Monitoring AI Agents in Production: Drift, Not Uptime

AI agent monitoring is about whether the agent's judgment is still right, not only whether the system is up. What drift is and how to catch it early.

, Lead Forward Deployed Engineer · 4 min read

On this page

AI agent monitoring is not only a question of whether the system is up. A voice agent can answer every call, with no errors in the logs, and still be wrong. The failure that matters in production is drift: the agent’s judgment slowly moving away from what you intended. Availability tells you the agent is running. Only monitoring the quality of its decisions tells you it is right.

We cover the Deskflow version of this in our post on monitoring AI judgment in production. This one is about voice and WhatsApp agents.

What drift is

Drift is a change in how the agent behaves with no change you deliberately made. It has several sources:

  • The inputs change. New products, new campaigns, a new way customers describe a problem.
  • Your rules change. A policy gets updated, but the agent’s instructions or the test cases do not.
  • The underlying model changes. A model update shifts how it phrases, decides or refuses.
  • Your systems change. A field is renamed, a response gets slower or a lookup returns different data.
  • The mix of callers changes. A new market, a new language variant, a new time of day.

None of these throws an error. The agent keeps answering. That is why “is it up?” is the wrong question.

Why uptime is not enough

Uptime answers: is the system running? For a voice agent, it also helps to know that the platform underneath is reliable, and it matters. But a running agent that routes calls to the wrong queue, offers something it should not, or fails to hand over when a customer is upset is a running failure.

The cost shows up late, in complaints, repeat contacts and rework, not in a dashboard that turned red. In the telecom routing case, misrouting was the cost multiplier: each wrong transfer was a human interaction that should not have happened. A drift in routing accuracy shows up as that cost returning.

What to watch instead

Monitor the outcomes your rules define.

Decision quality.

  • Share of calls routed correctly.
  • Share resolved versus handed over, and whether that ratio moves.
  • Handovers with missing context.

Rule adherence.

  • The AI and recording disclosure at the start of the call.
  • Identity verification before account details.
  • Offers made outside the allowed list.

Customer signals.

  • Repeat contacts about the same issue.
  • Requests for a human in the first seconds.
  • Abandoned calls and calls ending early.

Operational signals.

  • Response latency in live voice, which affects how natural the call feels.
  • Failures in the integrations the agent depends on.

How to catch drift early

Sample every day. A person reads a sample of conversations, not only the flagged ones. This is part of the rollout: we begin with 30–100 interactions a day with daily review, and the daily habit continues after launch in a lighter form.

Keep a fixed test set. Run the same set of cases against the agent regularly. If a case that used to pass now fails, something changed.

Compare against a baseline. Track the metrics from before the agent. A slow drift in a number is visible only if you remember what it was.

Alert on change, not only on failure. A sudden shift in the share of calls handed over is a signal even if each call looks fine.

Use AI to help review. A second model can score conversations against your rules and surface the ones a human should read. Human and AI evaluation work together.

In our Deskflow work, we caught drift in the agent’s judgment in about 15 minutes. That figure is for that system and that kind of check.

What to do when you find drift

  1. Pause or narrow. If the drift is serious, route affected call types to people while you investigate.
  2. Find the cause. Inputs, rules, model, systems or callers.
  3. Fix and re-test. Add the case to the fixed test set so it cannot return silently.
  4. Record it. Write down what changed and how it was found. Over time this is how a system improves after deployment.

Who owns it

Monitoring needs an owner. On your side, someone who knows the rules and can decide what correct means. On ours, the team running the agent. If nobody is responsible for reading the conversations, drift goes unnoticed until it is expensive.

What this means for buying an agent

Ask any vendor: what do you watch after launch? How quickly would you know if the agent’s decisions got worse? Who looks at the conversations? An answer that begins and ends with uptime is incomplete.

If you already run an agent and want to review how you would notice drift, bring your current metrics and we will go through them with you.

Related articles

Bring one process. We'll tell you what it's worth to automate.

Or hear the agent first: leave your number and it calls you in under 5 minutes.