Skip to content
ADVISORY RECONCILIATION · STRIPE ↔ ENTITLEMENTSRun the browser-only CSV audit
The journal
QualityNovember 14, 2026 · 7 min read

Why billing audits produce false positives

A false positive is not harmless. It trains the team to ignore the audit. Here are the failure modes worth designing against.

The credibility of a reconciliation system is built on precision, not volume. One convincing false finding can cause a team to ignore the next ten real ones. Most false positives come from a small set of avoidable mistakes.

The usual suspects

  • Partial snapshots: the Stripe export stopped at page two, so active subscriptions look missing.
  • Stale snapshots: a cancellation happened after the last sync; the finding is real in time but not now.
  • Identity drift: a customer changed their billing email or created a second customer record.
  • Timezone boundaries: a period ended at midnight UTC, while the application evaluates in local time.
  • Manual grants: support extended access and did not record the exception.
  • Multiple subscriptions: one customer has an old canceled subscription and a current active one.
  • Deleted test records: a test customer was removed from Stripe but remains in the application.

Require a snapshot contract

Each source should declare whether the snapshot is complete, when it was captured, and how many rows it contains. The reconciler should refuse to produce high-confidence findings from an incomplete snapshot. A partial run can still be useful for debugging, but it must be labeled as partial.

Deduplicate by fingerprint, then preserve history

A stable fingerprint prevents the same issue from generating a new row every run. But do not delete the finding when a later run no longer sees it. Mark it resolved, record when, and keep the evidence. If the issue reappears, reopen the existing record with a new detection timestamp.

The fingerprint should be built from stable identifiers, not display text or list position.

Make review cheap

Show the source timestamps, the policy applied, and the specific IDs involved. Give reviewers a way to mark an exception with an expiry date. Roll those exceptions into the next comparison instead of rediscovering them every run.

A useful audit is one a human can trust without redoing the entire query.

FROM READING TO EVIDENCE

Your exports are more interesting than our examples.

Run the local reconciliation and see which categories actually appear in your data. The audit needs no account.