Quick Answer
What is Agentforce Observability, and how does it help customers?
Agentforce Observability is Salesforce’s central application for analysing and improving AI agents through analytics, session tracing, intent clusters, quality scores, alerts and consumption insights. It helps teams understand where a customer journey fails and what to improve. KVP recommends measuring verified customer outcomes, calibrating automated evaluations with human review and making controlled changes — not treating a busy dashboard as proof of a successful AI programme.
Why monitoring AI agents needs more than conversation counts
A customer saying “thank you” does not prove an issue was resolved. A session without escalation could be a successful self-service interaction — or a customer who gave up. The Salesforce Observability overview positions the product around a deeper operational question: what happened inside the agent’s work, and how can it be improved?
Monitoring highlights symptoms. Observability adds the session, action and context needed to investigate causes. Salesforce’s related product update describes a shift from proxy signals to context-based evaluation, with structured session data supporting analytics and alerts. KVP’s view is that this should connect technical evidence to a business definition of success, not simply produce more metrics.
Core Agentforce Observability features
A central view of your AI agent fleet
Agentforce Analytics brings usage, effectiveness and customer-feedback signals together. Administrators can compare agents and investigate changes rather than judge the digital workforce only by conversation volume. Salesforce also describes unit-level analysis and cross-org data with DC1 support; confirm the required configuration before assuming all orgs are connected.
Session tracing and configuration inspection
Inspect a session down to individual interactions, subagents and actions to locate failure points. A slow answer might reflect an external service delay; an incorrect answer might reflect unsuitable knowledge or instructions. The trace helps distinguish these paths so teams change the right part of the solution.
Conversation clusters by intent
Group conversations that share an intent and inspect repeated patterns. If order-status enquiries repeatedly fail while returns work well, the priority is that specific journey, not a blanket rewrite. Review representative samples so frequent topics do not hide rare but high-risk failures.
Quality scores and context-based evaluation
Salesforce describes relevance and helpfulness scores, alongside newer context-based outcome evaluation and editable scoring prompts. Custom LLM-as-judge scores can reflect a business definition of success. These are evaluation signals, not infallible ground truth: validate them against human assessment and authoritative transaction records.
Proactive health monitoring and alerts
Configure near real-time alerts around important metrics and business logic. An alert should lead to a named responder, an investigation and a documented intervention. It does not, by itself, prove that the agent will automatically repair a configuration or that a service-level commitment has been met.
Consumption visibility and financial accountability
Review credit consumption and usage patterns by agent, alongside Digital Wallet information. Connect this to verified task completion instead of celebrating low cost per conversation when the customer still needs help. Standard observability metering and the cost of running agent actions are different questions.
Watch: Salesforce’s Agentforce Observability demonstration
Agentforce Observability: Manage & Improve Your AI Agent Fleet
Video credit: Salesforce. Embedded from Salesforce’s official Vidyard player; copyright remains with Salesforce. Loading it connects to Vidyard, which may process usage data. Watch on Salesforce’s website.
How it can help KVP customers and their teams
Customers: fewer repeated failures and more reliable handovers
Tracing recurring intents helps identify where an answer lacks grounding, a request stalls or escalation loses context. Fixing those issues can improve the customer journey. Measure repeat contacts, completion and customer feedback rather than assuming a lower escalation rate always means better service.
Service and sales leaders: evidence for prioritisation
Journey-level analysis helps leaders decide which issues deserve investment. For a service agent, examine resolution and repeat contact; for a sales assistant, examine accepted outputs and qualified handovers. Different agents need different success definitions, with a baseline established before any change.
Builders and administrators: a shorter path to diagnosis
Session-level evidence can help separate instructions, knowledge gaps, permissions and integration failures. Clustered incidents guide focused fixes instead of speculative prompt changes. Re-test representative successful sessions as well as the original failure to avoid introducing a regression.
Finance and governance teams: more transparent AI operations
Consumption data, documented evaluations and controlled changes create a clearer review trail. Cost per verified outcome can support investment decisions, while restricted session access helps protect sensitive information. Observability supports governance; it is not a compliance certificate or a guarantee of return.
These are potential benefits and recommended measures, not reported KVP customer results or guaranteed improvements.
KVP view: observability is an operating discipline, not a go-live extra
Make accountability part of adoption. An agent needs an owner, an agreed remit, an escalation path and a review cadence. Treat deployment as the beginning of a managed improvement loop. When the system flags a weakness, the team must know who can change the instructions, knowledge or integration and who approves that change.
Existing Salesforce customers have a useful starting point. Customer records, cases, opportunities and workflows can provide the business context for evaluating an agent. However, incomplete records, unclear ownership and outdated knowledge do not disappear when AI is added. Review those foundations before expanding the agent’s responsibilities.
Match the measures to the industry. A manufacturing service agent might be assessed on accurate equipment guidance and correct escalation; a real estate assistant on current inventory information and a properly confirmed site visit; a distributor assistant on a validated order handover. These are illustrative evaluation designs, not claims of completed deployments or measured outcomes.
