9.2 Agent Analytics and Agent Optimization
Key Takeaways
- Agent Analytics is built on the Session Tracing Data Model and shows Tableau Next dashboards for Service Agent Analytics and Employee Agent Analytics.
- Effectiveness metrics include deflection rate, escalation rate, abandonment rate, engagement rate, and success rate, and a session counts as engaged only when at least one action runs.
- Agent Optimization's Sessions & Intents page shows processed sessions with intent summaries, intent tags, quality scores, and session outcomes.
- Intent clusters form only when at least 10 semantically similar intents exist, and clustering runs about once a week.
- Session Outcome is Escalated, Abandoned, Deflected, Ambiguous, or NOT_SET, based on escalation status plus LLM-evaluated deflection and abandonment scores.
9.2 Agent Analytics and Agent Optimization
Quick Answer: Agent Analytics answers "what's trending?" It aggregates all sessions into dashboards for effectiveness, usage, quality, health, trust, and voice, with Performance Insights breakdowns by subagent, intent, or action. Agent Optimization answers "why?" Its Sessions & Intents page shows individual sessions, intents, intent tags, quality scores, session outcomes, and traces. Use analytics to find the problem, optimization to diagnose it, the builder to fix it, and Testing Center to verify the fix.
Agent Analytics
Agent Analytics is built on the unified Session Tracing Data Model (STDM), which captures every turn and event across sessions. It replaces legacy Agentforce Analytics, which used older Data 360 objects based on copilot events. Salesforce advises existing users to upgrade.
Navigating
In Agentforce Studio → Analytics (Observe & Optimize), choose the installed Service Agent Analytics or Employee Agent Analytics app, select an agent or All, and set filters such as timeframe, channel, and modality.
| Tab / dimension | Metrics | Data source |
|---|---|---|
| Overview | Summary cards and trends across areas | STDM + Audit & Feedback + RAG Quality |
| Effectiveness | Deflection, escalation, abandonment, engagement, success, and task resolution rates | STDM |
| Usage | Unique sessions, interactions, users, and average interactions per session | STDM |
| Quality | Quality score (1–5), answer faithfulness, answer relevance, context relevance | Audit & Feedback + RAG Quality |
| Health | Interaction latency, session duration, error rate | STDM |
| Trust | Instruction adherence response rate, average toxicity score | Audit & Feedback + Trust Layer |
| Voice | Interruption rate (voice agents) | STDM |
| User Satisfaction | Thumbs up/down counts and trends (Employee agents) | Audit & Feedback |
| Performance Insights | Break down any metric by subagent, intent, or action | STDM + Audit & Feedback |
Effectiveness metric definitions
| Metric | Definition |
|---|---|
| Deflection rate | Share of sessions that ended without escalation |
| Escalation rate | Share of sessions escalated to a human or another agent |
| Abandonment rate | Share of sessions abandoned, such as a timeout or drop-off. The default session timeout is two hours |
| Engagement rate | Share of sessions with a substantive reply tied to agent-triggered work. A session is engaged only when at least one action is triggered |
| Success rate | Share of interactions that completed agent action steps without errors |
Performance Insights is the step between seeing a trend and reviewing sessions. Break the metric down by subagent, intent, or action first, so you know where to look instead of reviewing sessions at random.
Sandbox note: Agent Analytics shows only data stored in the current org. Sandbox agent usage records are written to the production org's usage event data, tagged by sandbox org ID and environment type, and Digital Wallet in production is the source of truth for consumption.
Agent Optimization
Agent Optimization is the session-level component of Agentforce Observability. It helps you dig into unresolved interactions, find knowledge gaps, and analyze sessions.
| Feature | What it provides |
|---|---|
| Intents | Groups of interactions in a session that address one user request |
| Quality scores | LLM-calculated relevance of the agent's response to the request (High, Medium, or Low in the interaction summary, with reasoning) |
| Session analysis | Intent-level view of what users ask and how agents respond |
| Trend identification | Low-performing subagents by quality score, misinterpretations, poorly handled conversations |
Access: assign the Access Agentforce Optimization (Agent Optimization) and Data Cloud User permission sets. Setup requires Agentforce Session Tracing and Data Model and Agent Optimization turned on in Einstein Audit, Analytics, and Monitoring Setup.
