Product thinking / Case study

AI Meeting Insights Generator

Company: AECC Global | Role: PM - AI & Growth | Year: 2024

An AI-powered tool that automatically generates structured meeting summaries from counsellor-student sessions, personalised to each student and linked directly to follow-up actions in the CRM.


The Problem

After a counsellor meeting, two critical things needed to happen:

  1. The counsellor had to update the CRM with meeting notes
  2. The student needed a personalised follow-up with their shortlist and next steps Reality:
  • Counsellors spent 15–20 minutes post-meeting updating CRM notes - inconsistently and often incompletely
  • Follow-up emails were templated and generic (not personalised to what was discussed)
  • Students frequently said they “couldn’t remember” the course shortlist from the meeting
  • Counsellors had no structured record of student concerns, objections, or stated preferences - critical for re-engagement calls

Discovery

What I explored:

  • Observed 12 counsellor meetings end-to-end (including post-meeting CRM update)
  • Interviewed 15 students 48 hours after their meetings: what did they remember? what did they wish they had?
  • Ran a counsellor survey (n=38) on time spent on administrative tasks

Key findings:

  • 18 minutes average post-meeting admin time per counsellor per meeting
  • Only 43% of CRM notes had structured data (course shortlist, student concerns, next steps) - the rest was free text with varying quality
  • Students who got personalised written summaries converted at 2.1× the rate of those who got generic follow-ups (found in a retrospective data analysis)
  • Counsellors wanted voice-to-notes capability - most were more comfortable speaking than typing

Solution Design

Architecture

Meeting ends (video or in-person)
    ↓
Audio transcription (ASR layer)
    ↓
LLM processing:
  - Extract: course shortlist discussed, student concerns, decision stage
  - Generate: structured CRM note (fields: courses, objections, next steps, follow-up date)
  - Generate: personalised student summary email
    ↓
Counsellor review screen (30-second edit + approve)
    ↓
Auto-push to CRM + auto-send follow-up email

Key Design Decisions

Decision 1: Counsellor review step is non-negotiable Rationale: LLMs make errors. An incorrect meeting summary sent to a student would destroy trust. We built a mandatory 30-second review step. In practice, 78% of summaries required zero or one edit. Decision 2: Structured output schema (not free-form notes) Rationale: The value of this tool downstream (re-engagement, analytics, team handover) depended on structured, consistent data. I defined a strict JSON schema for the LLM output and validated against it. Decision 3: Start with video meetings only (v1) Rationale: In-person meeting transcription required hardware or phone use that created adoption friction. Video meetings had clean audio and a defined start/end. Scope discipline was critical here.

Results

  • Post-meeting admin time reduced by ~70% (18 min → 5 min including review)
  • CRM structured data completeness: 43% → 91%
  • Student follow-up personalisation rate: ~20% → 100% (all meeting summaries personalised)
  • Students who received AI summaries converted at 1.9× the rate of those who didn’t (confirmed in A/B across 2 offices)

What I Learned

  • The “human in the loop” step must be fast - if the review step takes >60 seconds, counsellors will skip it. We designed for 30 seconds.
  • Prompt schema design = product design - the quality of output was entirely determined by how precisely I defined the extraction schema and output format
  • Don’t underestimate the re-engagement value - the structured CRM notes became as valuable as the follow-up emails. Re-engagement calls improved significantly because counsellors had rich context on student concerns.
Explore more work