AI support vendor quoted 40% deflection, called 8% normal after 8 months

went live with an AI support bot last january. connected it to our help center, trained it on our top 12 ticket types, gave it 6 weeks to learn. by month 3 we were at 6% deflection. month 8 we hit 8% and stalled. our account manager kept sending benchmark decks showing 7-12% was "typical for complex B2B" and for a while we just believed it. we even renewed because the deflection numbers looked fine relative to whatever PDFs he was sending over. what actually cracked it open was a founder i met at SaaStr in may. his team was hitting 47% deflection on about 900 tickets a month, billing and onboarding questions mostly, same general product category as us. i assumed he was measuring it wrong. he wasn't. he walked me through the setup and the difference was architecture, not training or prompting. his tool was built around resolution from day one. ours was a ticketing system with an LLM wrapper on top and they called it "AI customer service." we started re-evaluating and every single demo ended up being the same conversation: is the AI the actual core of this thing or just a layer sitting on top of a routing system. completely different product philosophies, and apparently a 39-point deflection gap between them in practice. still haven't switched yet so i don't have a clean before/after. but if 8% is what most teams are actually hitting then either we bought something broken or this whole category is one big benchmark hallucination.