MeetFollow—MeetingActionItemExtraction&Follow-UpAgent

Illustrative Project

MeetFollow is an illustrative example demonstrating an AI workflow automation pattern, not a completed client engagement.

Overview

MeetFollow processes meeting transcripts, extracts concrete action items with their owner and implied deadline, and automatically checks in with the responsible person a few days before the deadline if there's been no status update — closing a loop that almost every team intends to follow up on and rarely actually does consistently.

The Challenge

Action items mentioned in meetings are often phrased casually and indirectly — "I'll take a look at that" rather than a clearly structured commitment. The system needed to reliably identify genuine commitments versus passing comments, attribute them to the right person even when referred to by name, nickname, or role, and infer a reasonable deadline when none was stated explicitly.

Architecture & Technical Decisions

Structured Extraction With Explicit Uncertainty

The extraction prompt asks the model to identify candidate action items along with a confidence level and, critically, to explicitly note when a deadline is inferred versus stated. Items with low confidence or ambiguous ownership are flagged for the meeting organizer to confirm rather than silently assigned.

Speaker Attribution Resolution

Transcripts often refer to people inconsistently — by first name in one place, by role in another. A lightweight resolution step matches mentioned names and roles against the meeting's actual attendee list (pulled from the calendar integration) to attribute action items to the correct person reliably.

Deadline-Aware Follow-Up Scheduling

Once action items are confirmed, a scheduled job checks each item's status a few days before its deadline. If there's no recorded update, the agent drafts a friendly, context-aware follow-up message — referencing the original meeting and commitment — rather than a generic reminder, and sends it through the team's existing communication channel.

• Confidence-flagged extraction with organizer confirmation for ambiguous items
• Attendee-list-aware speaker attribution rather than relying on transcript text alone
• Context-aware follow-up generation referencing the original commitment, not generic reminders
• Status tracking that closes the loop when an item is completed, not just when it's created

Results

• 97% action item capture rate when validated against manually reviewed transcripts
• 41% improvement in on-time action item completion after automated follow-ups were introduced
• 300+ meetings processed weekly at steady-state team usage
• 93% accuracy in correctly attributing action items to the right owner

What I Learned

The hardest part of this project wasn't extracting action items from clear, explicit statements — it was handling the ambiguous, casually-phrased commitments that make up a surprising share of real meeting conversation. Building in explicit confidence flagging and organizer confirmation for ambiguous cases, rather than forcing every extraction to be fully automated, was what made the system trustworthy enough for teams to actually rely on its output.