Ethical Guidance for Agentforce Sales Coach

Agent User Creation

When using personal names for AI agents, always make sure the agent’s function is explicitly mentioned to maintain transparency about its AI nature. Use a name that describes the function, job, or task rather than the agent performing it. For example, Acme’s AI Sales Coach.

Transcript-Based Evaluation

We recommend not instructing the agent to provide feedback on aspects such as tone, body language, or facial expressions, as Agentforce Sales Coach's evaluations are solely based on transcripts. The agent can accurately assess the content of the presentation, but it can’t evaluate nonverbal cues. Focusing on feedback related to the structure, clarity, and key message delivery ensures the most reliable results.

Presentation Scores

We recommend against instructing the agent to “score” a rep's talking points, as large language models (LLMs) can introduce variability and randomness in rubric-based scoring. While LLMs excel at analyzing patterns in data and providing valuable insights, they aren’t designed for consistent scoring of performance. Prompting the LLM score on performance can lead to variability in assessments. As a result, using an LLM to score a pitch or presentation can lead to inconsistent or unreliable assessments. For more accurate feedback, focus on using the agent for content evaluation, such as the clarity of messaging and the inclusion of key points.

Known AI Limitations

• Artificial hallucinations: AI can mistakenly fill in gaps or make assumptions based on patterns it has learned during training. For Agentforce Sales Coach, AI can invent information to fill in missing details or inaccurately assume context for a specific scenario.
• Misinterpretation of words: Different speech variations, homophones, and pronunciations can lead AI transcribers to misinterpret words, leading to incorrect transcriptions. The misinterpretation of words can lead to sentences that don’t make grammatical sense, which can affect the interpretation of the transcript and cause misunderstanding of the speaker’s intent.

  To address misinterpretation of words, users can review a live transcript during role-play sessions as they speak, allowing real-time verification of their transcription.

  

  Note This feature isn’t available for stand-and-deliver sessions.
• Errors: The use of flawed transcripts (such as from hallucinations or misinterpretation) by the LLM can lead to incorrect evaluations because the analysis is based on inaccurate or incomplete data.