Ada ditched containment for resolution — and doubled what AI can solve

Ada, the $1.2 billion AI-native customer service platform, is publicly breaking up with the industry’s favorite vanity metric: containment rate. CEO Mike Gozzo argues that boasting an 80-100% containment rate is meaningless if the chatbot is just trapping people in a loop of useless answers. It’s a blunt diagnosis of a problem most vendors prefer to ignore, and it explains why Ada went back to the drawing board in 2022, betting the company on a rebuild powered by OpenAI’s reasoning models.

That rebuild produced a custom evaluation framework, initially built on GPT-4, that judges conversations based on whether they were actually resolved — relevant, accurate, and safe, with no human handoff. In testing, Ada’s evaluator agreed with human judgment 80–90% of the time. That alignment gave them a north star metric that exposed just how hollow their old numbers were: the legacy system sat at 70% containment but a paltry 30% resolution rate. The new multi-agent system, which runs customer questions through multiple turns of OpenAI models before generating an answer, flips the script entirely.

Early migrations are seeing resolution rates hit 60%, with top performers exceeding 80%. That’s literally double the conversations getting resolved well, which Gozzo says translates directly into full-time equivalent savings, better retention, and more signups. The technical underpinnings are a multi-agent architecture — a central planner dispatching work to subagents — all running on OpenAI’s API. Ada stresses that no competitor model has beaten OpenAI on their internal evaluation set, and they are using fine-tuning to generate confidence scores that help suppress hallucinations across the toolchain.

Maybe the most telling shift isn’t technical but cultural. Gozzo notes that enterprise buyers have stopped asking if automated resolution is possible and started asking how fast they can deploy it. Ada’s public goal is now a 100% resolution rate, which the CEO frames as a matter of “when, not if.” That’s either refreshingly ambitious or a recipe for overpromising, depending on how much you trust synthetic test frameworks to mirror the chaos of real customer anger at 11 p.m. on a Saturday.