After cutover your team owns first-line operations. This page is what that job consists of.

​
What to watch

Four signals tell you whether the deployment is healthy. Everything else is diagnosis.
How long the caller waits before the agent starts speaking. It is the metric callers actually experience, and the first one to move when the deployment is under pressure. See Latency.
Peak concurrency as a proportion of what the deployment was sized for. This is your capacity early-warning: it rises for weeks before anything breaks.
The share of calls passed to a person. A rising rate usually means callers are asking for something the agent was never configured to do, which is a content problem, not an infrastructure one.
Actions that error or time out against your systems. These are the failures callers feel as an agent that promised something and did not deliver it. Alert on them separately from everything else.
A local dashboard, served from the deployment, shows usage and system health without sending anything outside your network to render it.

​
Licence renewal

The deployment validates a signed licence file locally, so it does not depend on reaching Voho to keep running. If a licence is not renewed in time the deployment degrades gracefully rather than cutting off, a line does not stop answering because a renewal was sitting in someone’s inbox. Treat that as a safety net, not a plan. Put the expiry date in the same calendar as your certificate renewals.

​
Logs and audit

Operational logs stay on your hosts under your log policy. Configuration changes, access events and exports form a separate audit trail that can be streamed to your SIEM, so evidence lives in the system your auditors already read. Keep the two separate. Debug logging is noisy and short-lived; the audit trail is neither.

​
Upgrades

Upgrades are yours to schedule. In a disconnected deployment they arrive as media you import; in a connected one they are still applied on your window rather than pushed.
  • Take the release notes and the upgrade runbook for the version you are moving to.
  • Apply to a non-production deployment first if you have one, or during a low-traffic window if you do not.
  • Keep the previous version available until the new one has held through a real peak, not just through a smoke test.
  • Re-run your test set afterwards. Voice regressions show up in dialect handling and number reading long before they show up in error rates. See Testing.

​
Capacity

Capacity is a planning exercise, not an alert. Review it on a schedule:
  • Track peak concurrency month over month against the sizing.
  • Add headroom before the season you already know is coming, not during it.
  • When you add hardware, re-measure time to first audio, capacity that does not hold latency has not been added.
The number to defend is peak concurrency during your busiest hour of the year. Average utilisation looks comfortable right up until the hour that matters.

​
Records and retention

You own the records, so you own the lifecycle: retention per data type, backup, and the ability to locate and delete one caller’s data across audio, transcripts, summaries and anything the actions wrote downstream. Test the deletion path before someone requests it. See Security and data handling.

​
When to escalate

Keep the boundary clear between what your team fixes and what goes back to Voho.