Where KVP can help: define the customer journey, review Salesforce readiness, map permissions and data sources, design evaluation scenarios and establish a practical operating model. Agree delivery scope and responsibilities first. Better telemetry is valuable only when it leads to safer actions and a better customer experience.
KVP’s six best practices for Agentforce monitoring and optimisation
- 01
Define success before configuring a dashboard
Agree the business event that counts as completion, such as a correctly answered policy question or a confirmed service request. Track quality, escalation, latency, repeat contact and consumption together. Do not optimise deflection at the expense of the customer.
- 02
Review a balanced sample of sessions
Include failures, successes, ambiguous requests, different languages and sensitive intents. Check the source evidence and actual action results. Compare automated evaluation with human judgement and investigate disagreements before using a score as a release gate.
- 03
Separate failure types and ownership
Route knowledge gaps to the content owner, integration errors to the technical owner and policy uncertainty to the business owner. Give critical alerts a severity, response expectation and escalation path. Avoid alert fatigue by tuning thresholds against a real baseline.
- 04
Protect conversations and trace data
Apply least-privilege access, appropriate retention and privacy review to session records. Restrict exports and avoid copying personal data into broadly shared incident reports. Test that the agent cannot use monitoring or troubleshooting as a reason to bypass access controls.
- 05
Make every change testable and reversible
Version instructions, actions, knowledge and scoring prompts. Reproduce an issue in a controlled environment, run regression checks and obtain the relevant approval. Record what changed, compare the resulting sessions and retain a rollback path.
- 06
Review outcomes and spending on a regular cadence
Begin with frequent reviews during the pilot, then agree an operating rhythm based on risk and traffic. Check both customer outcomes and the cost of agent execution. Confirm which standard analytics are unmetered and which net-new customisations introduce charges.
Licensing and Flex Credits: what is included, and what is not?
Salesforce’s Observability update states that standard session tracing, telemetry ingestion, dashboard metrics, out-of-the-box queries, drill-downs and alert evaluation are unmetered across Agentforce SKUs, at no additional Data Cloud cost. Usage can still appear in Digital Wallet without counting against Flex Credits.
The same source explicitly keeps net-new customisation metered: new dashboards, external-data ingestion and new semantic-data-model objects. Modifying or extending existing out-of-the-box dashboards is described as unmetered. This does not mean that agent execution, all Data Cloud use or a bespoke implementation is free.
Confirm your current contract, org access, regional availability and exact feature rollout with Salesforce. The update discusses capabilities rolling out from August; an announcement or demo is not proof that every capability is enabled in every customer org.
Five steps to start with Agentforce Observability
- 1
Choose a bounded pilot
Select one agent and one valuable customer journey. Name a business owner and a technical owner, and document actions that require human approval.
- 2
Confirm readiness and product access
Review agent permissions, knowledge quality, logging, available Observability capabilities and contractual entitlements. Establish how sensitive session data will be handled.
- 3
Baseline the journey
Record volumes, verified completion, escalation, latency, repeat contacts and relevant consumption. Use representative sessions to agree a scoring rubric.
- 4
Activate the review-and-improve loop
Review dashboards and traces, investigate the highest-priority intent clusters, configure actionable alerts and test targeted changes before release.
- 5
Prove the result before scaling
Compare equivalent periods and customer journeys after each change. Expand only when performance, cost and controls meet the agreed criteria; keep ownership in place after go-live.
Agentforce Observability: frequently asked questions
What is Agentforce Observability?
Agentforce Observability is Salesforce’s application for supervising and improving AI agents. It combines analytics, session traces, intent clusters, quality scores, health alerts and consumption visibility to help teams investigate performance and refine agent configurations.
What is the difference between AI agent monitoring and observability?
Monitoring flags a symptom such as increased latency or escalation. Observability adds the session, interaction and action-level context needed to investigate why it happened. Teams still need human review and controlled changes to confirm and resolve the cause.
Does Agentforce Observability consume Flex Credits?
Salesforce’s current Observability update says standard session tracing, telemetry ingestion, out-of-the-box dashboards, drill-downs and alerts run at no additional Data Cloud cost and do not count against Flex Credits. Net-new custom dashboards, external-data ingestion and new semantic-model objects remain metered. Agent execution is separate; confirm your agreement and actual entitlements.
Can quality scores prove an AI answer is correct?
No. Scores help prioritise reviews, but an LLM-based evaluator can also make mistakes. Calibrate scoring prompts against human-reviewed examples, inspect the evidence and confirm business outcomes in the authoritative system.
How can existing Salesforce customers get started?
Select one agent and a measurable customer journey, confirm access and available capabilities, establish a baseline, review traces and define alert ownership. Test changes in a controlled environment and compare quality, escalations, latency and cost per verified outcome before scaling.
Does observability replace security controls or human approval?
No. Monitoring does not replace least-privilege permissions, privacy controls, retention policies, knowledge governance or approval of sensitive actions. Limit access to session data, and retain human escalation for risky or uncertain situations.
Official sources and editorial scope
Reviewed October 14, 2026. Product capabilities, availability and commercial terms can change. Salesforce owns the linked demonstration. KVP commentary is independent implementation guidance, not a Salesforce warranty or a claim of realised customer results.
Agent observability