Read enough AI meeting summaries and they start to blur into one. “The team discussed the Q3 roadmap. Concerns were raised about timeline. Next steps were agreed.” Three sentences that describe the category of what happened without telling you a single thing you can act on. You were in the room. You already knew the team discussed the roadmap. What you needed was the part the summary quietly dropped.
Around the fire, nobody ever recapped the night as a bulleted list. People remembered who said what, who promised what, and whose story didn't quite add up. That's the shape memory actually takes — and it's the shape almost every modern meeting tool gets wrong.
The summary compresses the wrong axis.
A meeting is high-dimensional. It has speakers, commitments, conditions, disagreements, decisions, open questions, and changes-of-mind, all unfolding over time. A bullet-list summary flattens every one of those onto a single axis: topics discussed. It's a faithful answer to the question “what was this meeting about?” and a useless answer to every question you actually ask a week later.
The questions you actually ask are these. Who agreed to do the thing? By when? Did they hedge? Did the number we landed on match the number in the contract? Did she say something today that contradicts what she said last month? None of those survive the compression to bullets, because the summary was built to be readable, not to be load-bearing.
That is the structural mistake at the center of the category. The output looks like a deliverable, so it feels like the product is done. But the readable artifact and the useful artifact are different shapes, and the industry shipped the readable one.
What gets thrown away.
Watch what a bullet-list summary deletes, every single time:
The speaker. “It was agreed the proposal would go out Friday.” Agreed by whom? Who's on the hook? A commitment with no owner is a commitment no one keeps.
The condition. Half of all real commitments are conditional — “I'll send it once Mike signs off on pricing.” The summary records “will send proposal.” The condition was the whole point.
The exact words. “Roughly $40K” and “no more than $40K” and “around forty, give or take” are three different agreements. The summary picks one paraphrase and erases the other two. Three weeks later, in a contract dispute, the paraphrase is worthless.
The contradiction. Today she said the integration was “basically done.” Last month she said it “hadn't started.” A per-meeting summary can't see across meetings, so it can't see the gap — and the gap is often the most valuable thing in the room.
The pattern is consistent: a summary keeps the gist and discards the attribution, the qualifier, the source, and the cross-reference. Those four are exactly what you need when something goes wrong, which is the only time you go back to a recording at all.
A different output for a different job.
Bonfiyah starts from a different question. Not “what was this meeting about?” but “what do you need from this meeting three weeks from now?” — and the answer is almost never a paragraph.
Promise Tracker produces a structured list of commitments instead of a summary of topics. Every promise, yours and theirs, attributed to the speaker, with the deadline if there was one and the condition if there was one, anchored to the exact second in the audio where it was said. It tracks each one to closure. When Friday comes and the proposal hasn't, it surfaces the original quote — her words, her timestamp — not a paraphrase that's already drifted.
Truth Layer reasons across recordings, which is the move a per-meeting summary structurally cannot make. It's the feature that catches “basically done” today against “hadn't started” last month, and shows you both quotes side by side. Not to win an argument — to make sure the conversation is happening on the same set of facts. The attribution under both of these comes from cross-recording voice ID, so the same person is one identity from one meeting to the next.
Neither of these is a paragraph you skim and forget. They're objects you act on.
“But the summary is fine for most meetings.”
Sometimes it is. If the only thing you ever need from a recording is a vague reminder of the topic, a bullet list is enough, and Bonfiyah will happily give you one too. The point isn't that summaries are worthless. It's that they're positioned as the output when they're really the lowest-value layer — the thing that's easy to generate, easy to read, and quietly missing everything that costs you money when it's gone.
The tools that lead with the summary are built for the demo, where a clean bullet list looks impressive in fifteen seconds. The work doesn't happen in fifteen seconds. It happens three weeks later, when the deadline slips and you go looking for what was actually promised, and the summary has nothing to say.
Where this leaves you.
If your meeting tool hands you a tidy summary you've never once gone back and used, that's not a discipline problem on your end. It's the wrong shape. The information you needed was extracted, flattened, and thrown away before it ever reached you.
Install Bonfiyah, record three meetings, and let the Pro AI layer run on its free 7-day trial. Compare what you get — a list of attributed, dated, sourced commitments — against the bullet summary you'd normally file and forget. By the end of the week the difference in shape is the difference you'll feel. You can start free and let the trial do the arguing.
A good story isn't a list of topics. It's who said what, and whether it held.
— Richard
