Hey folks, I've been living in Read AI for the past few weeks, trying to get it to mesh with our analytics team's workflow, and something that keeps blowing my mind is how consistently useful its "next step" suggestions are after meetings. It's not just pulling out generic action items like "follow up on X." It gets weirdly specific. So I went down a rabbit hole trying to figure out *how* it does this, and I wanted to compare notes.
From what I can piece together (and this is just my theory based on their docs and a lot of testing), it's not one magic model. It feels like a multi-layered pipeline, kinda like an ETL process for conversation.
First, there's the obvious transcription layer. But then, I think it's doing a few key things on top of that:
* **Entity and Topic Chaining:** It's not just identifying "Project Aurora" once. It's linking every mention, seeing who spoke about it, what adjectives were used ("blocked," "on schedule," "needs review"), and across what time span in the meeting. It builds a little genealogy for each subject.
* **Sentiment & Intent Tagging:** This is the subtle part. When someone says, "Ugh, I still haven't gotten the data from finance," the model likely tags that with a frustration sentiment and, more importantly, an intent like "REQUEST_COMPLETION" or "BLOCKED_BY_EXTERNAL_TEAM." That intent is a huge clue for a next step.
* **Temporal and Ownership Cues:** It's *scarily* good at catching phrases like "I'll handle that," "Can you circle back next week?", or "We need to decide by Thursday." It parses these into a simple who/what/when tuple. Combine that with the topic chain, and you've got the raw material for a suggestion.
But here's the niche bit I geeked out on: I don't think it's just spitting out those parsed tuples. There seems to be a final step where it matches these patterns against a library of common "next step templates" for business contexts. So if it sees a pattern of [BLOCKED_BY_EXTERNAL_TEAM, DATA_REQUEST, PROJECT_AURORA], it might surface the template: "Follow up with the Finance team on the data request for Project Aurora."
I tested this by seeding a mock meeting with very obscure, internal jargon for a process. Its next steps were vaguer. But when I used common corporate language, it was razor-sharp. That suggests to me it's relying on fine-tuned models trained on a corpus of business meetings, not just a giant general-purpose LLM.
Has anyone else done similar poking around? I'm really curious if you've found ways to tweak or train it to better recognize your team's specific idioms for action items. It's a fantastic tool out-of-the-box, but if we could nudge it to recognize our weird internal project codenames as official "next step" triggers, it would be a game-changer.
—Jake
Spreadsheets > opinions