Agentforce Voice Best Practices
Voice-enabled agents require distinct conversational design patterns compared to text-based agents to handle spoken interaction effectively.
Required Editions
| Available in: Lightning Experience |
| Available in: Enterprise, Unlimited, and Developer Editions with Foundations or Agentforce 1 Editions, and Salesforce Voice add-ons. |
Here are some best practices that have emerged through our testing.
Conversational Design and Interaction
Conversation Flow
Apply these practices to the instructions provided to your subagents.
- Keep your welcome messages short to account for the fact that callers can't interrupt or skip greetings. Avoid context-heavy or dynamic greetings that repeat information the caller provided.
- Use closed-form prompts to elicit short, specific answers. For example, say "Are you familiar with the product?" rather than open-ended questions. This type of specificity in the instructions prevents the agent from interpreting mid-sentence pauses as the end of a turn.
- Manage latency with verbal cues such as, "Let me look that up" to set expectations. and fill the gap.
- Confirm transfers by including a definitive question from the agent. For example, you can instruct the agent to say "I’m going to transfer you now. Is that OK?", before handing it to a human rep.
- Instruct agent to invite callers to disconnect if they have no more questions, as agents typically end calls only after a period of silence.
Speech Optimizations
- Strategic repeat-back pattern: Use repeat-back instructions only when accuracy is critical or the input is complex (for example, flight details) to avoid being overly verbose and slow.
- Formatting Instructions: Specify clear instructions to ensure natural pronunciation.
Here are a few example instructions.
- Times: Use 12-hour formats and avoid "O'clock" (for example, "8 a.m.").
- Currency: Instruct the agent to omit ".00" for natural speech (for example, "Ten dollars").
- Case numbers: Read the case numbers as individual digits with pauses (for example, "one - two - three") rather than whole numbers.
- URLs: Strip "https://" from any URLs, as the engine reads the protocol literally.
- Pronunciation: Use the Pronunciation Dictionary setting to help agents pronounce certain unique terms. The pronunciation dictionary supports the International Phonetic Alphabet (IPA) and Carnegie Mellon University (CMU) formats. Preview the word as pronounced by the agent first to check if it's required as a dictionary entry.
- Transcribing: Use key-term prompting to make sure certain unique terms specific to your
organization are difficult to translate to text correctly. However, use keyword prompting
or boosting only when words or terms are misinterpreted or transcribed incorrectly by the
agent. Use the key-term boosting or promoting setting only when necessary as it can
interfere with the interpretation of similar sounding words.
These examples demonstrate how key-term prompting can significantly improve recognition accuracy and confidence scores for industry-specific terminology.
Note The confidence scores later are illustrative examples showing typical improvement patterns. Actual results can vary based on audio quality, accent, and context.Source Confidence Score - Before Confidence Score- After account number "word": "a count", "confidence": 0.82154321 "word": "account", "confidence": 0.97891234 representative "word": "represent a tip", "confidence": 0.79043218 "word": "representative", "confidence": 0.98654321 billing department "word": "building", "confidence": 0.81234567 "word": "billing", "confidence": 0.96789012 escalation "word": "escalate shin", "confidence": 0.76543210 "word": "escalation", "confidence": 0.98123456 customer service "word": "customer", "confidence": 0.84567890 "word": "customer", "confidence": 0.97234567 technical support "word": "tech nil call", "confidence": 0.83456789 "word": "technical", "confidence": 0.98345678
Best Practices for Setting Up Voice with Telephony
The choice of telephony connectivity is one of the most significant technical decisions for an Agentforce Voice deployment, affecting latency and metadata capabilities.
Connectivity Strategy
- SIP (recommended): SIP is the preferred path as it supports custom UUI (User-to-User Information) headers. These headers carry hex-encoded JSON payloads to pass routing data and caller context natively.
- PSTN Limitations: While common, PSTN lacks a native metadata passthrough and currently supports only US and Canada numbers. International or cross-region flows must use SIP to keep requests within the same region and minimize latency.
Technical implementation:
- Call Record Linking: For PSTN, turn on Context Linking Preference in setup. The Voice Call must be created in the IVR before the transfer and completed within 30 seconds.
- Call Recording: Unlike PSTN, SIP recording isn’t automatic in Salesforce. It must be explicitly configured with your telephony partner, and those recordings reside in the partner's system.
- SIP configuration: Create a SIP channel in Salesforce and turn on and configure a SIP trunk within the partner system to connect to that channel.
Best Practices for Planning and Implementing a Voice-Enabled Agent
Successful deployments require early alignment across technical, data, and legal workstreams.
Scope and data preparation:
- Prioritize Low-Complexity Use Cases: Start with routine, high-volume calls like FAQs or status checks. Avoid high-acuity scenarios like clinical triage during the initial phases.
- Audit CRM and Knowledge Data: Agent performance depends on the quality of underlying data. Resolve duplicate records and make sure that knowledge articles are well-structured and current. To improve response speed for knowledge retrieval, use faster output generation models in your prompt template. For voice use cases, Gemini Flash models are typically faster.
- Verify capacity: Ensure your expected peak volumes fit within the default limit of 500 concurrent voice calls per org.
Security, Compliance, and Escalation
- Legal Engagement: Define a voice recording retention policy immediately. If full retention is restricted, negotiate a short window (for example, 24 hours) for troubleshooting.
- Redaction and masking: Use Sensitive Data Rules to mask PII or PHI in transcripts. For PHI, verify that your Salesforce Business Associate Agreement (BAA) covers the entire voice pipeline and turn on Shield Platform Encryption. BAA is a legal contract under HIPPA.
- Design Fallbacks: Always configure a fallback queue or voicemail to make sure that callers never reach a dead-end. Handle error scenarios to prevent accidental call disconnection.
Did this article solve your issue?
Let us know so we can improve!
