Rehearse the customer journey, not only request volume
Model the countdown page, live chat, product assistant, cart, checkout, and staff console as separate paths. Name the expected peak, customer success measure, AI budget ceiling, and paths that must remain usable. A generic load test can pass while the one revenue-critical assistant route is misconfigured.
Run the rehearsal in a production-like environment with the same cache, proxy, WordPress hooks, model settings, alerts, and rollback permissions. Use synthetic customer data and clearly labeled test orders. Confirm that test traffic cannot contaminate finance or marketing reports.
Gate 1: genuine demand at 1× and 5×
At baseline, verify one trace from browser through provider and one completed checkout. At five times expected genuine demand, measure AI completion, checkout completion, payment failures, p95 latency, provider headroom, tokens per order, and queue depth.
Pass only if customer outcomes remain inside named thresholds and the budget forecast stays below the approved ceiling. If the environment cannot generate representative model work, mark the test partial and do not infer cost performance from HTTP throughput alone.
| Gate | Scenario | Required evidence | Pass decision |
|---|---|---|---|
| Demand | 1× then 5× genuine journeys | checkout, latency, tokens/order | all named customer thresholds met |
| Anomaly | 20 identical requests/min | narrow control and unaffected checkout | exposure falls without broad stop |
| Alert | limit approach and provider 5xx | email/Slack receipt and owner ack | delivered inside target |
| Rollback | bad production-like rule | restore time and post-check | under 5 minutes with evidence |
Gate 2: anomaly without collateral damage
Send 20 identical requests per minute to one non-checkout assistant path while ordinary journeys continue. Verify that the intended repetition or source control activates and that checkout help, cart, and unique product questions remain available.
Inspect Human, Bot, and Unknown signals but score the gate on observable behavior and outcomes. A successful rehearsal does not require every synthetic client to receive a preferred label; it requires the narrow, explainable rule to reduce the intended exposure.
Gate 3 and 4: alerts and rollback
Inject a provider 5xx condition and a budget-approach condition separately. Confirm email or Slack delivery, incident details, recipient acknowledgement, and an alternative route if the primary channel fails. A log entry no one receives is not an operational alert.
Deploy one deliberately bad but safe rule and execute the documented rollback. Time from decision to restored customer path must be under five minutes, then verify settings, one customer journey, and one provider record. The person approving Enforcement should not be the only person able to revert it.
Make go-live conditional on evidence
The checklist must record test version, time, participants, exact inputs, metrics, screenshots or exports, failures, corrective actions, retest, and signatures for all four gates. A failed gate remains visible; do not overwrite it with the successful retest.
Approve production Enforcement only when every gate passes, rollback access is confirmed, temporary launch thresholds have an expiry, and an on-call handoff exists. The guide should receive readers who need the full implementation sequence, while the checklist remains the launch-day control document.
Use the livestream go-live checklist from “Rehearse a livestream product launch before the audience arrives” on a real first installation. Download AI Cost Circuit Breaker for free, begin in Monitoring, and move to enforcement only after the expected signals and rollback are verified.