They understood that "add AI" doesn't mean wrapping an API call around existing logic. The agentic architecture their team proposed was actually designed around how our infrastructure behaves under load. Their ML engineers (around 3+) spent real time learning our failure modes before writing a line of code. That matters when the system has to make autonomous decisions at scale without a human in the loop. Review collected by and hosted on G2.com.
Scope expanded mid-project. We kept surfacing edge cases in the routing logic that needed handling, and that added time and a bit of extra money we hadn't planned for. Review collected by and hosted on G2.com.