February 2, 2026
The seven customer support metrics to monitor in Agentforce are first contact resolution, escalation and hand-off rate, CSAT, first response time, task completion rate, average handling time, and containment rate. Track them together, never alone, because every one of them can look excellent while your service is getting worse.
Two things make this different from measuring a human support team. Salesforce publishes its own metric framework based on running Agentforce on Salesforce Help, and it includes measures most lists leave out. And agents cost money per action, so there is a metric here that no other support channel has: cost per resolution. Both are covered below, along with where each number actually lives in your org.
Start with the summary, then read the detail on the ones you are not tracking yet.
Every row in that third column exists for a reason. A metric read alone is a metric that can be gamed, and there is a section on exactly how further down.
What it measures: the share of issues closed in the first interaction with no follow-up, callback, or escalation.
FCR tells you whether the agent has the logic and the grounded knowledge to finish a job. Low FCR usually means one of two things: the knowledge sources are thin, or the flows are too complex and the agent loses the thread partway through.
One correction worth making to how most teams count this. Salesforce distinguishes a resolution (the conversation ended without a hand-off) from a customer-confirmed resolution (the customer said it was solved). Those are very different numbers. A conversation that ends because the customer gave up counts as a resolution under the first definition.
What it measures: how often a human takes over a conversation the agent started.
This is the clearest signal of agent maturity, and it needs one split to be useful. Salesforce separates escalations, where the agent could not deliver, from immediate customer-initiated hand-offs, where the customer asked for a person before giving the agent a chance.
Those two numbers point at different problems. A high escalation rate is a capability gap you fix with better grounding or clearer flows. A high rate of immediate hand-offs is a trust problem, and no amount of agent tuning fixes it. You fix that with better framing, better first impressions, and honesty about what the agent can do.
What it measures: how customers rate the interaction, captured right after it ends.
CSAT is the check on every efficiency metric. An agent can close a case in forty seconds and leave the customer irritated, and only CSAT catches that.
Watch the response rate alongside the score. If only a small fraction of customers respond, and they are the annoyed ones, your CSAT is measuring complaint volume rather than satisfaction. Track how many surveys come back, not just the average of the ones that do.
What it measures: how long before the agent responds after a customer submits a question.
Instant response is the reason customers accept an agent instead of a person. When this degrades, the value proposition degrades with it.
Look at the distribution, not the mean. An average of two seconds hides the Monday-morning spike where some customers waited thirty. Peak-hour performance is what your customers actually experience.
What it measures: the share of tasks the agent finishes without technical failure.
Agents follow defined logic, so failures cluster around missing data, broken integrations, and ambiguous inputs. Conversations that stop mid-process are the ones to inspect, because each one is a customer left hanging.
The trap here is treating completion as correctness. A flow that runs end to end and gives the wrong answer completes successfully. Sample real transcripts regularly rather than trusting the completion percentage on its own. Practical reporting walkthroughs on Salesforce Codex are helpful for building the underlying reports.
What it measures: the average time to resolve a query end to end.
Agents cut AHT by retrieving context instantly instead of searching multiple systems. That is a real gain, and it is also the metric most likely to be misread.
Falling AHT with falling CSAT is not efficiency. It is the agent closing conversations before the customer is done. Always read the two together.
What it measures: the share of contacts handled entirely by the agent with no human involvement.
This is the number that appears in ROI decks, and it deserves a precision note. Deflection and containment get used interchangeably and are not the same thing. Deflection often counts a customer who left without asking anything, which is a failure dressed as a win. Containment counts conversations the agent genuinely finished.
Measure containment. Then check it against repeat contact rate, because a contained conversation followed by the same customer calling tomorrow was never contained at all.
Salesforce runs Agentforce on its own Help site and publishes the framework it uses, split into adoption and effectiveness. Three of those measures rarely appear in third-party lists, and each closes a real blind spot.
Abandons is the one to add first. A customer who opens the agent and closes it without typing anything tells you the entry experience is failing, and no resolution metric will ever surface that.
Adoption as a denominator matters just as much. Containment of 80% means one thing across 10,000 conversations and nothing at all across 40. Sessions and conversations give your effectiveness numbers a scale to sit against.
