## TL;DR
Three hundred false flags have a pattern, and the pattern here is the agent reading billing interval from the wrong evidence. Annual accounts misread as monthly look perpetually overdue for renewal behaviour, quiet in the wrong months, and one interval away from expiry, so every heuristic built for monthly accounts fires on them at once. Find where the agent derives interval, rebuild it from the billing systems price and subscription facts, clear the false flags, and segment every cadence based rule from now on.

```text
agent misclassified annual accounts as monthly and flagged false churn risk for 300 accounts
```

## Use this when
- False churn risk flags concentrate on annual accounts while monthly accounts score sensibly
- The flagged count is large and sudden, here about 300 accounts, after an agent or data change
- Billing shows healthy annual subscriptions with months left before renewal for the flagged accounts

## Not for this skill when
- The annual accounts really are near renewal because they were signed in one large cohort
- The agent flags all cadences equally, suggesting a broader scoring problem

## Steps
1. Sample flagged accounts and establish their true interval from billing. For twenty flagged accounts, record the price, billing interval, current period dates, and next renewal from Stripe or Chargebee directly, and confirm how many are annual in fact, because the billing system is the only opinion that counts here and the sample sizes the defect honestly. Verify: the sample true intervals are documented from billing records
2. Find where the agent gets interval. Trace the field or inference it uses, a CRM plan text field, a price name pattern, an amount heuristic such as assuming large invoices are annual, or a default in code that labels anything unstated as monthly, and reproduce the misread for a sample account by following the agents exact inputs, not the billing UI a human would check. Verify: the agents interval source and its failure on the sample are pinpointed
3. Audit the true blast radius. Classify every account the agent tracks by its billing derived interval and the agents interval, count agreements and each direction of disagreement, and identify every rule, score, or alert that consumed the wrong interval, because renewal timing, silence windows, and payment expectations all inherit the error. Verify: the full interval mismatch counts and the consuming rules are listed
4. Rebuild interval from billing facts. Derive it from the subscription price interval and current period length in the billing system, sync it as a single governed field with a last verified timestamp, and make the agent read only that field, refusing to infer cadence from names, amounts, or CRM text that sales processes edit freely. Verify: the agent reads the governed field and a test annual account classifies correctly
5. Clear the false risk properly. Rescore the affected accounts with corrected intervals, close the false flags with a reason that names this defect so reporting does not count them as worked saves, and notify the CSMs who received flags that the cohort was invalid, because otherwise they will waste the week or, worse, discount healthy annual accounts to save them from a phantom risk. Verify: false flags are closed with the defect reason and CSMs are notified with the account list
6. Add an interval reconciliation check. Compare the agents interval against billing weekly, alert on any mismatch beyond a tiny tolerance, and require cadence segmentation in the test cases for every future churn rule, because cadence confusion is a permanent species of bug in subscription businesses, not a one time event. Verify: the weekly reconciliation runs and flags a planted interval mismatch

## Why it happens
Billing interval is a structured fact in the billing system and a folk fact everywhere else, written in plan names, remembered by reps, and guessed from invoice sizes. An agent consuming the folk version inherits every inconsistency, and because monthly accounts are usually the majority, a monthly default looks right in testing while being wrong for the entire annual book. The heuristics layered on top then convert one wrong field into hundreds of confident, wrong risk flags.

Compatibility: Churn monitoring agents, billing interval data from Stripe or Chargebee subscriptions, and CRM plan fields, in current agent and billing platform versions.

## Variants and edge cases
### agent misclassified annual as monthly churn risk
['The direct phrasing. Billing period facts are the ground truth, inferred cadence from names or amounts will always eventually lie.']

### 300 accounts false churn flags billing cadence
['The incident variant. Closing false flags with an honest defect reason protects save rate reporting from counting phantom saves later.']

### annual accounts flagged monthly heuristics
['The rules variant. Audit every cadence sensitive rule the agent runs, silence windows and renewal countdowns break together when interval is wrong.']

Edge cases to work through before you close the incident:
- Price names can contain the word annual for monthly billed annual commitments and other commercial inventions. Never classify on names when the price object carries the real interval.
- Accounts can hold mixed intervals across products. The agent needs a per subscription interval and an account level rule for which subscription drives renewal risk.
- Interval can change at renewal by design. The governed field needs effective dating or every historical score explanation becomes wrong after a switch.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_kaa8yKaddlfFqbQPpGPtDQ