My results after 100 sales calls: 78% summary accuracy score.

Okay, so I've been A/B testing MeetGeek against my own manual notes for the last few months. Context: I'm in a sales-adjacent marketing role, and I wanted to see if the AI summaries could actually replace my frantic note-taking during discovery calls and demos.

I set up a simple scoring system for my own reviews. After each of the last 100 sales calls, I'd:
* Let MeetGeek generate the summary and transcript.
* Compare it to my own handwritten notes and memory of the call's key points.
* Score it on a simple "Accurate" or "Not Accurate" for capturing the **core next steps, objections, and specific feature requests**.

The raw result? **78 out of 100 calls were scored "Accurate."** That's a 78% accuracy score in my book.

Here's the breakdown of what it got right and where it stumbled:

**What MeetGeek nailed consistently:**
* Capturing clear action items and next steps verbatim.
* Listing specific questions asked by the prospect.
* The general flow/topics of the conversation.
* Identifying the company name and participant roles (usually).

**Where it missed the mark (the 22%):**
* **Nuanced objections.** If a prospect said, "I'm worried about the implementation time," it got it. But if the objection was more subtle, like "We'd need to get legal to review the API terms," it sometimes buried that in the transcript without surfacing it in the summary.
* **Specific technical details.** Mention a specific integration (e.g., "the Braze to Segment sync") and it's fine. Describe a *custom* workflow without naming the tools? Often too vague.
* **Humor and sarcasm.** Obviously. This led to a few weird summary lines where a joke was taken as a serious statement.

My takeaway? It's an incredible tool for freeing up mental space and ensuring you don't drop obvious next steps. But for complex sales where reading between the lines is key, you still need to skim the transcript yourself. It's like having a 78% reliable junior sales rep taking notes.

Has anyone else done a similar accuracy test? I'm curious if others have found workarounds for capturing those nuanced objections better.

—Liam