Week 0: define what an "error" and "exception" means, and add two cheap telemetry points
Before you sample, write one clear sentence that says what counts as an error for this automation (wrong contact, wrong invoice, duplicate order, wrong lifecycle stage). Also define an exception (system timeout, API error, missing field) so reviewers can separate process faults from infrastructure problems.
Add two lightweight telemetry items your automation can set: a short processing flag (e.g. processing_state = queued|processing|done) and a single outcome tag (outcome = success|error|exception). These can be custom fields in HubSpot/Salesforce, a small JSON payload in a log, or a single column in a shared spreadsheet. The operational principle is platform‑neutral: one state flag + one outcome marker makes sampling and counting reliable.
Decide the sample method: either a 1% canary (randomly let 1% of live records go through the automation) or a time‑boxed sample (run the automation for 2–3 business days). Aim to collect 50–200 processed cases; that gives a usable error signal in a fortnight for most small teams.