Salesforce Agentforce service triage (L4)

The steps

1. 01

   Sanitize the Case Data Model

   Before touching Agentforce, you must normalize your historical Case data so the model has a clean 'source of truth' to learn from. In Salesforce Setup, go to the Object Manager and select 'Case.' Audit your 'Type,' 'Reason,' and 'Status' picklist values. If you have 50+ reasons or overlapping categories (e.g., 'Billing Issue' and 'Payment Problem'), consolidate them. Agentforce relies on distinct classification to route correctly. • Create a custom checkbox field on the Case object called 'Is_AI_Ready' to flag high-quality historical cases for training. • Use Data Loader or an Apex script to clean up trailing spaces or 'null' values in subject lines. • Owner: RevOps or Salesforce Admin. • Time Estimate: 4-6 hours. • Pitfall: Skipping this leads to 'hallucinated' classifications where the AI guesses between two similar categories. • Definition of Done: A picklist of <15 distinct 'Reason' values and a report showing 95% of cases from the last 30 days have these fields populated.

2. 02

   Configure the Agentforce Foundation

   Navigate to 'Einstein Setup' in Salesforce and ensure 'Einstein Classification' and 'Service Cloud Einstein' are toggled on. Next, go to 'Agentforce Service Agent' configuration (formerly Einstein Copilot). You need to define the 'Service Agent's scope. • Under 'Einstein Case Classification,' create a new model. Select 'Case' as the object and choose your newly cleaned 'Reason' and 'Type' fields as the ones to predict. • Set the 'Confidence Threshold' to a conservative 80% initially. Anything below this should stay in the standard queue rather than being auto-triaged. • Example Config: Choose 'Automate Field Updates' so the AI doesn't just suggest the value, but actually fills it in. • Owner: Salesforce Admin. • Time Estimate: 2 hours. • Prerequisites: Einstein permissions assigned to your user profile. • Pitfall: Setting the confidence threshold too low (e.g., 50%) results in messy routing that frustrates the support team. • Definition of Done: The classification model is built and 'Active' status is displayed in Einstein Setup.

3. 03

   Build the Auto-Drafting Logic

   Within the Agentforce builder, you must create 'Actions' that the agent can perform, such as drafting a response. Go to 'Einstein Service Replies' settings. Enable 'Service Replies' and link it to your 'Knowledge' base. • If you don't have a Knowledge Base, you must create at least 10-20 'Quick Text' or Knowledge articles for the AI to reference. • Configure the 'Drafting' prompt: Use the 'Template' builder to instruct the agent on tone. Example Prompt: 'You are a helpful support agent. Use the context from our Knowledge articles to draft a response. Do not promise refunds without a manager's approval. Stay concise.' • Map the 'Draft' output to a custom field on the Case called 'AI_Proposed_Response__c' so human agents can review it before it goes live. • Owner: CS Manager. • Time Estimate: 3 hours. • Pitfall: Giving the AI 'write' access to the customer email field directly too early. • Definition of Done: An incoming case automatically generates a draft in the hidden internal field.

4. 04

   Pilot via Targeted Routing Flows

   Instead of applying AI to every support ticket, create a specific 'Pilot Queue' in Salesforce. Go to 'Queues' in Setup and create 'AI_Pilot_Queue.' • Update your Lead Assignment Rules or Case Assignment Rules (Setup > Case Assignment Rules) to route only specific ticket types to this queue (e.g., 'Account Access' or 'Billing Questions'). • Use 'Flow Builder' to trigger the Agentforce Service Agent only when a Case enters this specific queue. Formula: '{!$Record.Status} == "New" && {!$Record.Queue.Name} == "AI_Pilot_Queue"'. • This limits your risk and allows you to compare AI performance against a 'Control Group' of human-only queues. • Owner: Salesforce Admin. • Time Estimate: 2 hours. • Pitfall: Choosing the most complex or 'angry' customer tickets for the pilot. • Definition of Done: Only cases meeting specific criteria are processed by Agentforce, while others follow the standard flow.

5. 05

   Deploy Dashboard for AI Observability

   To measure success, you need a side-by-side comparison. Create a new Salesforce Dashboard called 'Agentforce vs. Human Performance.' • Create a Report on Cases, grouped by 'Owner' (AI Queue vs. Human Queues). • Add three specific components: 1. Average Time to First Draft, 2. Case Deflection Rate (if the AI closed it without a human), and 3. CSAT Score. • For the CSAT, use Salesforce Surveys to send an automated email post-close. Ensure the survey link includes a hidden field identifying if Agentforce handled the case. • SQL/Query Logic: 'SELECT AVG(Case_Resolution_Time__c) FROM Case GROUP BY Escalated_By_AI__c'. • Owner: RevOps/Support Lead. • Time Estimate: 3 hours. • Pitfall: Forgetting to track 'AI Rejects' (how many times a human had to fix a wrong AI classification). • Definition of Done: A live dashboard showing if AI is actually faster and if customers are as happy as they are with human agents.

Next playbooks

• CSM call summaries via Gainsight + AI (L2)
• AI-generated QBR decks (L3)
• Intercom Fin AI for support deflection (L3)
• Support deflection with AI search (L3)