We ran a commitment ledger over 15 real meetings. It found four failures the note taker missed.

Fifteen meetings were recorded and transcribed over roughly six weeks by a commercial note taker. Every one produced a tidy summary and a list of action items. Every item was marked Pending.

We extracted the commitments and stored them differently: not as a status, but as a history. Same commitment, tracked across meetings, keeping every version of its owner and its date.

That single change surfaced 111 commitments, nine of which had never been assigned an owner at all, and four failures that had been invisible.

One: the task that changed hands without a handover

A launch page was assigned to the content lead on 22 June. By 30 June the same task sat with the ops assistant, due date already five days gone.

No sentence in either meeting hands it over. Nobody said "I will take that." The owner simply differs between two records. The note taker wrote Pending twice and had no way to notice, because noticing requires holding both meetings at once.

Two: two money commitments that evaporated

On 22 June, two commitments to transfer a platform balance to the founder. On 30 June, neither appears anywhere in the list.

Not completed. Not cancelled. Not deferred. Absent. And absence is the one thing a per-meeting summary structurally cannot report, because it only knows what was said, never what stopped being said.

Three: a deadline whose arithmetic was already dead

A training course at 16 of 649 sections, due in six days, named in the room as the blocker gating an entire ad campaign.

That is 105 sections a day. Nobody in the meeting did the division, and neither did the summary, because summarising is not the same as checking whether a plan is possible.

Four: one person holding the launch

Eleven of thirteen action items on 22 June belonged to one teammate. Ten of sixteen on 30 June, plus two items with no owner recorded.

That teammate resigned the following month. Nothing had a second owner. The concentration was visible in the data the whole time; it was just never counted.

Why extraction is not the hard part

Turning speech into action items is close to solved. Several products do it well, and this is not a criticism of them.

The gap is that every one of them treats a meeting as an island. The list is generated, delivered, and then the next meeting generates a fresh one. Nothing compares Tuesday to three Tuesdays ago.

Every failure above is only visible in the comparison. Owner drift needs two records. A silent drop needs an absence. Impossible arithmetic needs the progress number and the deadline in the same frame. Load risk needs a count across the whole set.

What we actually built

Five detectors, each one a question you would ask if you had the time to read every transcript side by side every week:

• Silent slip. Open, past due, and unmentioned in the next meeting.
• Owner drift. Same commitment, different owner, no handover said out loud.
• Silent drop. Present, then absent twice, never marked done.
• Undecided change. A date or estimate moved with no decision beside it.
• Load risk. Too much open work on one person, or a keystone with no backup.

Every commitment is also tagged by what it touches: scope, a date, a dependency, or money. Anything touching none of them sinks. A to-do app does not care that a date moved without a decision. A product manager is judged on exactly that.

The honest limits

This is one operator's fifteen meetings, one team, one company. That is enough to show the failure mode is real and expensive. It is not enough to tell you how often it happens to you.

We do not publish a time-saving number, here or anywhere else on this site, because we have not measured one under a method we could show you. The claim is coverage: four out of four, on records you could audit if they were yours.