Skip to content
S SiPesan Docs
ID EN

AI Analytics

Once your AI Agent is running, the AI Analytics page answers one big question: how well is it actually doing? It's separate from the general Analytics page and shows AI numbers only, for the last 30 days. You'll find it in the AI Agent group, and only admins can open it.

The Headline Numbers

The summary row up top gives you a quick read:

  • Total sessions — how many conversations the AI handled.
  • Deflection rate — the share of sessions that finished without a human agent. The higher it is, the more load the AI takes off your team.
  • Escalation rate — the flip side: the share of sessions that ended up handed to an agent.
  • Success rate — the portion of interactions that ran cleanly, with no errors.
  • Average latency — how fast the AI puts together a reply, in milliseconds.
  • Average turns — how many back-and-forth messages before a session wraps up.
  • Total messages and total tokens — the scale of usage.

Tokens here are a measure of usage, not a bill. SIPESAN doesn't charge for tokens; the figure helps you see how "heavy" AI conversations are, not what they cost.

Charts Worth Watching

  • Handling mix — how messages split across the AI, agents, and the chatbot. Useful for seeing who's really doing the most work.
  • Sentiment breakdown — the share of conversations that read positive, neutral, and negative. A spike in negative is your cue to check the AI's answers.
  • Volume trend — daily sessions and escalations across 30 days. Helps you tie traffic spikes to real events (a promo, an outage).
  • Deflection & escalation trend — whether the AI is growing more self-sufficient over time, or giving up more often.
  • Token & latency trend — daily usage and speed. A sharp jump in latency usually points to a problem on the provider's side.
  • Escalation reasons — why sessions moved to a human (turn limit reached, angry sentiment, and so on). This is the most practical map for improving the AI: the reason that shows up most is the one most worth fixing.
  • Provider/model mix — which model gets used most, if you run more than one.

AI Agent Leaderboard

If you run several AI Agents, the leaderboard table lines them up side by side: session count, deflection rate, average turns, and escalations per agent. It shows which persona or setup works best, and which needs work.

How to Read It

Raw numbers don't mean much without direction. A few simple rules of thumb:

  • Deflection up while sentiment stays healthy means the AI is carrying load well. Widen its scope.
  • Escalations piling up on one reason means there's a clear gap — fix the persona instructions, add documents to the Knowledge Base, or wire up a Connector for the missing data.
  • Negative sentiment climbing means review the transcripts in the AI Interaction Log before you widen the AI's scope.