Running architecture in house meant our own team, not an outside firm's second set of eyes, was responsible for catching our own errors before they became expensive change orders in the field.
I built an adversarial review layer into the in-house design process. Specialist agents whose job was specifically to look for contradictions in the model, not just help produce it faster.
Speed without a check is how a fast in-house process turns into an expensive field mistake. If you're going to remove the outside firm's second opinion, you have to build your own second opinion back into the process on purpose, structured so it can't just agree with the first pass.
The review system self-caught a nine-and-a-half-foot bay error and a wrong setback, worth roughly $250,000 in avoided rework, without a human finding either mistake first.
If you remove an outside party's independent check to move faster, you have to deliberately build your own independent check back in, or speed becomes the thing that costs you.
A human still owns the decision to trust or override what the system flags. The tool caught the error, but a person still had to act on it. The system's in service of better judgment, not a replacement for it.