LHC Run 3, ATLAShas mitigated these computational demands by introducing a fastneural-network-based b-tagger, which acts as a low-precisionfilter using input from hadronic jets and tracks. It runs after ahardware trigger and before the remaining high-level-triggerreconstruction. This design relies on the negligible cost ofneural-network inference as compared to track reconstruction, andthe cost reduction from limiting tracking to specific regions of thedetector. In the case of Standard Model HH → bb̅bb̅, a key signature relying on b-jet triggers, the filter lowers the input rate to the remaininghigh-level trigger by a factor of five at the small cost of reducingthe overall signal efficiency by roughly 2%.