What AI is good at here
- Spotting risk in language: a champion who sounds unsure, a "third time this month" in a ticket, a competitor mentioned in passing.
- Connecting signs across systems: a usage drop, a reopened ticket and a quiet champion on the same account.
- Drafting the next step: a short email that cites what's happening on the account.
For how AI compares with health score rules and machine learning, see churn prediction for B2B SaaS.
Where it goes wrong
- Unsupported claims. A model can describe a risk that isn't in the data.
- Inconsistency. The same account can get a different answer on a different day.
- Duplicates. The same issue gets raised three times in different words.
- Overreach. The AI closes a risk or emails a customer without a person checking.
Guardrails that make it trustworthy
- No quote, no signal. Every risk must cite dated evidence, or it's thrown away.
- AI finds evidence; rules do the scoring. Keep the arithmetic out of the model so the score is stable and explainable.
- Merge duplicates by evidence. Two signals that cite the same call or ticket are one signal, however they're worded.
- People close risks. AI can suggest a risk looks resolved; a person decides.
- People approve customer email. Drafts are reviewed and sent by the account owner.
- Respect permissions. Each person sees only the accounts and data they can already see.
- Measure it. Compare accounts you acted on with a random holdout, so you know whether acting on the signals changed renewals.
- Know where the data goes. Check which model provider processes your calls and email, whether it's used for training, and how long it's kept.
A rollout plan
- Connect the CRM and one conversation source (calls or tickets).
- Review the first week of signals with the team: which were right, which weren't, and why.
- Add the other sources, then turn on scoring.
- Start one playbook, with a holdout, and review results after one renewal cycle.
- Keep tracking the share of signals the team accepts or dismisses. If it falls, find out which kind of signal is wrong.
This is how Horizon is built: see renewal risk signals and playbooks and outreach.
More on this topic: churn, definitions, guides and tools.
Updated November 3, 2026