Autopilot populates and updates Salesforce fields automatically using AI extraction from call transcripts and emails. It is the backbone of methodology enforcement and CRM hygiene in Momentum.
Three variants — choose based on trigger
Field type rules — this is where most errors happen
Autopilot auto-adapts output based on the Salesforce field type. Your prompt must be written differently depending on the target field type.
Picklist fields
Picklist rule: exact match only. Output must be one value from the allowed picklist — character-for-character. No explanation. No surrounding text. If the correct value cannot be determined: output nothing (never guess). A wrong picklist value is worse than an empty field — it corrupts CRM data.
- Define each value, don’t just list it. The AI needs to know what each option means, not just what it’s called. A label like “Partner’s End User” is ambiguous without a definition.
- Add few-shot examples. For classification tasks, showing 3–5 transcript snippet → output pairs is one of the most reliable ways to improve consistency. The AI learns the pattern from the examples rather than inferring it from your rules alone.
- Specify whose statements count. Rep statements, hypotheticals, future plans, and guesses don’t qualify. Only explicit statements from the customer or prospect should drive classification.
- Write tiebreaker rules. When two values could apply, tell the AI which wins. When things are ambiguous or contradictory, always resolve to your null value (None, Unknown, etc.) rather than guessing.
Textarea fields
Textarea rules: plain text, no preamble. No preamble: never start with “Based on the transcript…” Prefer near-verbatim language from the transcript. Always define the absent-data behavior — see the Overview for silence rule patterns including how to prevent trailing statements. Field type quick reference
Autopilot examples
The examples below show both prompt styles where it applies. For Textarea fields, a plain-language prompt often works well to start. For Picklist fields, the structured format is strongly recommended — value definitions, source scoping, and tiebreaker rules each directly improve classification accuracy in ways plain language can’t reliably replicate.
Example 1 — Next Steps (Textarea, Classic)
Natural language equivalent — same task, simpler prompt. For a straightforward extraction like next steps, a plain-language version often gets you 80% of the way there and is a good starting point before adding rules:“Extract any next steps agreed on this call, including who owns them and when. Format as a bullet list: [Owner]: [Action] — [Timeline]. If nothing was agreed, return nothing.”Add structure when you see the AI including vague commitments, rep-stated actions the customer didn’t confirm, or adding trailing statements when nothing was found.
Example 2 — Site Management Model (Picklist, Classic)
Example 3 — Closed Won Story (Textarea, Retropilot)
Triggered on Opportunity stage change to Closed Won. Analyzes the full conversation history within the deal window and writes a CRM-ready win story to a dedicated Salesforce field.
Example 4 — Closed Lost Reason (Picklist, Batch)
Run manually across a set of closed-lost records to backfill a reason classification field — a classic Batch use case for retroactive data cleanup or methodology enforcement.
Autopilot — QA checklist
- Correct variant: Classic for per-call, Retropilot for event-triggered history, Batch for bulk or one-off
- Picklist: allowed values listed, output = exact value only
- Multipicklist: semicolon separator specified in output format
- Textarea / Rich Text: no preamble, no markdown symbols (they render as literal characters in Rich Text fields), silence behavior defined
- Silence rule explicitly stated
- No over-extraction (inferring content not in the transcript)
- Save behavior matches current confidence level
