Skip to main content
Availability2026.3 or later

What GermainUX Watches

Turns, errors, ratings and AI findings

Find the Cause

Four worked examples

Alert and Act

Catch it next time

Set It Up

Monitoring, AI and markers

Overview

You rolled out Agentforce to your sales and service teams, and now the questions start. Why did the service team stop trusting it this week? Why does one team barely use it? Is it Salesforce, the agent, or the way people ask? CRM owners, IT operations and AI teams each see part of the answer. Salesforce counts the Agentforce requests that ran. That shows how much Agentforce was used, not why an answer failed, why an employee rejected it, or why a team stopped asking. The cause can be an agent update that broke an action, a Salesforce limit, or questions the agent was never set up to answer. Each needs a different team and a different fix. GermainUX answers the question from what each employee experienced. It records every Agentforce question and answer in the employee’s browser, with its response time, error type, rating and the AI’s reading of the conversation. It compares a period with its baseline, ranks the error types, frictions and employees that moved with it, lets you narrow it down to one agent or team, and marks agent releases on the same timeline. Then it shows you the conversation and the session around it, so you see what the employee saw.
Errors or Dislikes RiseAggregate AnalysisAgent, Error Type or TeamConversationSession Replay
This page uses Agentforce in Salesforce Lightning as the example. The same approach works for any AI assistant or chatbot whose conversations GermainUX monitors, see Chat Analysis.

Questions It Answers

  • Why did Agentforce start failing today?
  • Is it Salesforce, a request limit, or one of our agents?
  • Which agent release made the answers worse, and did the fix work?
  • Which teams don’t use Agentforce, and why did they stop?
  • What did employees ask, and what did Agentforce answer?
  • What happened with the request to escalate case 00048213?

What GermainUX Watches

GermainUX collects the signals that explain an Agentforce problem in one place, so the analysis can compare them with each other. All of them come from Agentforce Monitoring unless the row says otherwise.

Find the Cause

The investigation follows the same path every time, whatever the cause turns out to be:
  1. Open Aggregate Analysis on the Agentforce errors, AI frictions or conversations for the period. VS RECENT compares the period with the same hours on the previous days.
  2. Read the Leading Factors. The error type, the friction and the employees that rank show where the problem is concentrated. Filter on the agent (Responder) or on a team to narrow it down.
  3. Check Relationships for the hours when it started and the related KPIs that moved with it. Add your agent releases as markers to see which release it follows.
  4. Rule out the signals that stayed normal. They tell the team what not to fix.
  5. Open one affected conversation and its session in Session Replay as evidence.
The four examples below come from simulated data for 60 employees in three teams, using two agents: a Service Agent for the service team and a Sales Agent for the two sales teams. Their numbers belong to the example. Each one ends at a different kind of answer.

Example: An Agent Update Can’t Find Cases

On Tuesday morning, the Salesforce team activates a new version of the Service Agent with rewritten case-management instructions. A few minutes later, the agent stops finding the cases employees ask about.
1

Spot the failures

Open the analysis of the Agentforce errors for Tuesday morning. In this example, 22 turns fail against 5 in the same hours of the previous days, 340% more, and the errors start within the hour after the release.
Analysis of Agentforce errors on Tuesday morning: 22 errors against 5 on the previous days, with Agent Failure ranking in Leading Factors

Agentforce Errors After an Agent Release - GermainUX

2

Read what ranks

Leading Factors rank the error type Agent Failure for all 22 errors. The request reached Agentforce and Salesforce answered, but the agent had no answer.
3

Compare the agents

Compare the turns of the two agents for the same hours, with a filter on the agent (Responder) on each side. Top Differences show what sets them apart: in this example, 79% of the Service Agent’s turns end with an answer (Inform), against 95% for the Sales Agent. Response times are about the same, so the agent isn’t slow. It answers less. Only one agent’s answers changed.
Comparison of Agentforce turns on Tuesday morning: Top Differences show 79.2% of Service Agent turns answered (Inform) against 95.3% for the Sales Agent

Service Agent Turns Compared with Sales Agent Turns - GermainUX

4

Read the frictions

Open the AI frictions for the same hours: 46 against 19. Record Not Found (20) and Request Repeated or Rephrased (15) lead. Employees ask about the case on their screen, the agent answers that it can’t find it, and they ask again.
Analysis of AI Chat Friction events for Agentforce on Tuesday morning: 46 frictions against 19, with Record Not Found and Request Repeated or Rephrased ranking in Leading Factors

AI Frictions in Agentforce Conversations - GermainUX

5

Rule out the rest

There are no Service Error or Rate Limited errors in these hours, and the Salesforce instance status stays OK. Salesforce works.
6

Watch one conversation

Employee Jordan Lee ranks among the users with the most frictions. Open their conversation: they ask Agentforce to escalate case 00048213, get “I couldn’t find a case with that number”, rephrase, get “Sorry, I couldn’t find that record”, and dislike the answer. Their session continues with Click on Edit and Click on Save on the same case: they escalated it by hand. Session Replay shows what they saw.

