The signals that actually precede churn — what health scores hide, and what we keep learning from the conversations teams can't afford to miss.
An AI scorer can grade every support conversation, but its score doesn't show whether the verdicts hold still or match your reviewers. Three tests to run before the scores reach a coaching session or a team average.
ai agentsA support person who isn't sure of an answer can check the account or ask a colleague. An AI agent answers from what it was given and sounds the same either way. Two public cases, and a ten-conversation check you can run on your own agent.
qa scorecardsMore support teams now have an AI agent answering customers alongside their people. Which QA criteria carry over to the bot, which ones don't, what to add, and why its scores belong outside the team average.
churnAn account cancels and the CRM records "budget." Two months earlier, the same account contacted support three times about the same problem. How to check a cancel reason against the support history before acting on it.
support qaA QA scorecard lists criteria and points. The rubric is the written definition that tells an evaluator what earns those points. How to write criteria, scales, anchors and rules that two evaluators apply the same way, and how to check that they do.
customer evidenceA customer conversation gets filed by where it happened, not by what it contains. When each team reads only its own slice, a renewal risk can sit in a closed support ticket that the CSM never opens.
support qaA QA scorecard turns a conversation into one number. How that number is calculated, why two teams reporting 85 percent cannot assume they measured the same thing, and what to report instead.
support qaMost QA programs score four to eight conversations per agent per month and use the result to rank people. At that size, the sample supports coaching from specific conversations but not rankings or trend lines.
benchmarkGPT-5 ranked our benchmark accounts at 0.75 AUC reading raw conversations, and at 0.93 reading the distilled evidence from the same accounts. One model, one change, and the change was to the input.
support qaA weight on a QA criterion is a bet about what matters to the customer, and almost nobody checks the bet. How to test a scorecard against what the account did next.
promise-experience gapThe gap between what a customer was promised and what they experience is recorded in two systems that do not reference each other. A manual audit for reading the support queue against the sales narrative, and where the evidence already lives.
Customer IntelligenceStibo Systems and Accenture spent twenty-one pages on the data an AI agent needs before it acts. All of it describes what a customer is. None of it describes how the relationship is changing.
Customer IntelligenceEvery governance answer for AI agents so far is a permissions answer: OAuth, RBAC, scoped tokens, audit logs. All of it describes the request. Almost nobody is specifying what the response contains.
Customer IntelligenceWe pre-registered a churn benchmark, with success bars committed before the test data existed, then ran GPT-5, Claude Opus 4.8, Gemini 2.5 Pro, DeepSeek, Kimi K2, and GLM-4.6 against our own engine on a frozen 200-account corpus, publishing the results and the data to check them.
Customer SuccessYour support team watches accounts drift toward the exit every week: in flattened tone, repeat tickets, and CC lines that go quiet. A field guide to seven signals in the queue, and the evidence to pull for each one.
Customer SuccessEvery tool in your stack holds a piece of the customer's story. Nobody holds the whole thing. What fragmentation looks like inside a CS org, why it happens, what it costs, and a self-assessment you can run this week.
Customer SuccessThe renewal that died was lost in a sales call customer success never heard, months before the QBR where anyone noticed. The evidence of the gap was sitting in two teams' tools the whole time.
Customer IntelligenceA risk score nobody can trace is just an opinion with a number on it. The next question after any signal has to be 'show me where.' Most CS tools can't answer it.
Customer SuccessThe account that churns rarely surprises the people who held the pieces. It surprises everyone because nobody held all of them. That's a fragmentation problem, not an understanding problem.