NOTE: Demo visuals use either blurred real data or synthetic placeholders to protect customer privacy.

The AI Meeting Prep Bot still aims to achieve the same outcome: save GTM teams two to three hours per week by replacing scattered, inconsistent research with a consistent briefing. But as the system matured, one dependency became clearer. The quality of the bot’s output is bounded by the quality of the underlying CRM fields.

The specific issue was the Salesforce “Next Steps” field, a lightweight entry that should be filled after every meeting. In a scan of roughly 500 open opportunities, around 15% had an empty Next Steps value. That gap matters beyond the Prep Bot. Any agent that reasons over an opportunity state inherits the same “blank” truth.

Three frictions showed up:

1. Briefings miss what’s coming next when the field is empty, which weakens recommended actions and sequencing.

2. Reps can go into calls blind because the CRM does not reflect the last real outcome or commitment.

3. Other AI tools inherit the same gap, since many downstream agents rely on the same Salesforce fields to reason about deal state.

The bot already had fallbacks that pulled context from Gong and emails. But relying on fallbacks is not the same as having clean source-of-truth data.

The solution treated enrichment as both retrieval and writing. The first step was to integrate with Gong: if Next Steps is empty, review the most recent meeting and recover what was said but not logged. That alone is not enough, because Gong isn’t always the most up-to-date source of truth. Sometimes the real update is captured in a follow-up email. Sometimes Salesforce has already been updated elsewhere in the record, while the Next Steps field stayed stale. The system design needed to account for those realities.

So the workflow “smushes it all together” and lets the agent adjudicate:

• It pulls candidate “good” Next Steps entries that already exist and uses them as style guides.

• It checks the most recent Gong call for stated actions and commitments.

• It reviews recent emails to confirm what actually changed after the call.

• It learns the expected language and length, so the final output is concise and AE-native.

As Haren explained: “We allowed our AI agent to choose from the next steps fields that were pre-filled… look at the most recent Gong calls, look at the most recent emails… and also kind of learn the language and length.”

The output is intentionally less verbose and closer to how a strong rep would log it. It follows conventions such as date-first formatting and the acronyms AEs commonly use. The agent can also improve existing entries by adding missing specificity when the content is “good” but incomplete.

This update makes the Prep Bot more reliable without adding work for reps. It also improves the broader AI ecosystem by cleaning a shared dependency.

• For AEs: clearer next actions, fewer blind spots, and a briefing that reflects reality, not just what was logged.

• For GTM leaders: more consistent deal hygiene without chasing updates manually.

• For other AI bots: a stronger Salesforce substrate to build on, reducing “garbage in, garbage out” failure modes.

Going forward, Haren aims to expand this enrichment pattern to other high-friction Salesforce fields, prioritizing the ones that most often degrade briefing quality and push the system from “95% coverage” toward “99%.”