Example: Agentforce Hits a Request Limit

On the same afternoon, Agentforce stops answering many employees. In this example, 36 turns fail with Rate Limited.
1

Spot the spike

Open the analysis of the Agentforce errors for Tuesday afternoon. In this example, 48 turns fail against 4 in the same hours of the previous days, most of them in one spike.
2

Put the failed calls on the chart

Under Related KPIs, filter the Outbound HTTP Request card to the calls that broke their SLA (Color is RED) on the Agent API host (api.salesforce.com). Pivot it by HTTP Status and select the 429 row to add it to the chart. In this example, all 36 failed calls are 429, and their line sits right on the error spike.
Analysis of Agentforce errors on Tuesday afternoon with the Agent API calls that returned HTTP 429 drawn on the Relationships chart over the error spike, and the Outbound HTTP Request card pivoted by HTTP Status showing 36 calls with status 429

Agentforce Errors with the Rejected Agent API Calls - GermainUX

3

Read what ranks

Leading Factors split the errors by type: Rate Limited (36) is new, next to the Agent Failure (12) from the previous example. Salesforce rejected the requests with HTTP 429 because the org reached its Agentforce request limit.
4

Rule out the rest

The Salesforce instance status stays OK, and turns that get through are answered as fast as usual. Salesforce isn’t down. The org used up its Agentforce capacity.
Ask the Salesforce admin what else sent Agentforce requests in that window, such as a flow or an integration that calls agents in bulk, and schedule it outside working hours or raise the limit. Keep an alert on Rate Limited errors so the next one is caught in minutes.

Example: A Team Stops Using Agentforce

Nothing is broken, but fewer reps of the Field Sales team use Agentforce every day.
1

Spot the team

Open the analysis of the Agentforce conversations, filtered on the team. In this example, Field Sales has 11 conversations from 4 reps on Wednesday, against 15 conversations from 7 reps on the previous days.
Analysis of Field Sales Agentforce conversations on Wednesday: 11 conversations from 4 users against 15 from 7 users, 72.7% resolved against 50.5%, with the top conversation topics

Agentforce Conversations of One Team - GermainUX

2

Read who stayed

The reps who still use Agentforce get answers: 73% of their conversations are resolved, against 51% on the previous days, and fewer have a friction (27% against 47%). Their topics are spread across summaries, opportunities, reports and emails, and they type their own questions.
3

Find out how the others asked

Open the earlier conversations of the reps who stopped. Most start with a welcome suggestion, such as “Summarize this record” on an Account page. Their AI outcome is Partially Resolved, with the friction Generic or Incomplete Answer: Agentforce returned a generic description of the account.
4

Rule out the rest

Response times and errors for Field Sales are the same as for the other teams. The agent works, but it doesn’t help with what field reps need before a customer meeting.
This is a business cause, not a technical one. Give the Sales Agent a topic for account and meeting preparation, and show field reps the questions their colleagues get good answers to. Then compare the team’s conversations and reps the following week.

Example: The Fixed Agent Brings Answers Back

On Wednesday morning, the Salesforce team activates a fixed version of the Service Agent. The same analysis that found the failures shows whether the fix worked.
1

Spot the drop

Open the analysis of the Agentforce errors for Wednesday, filtered on the Service Agent. In this example, 3 turns fail against 27 in the same hours of the previous days, 89% fewer. The previous days include Tuesday’s failures, so also compare with a day before the update.
Analysis of Service Agent errors on Wednesday: 3 errors against 27 on the previous days, with Connection Lost in Leading Factors

Service Agent Errors After the Fix - GermainUX

2

Read what's left

Leading Factors show Connection Lost for 2 of the 3: the employee left the page before the answer arrived. These are everyday errors, not the case failures of Tuesday, so the agent finds the cases again.
3

Watch one conversation

Open a case update from Wednesday: the employee asks Agentforce to change the case status, and Agentforce does it in one turn.
Share the result with the team that changed the agent, and keep the alert on Agentforce errors so a regression shows up quickly.

Alert and Act

Each example points to a signal you can watch from now on:
  • SLAs and outlier detection on Agentforce errors and Bot Response Time, per agent, so the next failure raises an alert
  • Alerts on Rate Limited errors, so you hear about a request limit before employees complain
  • Release markers for every agent version you activate, so a change in answers lines up with the release that caused it
  • Scheduled reports on conversations, outcomes and frictions by team, for the CRM owner and the team leads

Set It Up

The analysis needs the Agentforce data, the AI analysis of conversations, and markers for your agent releases.
1

Monitor Salesforce Lightning

Deploy Salesforce monitoring for your Lightning users. Agentforce turns, errors, ratings and panel use come with the Salesforce UX Monitoring Profile, see Turn On Agentforce Monitoring.
2

Turn on AI conversation analysis

Turn on the chat processor and the prompt for Agentforce conversations, so GermainUX records each conversation’s outcome, topic, summary and frictions. See Turn On AI Conversation Analysis.
3

Add agent release markers

Record each agent version you activate as a release, and configure a KPI relationship from Chat Event and Chat Transaction to it, shown as a marker.
4

Set SLAs and alerts

Set SLAs on Agentforce errors and response time, and alerts on the error types that matter most to you.

Additional Resources

Agentforce Monitoring

Turns, errors and AI analysis

Analysis Dashboard

Aggregate and instance analysis

Use Cases

More Use Cases

Get Help

Every org has its own agents, topics and teams. The GermainUX team can help you set up Agentforce monitoring, the AI conversation analysis, the release markers and the alerts that fit your rollout.

Contact GermainUX Support

The GermainUX team helps you get up and running faster, empower your team, turn data into actionable insights, and optimize performance to deliver measurable business outcomes.

Can't Find What You Need?

Search all GermainUX docs with Ctrl + K, or use the feedback buttons at the bottom of this page to tell us what is missing.

The content of this document is proprietary to GERMAIN SOFTWARE.