Human support costs are largely fixed and per-hour. Agent costs are variable and per-action. That is a genuinely new thing to measure, and it belongs on your dashboard.
Salesforce prices Flex Credits at $500 per 100,000 credits, with a standard agent action consuming 20 credits, roughly $0.10, and voice actions consuming more. Pricing has changed more than once, so confirm the current rate card before you build a model on it.
The arithmetic is simple once you know actions per conversation.
Two consequences fall out of this. Escalations become more expensive than they look, because you pay the agent cost and then the human cost for the same contact. And a low resolution rate raises cost per resolution even when cost per conversation looks fine.
We walk through the same per-action arithmetic in a regulated context in our guide to Agentforce for insurance claims management, where the volume of status inquiries makes the math especially stark.
Listing metrics is easy. Knowing where they live is what makes them trackable.
That last row is the one teams skip. No automated metric tells you whether the agent was right, so build a weekly sample of real transcripts into someone's job. Grounding quality is usually the root cause when accuracy slips, which is why a properly modeled Data Cloud foundation does more for these numbers than any prompt tuning.
Building the reports themselves is ordinary Salesforce administration work, and it is worth doing properly once rather than rebuilding a dashboard every quarter. Community tutorials on Salesforce Geek cover the reporting mechanics well.
Not maliciously, usually. It happens when a number becomes a target and someone optimizes for it.
The pattern is consistent. Every efficiency metric can be improved by making the customer experience worse, which is why each one needs a quality metric beside it.
Chasing published benchmarks is a mistake. Containment on billing questions and containment on technical troubleshooting are not comparable numbers, and neither is comparable to someone else's org.
Capture your own baseline before the agent goes live: contact volume by reason, average handle time, first contact resolution, CSAT, and fully loaded cost per contact. That last one is what makes any savings claim defensible.
Then review on a rhythm. Weekly for the first month while you are still fixing flows, monthly for the next six, quarterly after that. Setting these targets during the build rather than after go-live is part of a disciplined Salesforce implementation, not a reporting afterthought.
Segment by intent when you review. One aggregate containment number hides the fact that the agent is excellent at password resets and hopeless at billing disputes, and only the segmented view tells you what to fix next. Architecture-level write-ups on Jitendra Zaa and practical tips on SFDCStop are useful when you get into building the segmented reporting layer.
Customer-confirmed resolution, escalation rate split by cause, CSAT, and cost per resolution. Those four together tell you whether the agent works, whether customers accept it, and whether it pays for itself. The rest are diagnostics for when one of the four moves.
There is no universal number, because it depends entirely on what you ask the agent to handle. Password resets and status lookups contain at a much higher rate than billing disputes. Baseline your own org by intent type rather than chasing a published figure.
Deflection often counts anyone who did not reach a human, including customers who gave up without asking a question. Containment counts conversations the agent genuinely completed. Containment is the honest metric; deflection flatters your dashboard.
Multiply actions per conversation by cost per action, then divide by your resolution rate to get cost per resolution. Salesforce prices Flex Credits at $500 per 100,000 credits, with a standard action at about 20 credits or $0.10. Confirm current pricing before modeling.
Almost always because conversations are closing before customers are finished. Add a customer-confirmation step before the agent marks anything resolved, and check your repeat contact rate for the same customers.
Weekly for the first month, monthly for the next six, then quarterly. Early on you are still fixing flows and grounding, so the feedback loop needs to be tight enough to act on.
Most Agentforce dashboards report what the agent did. The useful ones report what the customer got. That difference is the gap between a containment number you present to a board and a service operation that genuinely works.
Track the seven, add abandons and confirmed resolutions, put a cost per resolution next to them, and pair every efficiency metric with a quality one. Then sample real transcripts, because no dashboard will tell you the agent was wrong.
At Minuscule Technologies we build the reporting layer alongside the agents, grounded in properly modeled data, so the numbers mean something when someone asks. Talk to our Agentforce team about instrumenting your agents before you scale them.
You've seen what's possible. Now, let's make it happen for your business. Whether you need an end-to-end Salesforce solution, a complex integration, or ongoing managed services, our team is ready to deliver.
Schedule a Free Strategic Call
