Agentforce KPIs after go-live: does the agent pay off?

Agentforce KPIs after go-live answer one question on day 90: does the agent resolve dealer enquiries at a lower cost than your service desk, without handing the difficult ones over in worse shape? Five numbers decide it: escalation rate, task resolution rate, cost per resolved enquiry, handle time on escalated enquiries and the share of enquiries answered outside business hours. Agent Analytics, Digital Wallet and Service Cloud supply them. The switch-off rule goes on paper before go-live, agreed with the head of service, and it applies per subagent as well as to the whole agent.

Which KPIs show whether an Agentforce agent pays off?

Salesforce’s own customer figures point at the right kind of number. The Agentforce 3 press release1 reports 15 percent less average case handle time at Engine and 70 percent of 1-800Accountant’s administrative chat engagements resolved autonomously during the 2025 tax weeks. Neither is a manufacturer, and both are vendor figures, but both measure work that no longer lands on a person.

For a machine builder’s dealer agent I use five KPIs:

KPI	What it measures	Where the number comes from
Escalation rate	Share of sessions handed to a service rep or another agent	Agent Analytics
Task resolution rate	Share of sessions judged fully resolved	Agent Analytics, Trust Layer detector
Cost per resolved enquiry	Credits plus run costs, divided by resolved sessions	Digital Wallet plus your own cost records
Handle time on escalated enquiries	Minutes a service rep spends after the handover	Service Cloud, Omni-Channel
Out-of-hours share	Resolved enquiries that arrived while the service desk was closed	Session start times against business hours

Where the numbers come from

Agent Analytics. It sits in Agentforce Studio under Observe & Optimize. The Trailhead module on agent analytics2 defines escalation rate as the “share of sessions that escalated to a support representative or another agent” and deflection rate as the share that “ended without escalation”. Abandonment counts sessions that timed out, by default after two hours. Task resolution rate is the share “evaluated as fully resolved by the Trust Layer Task Resolution detector”. Quality scores such as answer faithfulness require Audit and Feedback in Setup. Performance Insights breaks each metric down by subagent, intent or action, which is where the switch-off decision gets made.

Reading them costs no credits: Salesforce’s Observability post of July 20263 says Observability “is now unmetered for every customer across all Agentforce SKUs”.

Digital Wallet. Digital Wallet4 breaks consumption down “to the agent, action, and channel level” and supports custom tags. It requires Data 360. One agent’s credits divided by its resolved sessions give the consumption part of the cost per resolved enquiry. Knowledge upkeep, regression tests and partner days sit in your own books, and in the worked example below they are the larger part.

Service Cloud. Escalations routed through Omni-Channel become AgentWork records, a work assignment routed to an agent5, here meaning the human rep. Report handle time after the handover against the baseline from before go-live. A clean summary from the agent should lower it; dealers who arrive annoyed after three failed attempts raise it.

Out-of-hours share. Salesforce offers no ready-made metric. I count sessions that start outside the service desk’s hours, per country, and how many were resolved. For an exporter with dealers in Asia or the Americas this can be the strongest argument.

Agent Health Monitoring, listed as GA in the Dreamforce 2026 announcements6 with “16+ built-in metrics”, covers errors and latency. It shows whether the agent runs, not whether it pays off.

Why does a deflected session not count as resolved?

Deflection is the number vendors like to show, and the weakest of the set. A dealer who gets a wrong delivery date and closes the window counts as deflected. Salesforce writes in the Observability post that these rates “typically measure proxy signals” and is moving deflection and abandonment to an LLM that evaluates the full session with an editable prompt. That improves the metric.

Billing adds a third definition. For resolution-based Help Agent pricing, according to Salesforce Ben7, a messaging session with more than two turns, no feedback and no escalation request that runs past two hours is billed as resolved. That model is compared with Flex Credits in what a dealer service agent costs.

So I count an enquiry as resolved only when the task resolution detector says so, and a senior service rep reads 50 sessions a month. Wrong answers in sessions marked as resolved mean the KPI is too optimistic, and that gap goes into the report.

Worked example: a dealer agent after 90 days

Assume a machine builder with 350 dealers in 30 countries, SAP as ERP and an agent in the dealer portal with two subagents: order status and spare part availability. All figures below are assumptions to show the method; only the credit multipliers and the list price come from Salesforce.

Month three. 2,400 sessions, 1,500 on order status and 900 on spare parts. Escalation rate 30 percent (720 sessions), abandonment 8 percent, task resolution rate 52 percent (1,248 sessions). The monthly sample of 50 shows three wrong delivery dates.

