Use stage gates rather than a calendar-based commitment to rollout. This is our suggested operating plan; it is not a claim that every deployment can safely finish in 90 days.
| Period | Work | Evidence required to proceed |
|---|---|---|
| January: define and test | Baseline current handling; approve permissions; map data; execute synthetic and representative tests | Agreed outcome definition, completed risk review, functioning fallback and reproducible test results |
| February: controlled live traffic | Route limited eligible traffic; review failures; reconcile claimed outcomes to records | Outcome accuracy, caller effort, transfer acceptance and fully loaded cost remain within approved limits |
| March: decide | Compare against the baseline and a credible alternative; review peak demand and exceptions | Explicit expand, repair, hold or stop decision signed by operations and risk owners |
What should the measurement ledger contain?
For each unique task, record:
- Eligibility and reason for exclusion, if excluded.
- Final outcome: autonomous completion, human-rescued completion, unsuccessful or still pending.
- Evidence in the scheduling, CRM or service system.
- Permission or policy violations, including those later corrected.
- Repeat contacts within a defined, workflow-appropriate observation window.
- AI, telephony, implementation, monitoring and human-rescue costs.
Our proposed governing metric: total workflow cost divided by unique, correctly completed, policy-compliant tasks. Publish autonomous completion separately so human rescue cannot disappear inside an attractive blended result.
For missed-call recovery, estimate incremental completed jobs against what existing callbacks would have recovered. Keep recovered contribution separate from payroll savings and avoided hiring.