The Sessions & Intents page
- Processed sessions have passed through the pipeline: intents extracted, clustered, and scored. They're typically available 4–5 hours after the session closes.
- Unprocessed sessions are active or recent. The Trace tab is available immediately, so you can debug without waiting.
- Columns include session ID, duration, outcome, custom scorer labels, intent summary, response summary, subagents and actions triggered, intent tag, and quality score.
- Filter by modality (voice, chat, or all), review voice recordings, export transcripts, and view average agent response latency (from the end of the user's message to the first token of the response).
- Email sessions created in real time aren't reflected in the Sessions & Intents table.
Intent pipeline and clustering
| Parameter | Value |
|---|---|
| Session considered closed | Explicit close, or after 3 hours of inactivity |
| Intent scheduler | Runs every 3 hours |
| Intent pipeline duration | Typically 4–5 hours, up to 24 |
| Clustering cadence | About once a week, over roughly the last month of intents |
| Minimum cluster size | 10 semantically similar intents for a cluster and tag |
| Untagged intents | Don't appear in the session trace; may be reconsidered in later weekly runs |
Session outcome logic
At session close, Agentforce evaluates a Deflection Score (1–5: were the user's goals achieved?) and an Abandonment Score (TRUE, FALSE, or UNSURE: did the user disengage because of dissatisfaction or an unresolved issue?).
| Outcome | Criteria |
|---|---|
| Escalated | Transferred to a human agent |
| Abandoned | Not escalated, and Abandonment Score is TRUE or Deflection Score is below 3 |
| Deflected | Not escalated, not abandoned, and Deflection Score is 4 or higher |
| Ambiguous | Not escalated, not abandoned, and Deflection Score is exactly 3 |
| NOT_SET | No session end signal received |
Session Outcome was introduced in Spring '26 and uses LLM-based scoring starting in Summer '26. Both scores include reasoning you can review, and you can customize the evaluation with scorers.
Reviewing and feedback
From a session, select an intent to see its tag, quality score, feedback, triggered subagents and actions, and evaluations. Reviewers can submit feedback using intent-level or session-level feedback templates (labels, true/false, free text, number, date, or URL).
Agent Optimizer (beta)
Agent Optimizer lets you ask natural-language questions in Agentforce Studio Analytics ("Sample sessions from last week and identify failure patterns"). It samples sessions (about 100 by default, up to 1,000), identifies patterns and possible causes, and suggests improvements. Diagnosis is sample-based, and sessions less than about 30 minutes old are excluded.
Worked Example
Symptom: Service Agent Analytics shows the escalation rate rising from 12% to 21% over two weeks.
- Performance Insights shows most escalations in the Returns subagent.
- Sessions & Intents, filtered to the Returns intent tag and Escalated outcome, shows low quality scores. The reasoning says customers ask about returning opened items.
- The Trace shows the agent calling Answer Questions with Knowledge. The retriever returns chunks about unopened items only.
- RAG metrics show high faithfulness but low context relevance, which points to retrieval (section 6.4). The opened-item policy article was never published.
- Fix: publish the article, then confirm the Knowledge data library updates.
- Retest with Testing Center cases generated from knowledge sources, then monitor the escalation rate.
Exam Traps
- Analytics = aggregated trends; Optimization = individual sessions and intents.
- Engaged sessions require at least one triggered action.
- Clusters need at least 10 similar intents. Low volume means no tags yet.
- Deflection score of exactly 3 = Ambiguous; 4 or higher (and not abandoned) = Deflected.
- Real-time email sessions don't appear in Sessions & Intents.
A manager sees the escalation rate rising in Agent Analytics and wants to know which subagent drives it before reviewing individual sessions. What should they use next?
In Agent Analytics, when is a Service agent session considered engaged?
A session wasn't escalated, the Abandonment Score is FALSE, and the Deflection Score is 3. What Session Outcome does Agent Optimization assign?
A new agent handles only a few similar requests per week, and its intents show no intent tags in Agent Optimization. What is the most likely reason?