Cost. Three actions at 20 credits and two Standard Prompts at 4 credits, per the rate card of 31 August 20268, make 68 credits per session and 163,200 a month. At the list price of $500 per 100,000 credits9 that is $816. Add three partner days at $1,000 for knowledge upkeep, fixes and regression tests: $3,816 a month, or $3.06 per resolved enquiry. Before go-live a service rep needed six minutes per enquiry at $60 an hour, so $6.

Handle time. The 720 escalated enquiries take 7.5 minutes each, longer than before, because only the difficult ones arrive: 90 rep hours against 240 before go-live. If half of the 192 abandoned dealers call afterwards, add 10 hours. The net saving is about 140 hours, $8,400 a month against $3,816 in agent costs.

Out of hours. 26 percent of sessions start outside 7:00 to 18:00 CET, and 58 percent of those are resolved: about 360 dealers a month who no longer wait for the next morning.

Per subagent. Order status: escalation 14 percent, task resolution 70 percent. Spare parts: escalation 57 percent, task resolution 22 percent, because parts outside the central warehouse need a manual check by the parts desk. The spare parts subagent uses $306 in credits and caused most of the fixes, so it carries half of the run costs: $1,806 for 198 resolved enquiries, or $9.12 each, more than a service rep.

$3.06

per resolved enquiry, whole agent

$6

per enquiry for a service rep before go-live

$9.12

per resolved enquiry, spare parts subagent

The decision on day 90: order status stays and goes to the next dealer group. Spare parts comes out of the agent and goes to the parts desk through a structured portal form until regional stock data reaches Salesforce. Part identification is covered in spare parts by serial number.

When should you switch an agent off?

Gartner expects10 more than 40 percent of agentic AI projects to be cancelled by the end of 2027, citing “escalating costs, unclear business value or inadequate risk controls”. A stop rule written down before go-live makes that a routine decision. I would agree these triggers with the head of service and apply them per subagent:

1. Immediately: a confirmed breach of a hard rule, such as a wrong warranty statement or another dealer’s order data. The subagent stays off until the fix passes the regression tests from testing before go-live.
2. After two improvement cycles of 30 days: escalation rate above 40 percent or task resolution below 40 percent.
3. On day 90: cost per resolved enquiry above the cost of a service rep per enquiry.
4. At any time: handle time on escalated enquiries clearly above the baseline for the same enquiry type, because then the agent shifts work instead of saving it.

The 40 percent values are my starting points, not Salesforce figures. Switching off does not have to mean ending: narrow the agent to one subagent, run it only while the service desk is closed, or fix and retest. For the last option, Agent Optimizer ranks recurring failure patterns from production sessions by impact. Salesforce’s September 2026 post11 calls it available in beta, and the Dreamforce list announces GA for October. A/B Experimentation lets a fix compete with the current version on split traffic and is announced as beta for October, so check the status in your org.

Prepare the 90-day review with three months of baseline data

1. Measure the baseline for three months: dealer enquiries by type, handle time, arrival time by country.
2. Switch on Audit and Feedback so that quality and trust scores exist from day one.
3. Tag the agent in Digital Wallet so its credits stay separate from other agents in the org.
4. Write down the thresholds and the decision owner, the head of service, not IT.
5. Book the sample reading of 50 sessions a month into a named person’s calendar.
6. Fix the review dates: day 30, day 60, decision on day 90.

The KPIs need a business case from before go-live, which is why I put business value and data quality first. Deciding which agents are worth their costs is part of my make-or-buy and licence planning work.

Sources11

1. 1Agentforce 3 press release (opens in a new tab)salesforce.com
2. 2Trailhead module on agent analytics (opens in a new tab)trailhead.salesforce.com
3. 3Observability post of July 2026 (opens in a new tab)salesforce.com
4. 4Digital Wallet (opens in a new tab)salesforce.com
5. 5work assignment routed to an agent (opens in a new tab)developer.salesforce.com
6. 6Dreamforce 2026 announcements (opens in a new tab)salesforce.com
7. 7Salesforce Ben (opens in a new tab)salesforceben.com
8. 8rate card of 31 August 2026 (opens in a new tab)salesforce.com
9. 9list price of $500 per 100,000 credits (opens in a new tab)salesforce.com
10. 10expects (opens in a new tab)gartner.com
11. 11September 2026 post (opens in a new tab)salesforce.com