How We Make Sure Your Call Transcripts Are Reliable

One line in our own transcript archive reads, "You've done a great job." The file puts that sentence next to the name of someone who works at a client, and for a while it sat in a folder of things clients had said about our work.

I said it. I was praising their spreadsheets.

Every word in the sentence is correct. The only thing wrong is the name on it, and the name is the entire reason the sentence was worth keeping.

That line came out of a larger exercise. In September we read back everything we had, 687 call transcripts, 38 meeting notes and my own mailbox, to build one file of the outcome numbers this firm can defend in public. The file opens with two hazards, and the first governs every number under it: speaker labels are unreliable in roughly a third of the transcripts.

The labels fail in a way you cannot see

I had assumed the failure would be a clean swap. Two people on a call, the labels on backwards, obvious in the first paragraph.

It is not that. What happens is absorption. One person's lines get pulled under another person's label, so a question and its answer are credited to the same speaker and the block reads as one person holding forth. Long back-and-forth exchanges collapse into a single labeled chunk. The worst files are the ones where somebody who demonstrably attended has no label at all and their words are parceled out among the people who do. In one of ours the person with no label is me, which is how a client ends up quoted saying something I said about them.

A mislabeled transcript is not corrupt, not empty and not obviously odd. It reads clean. It is the same hazard as a number that is wrong in four systems and consistent in all four: it passes every check a reader thinks to make. Reconciling one number across systems is the structured version of this problem. Transcripts are the version nobody reconciles, because they do not look like data at all.

Who said it decides what it is worth

Attribution, not wording, is what nearly sank the exercise.

The file grades every candidate figure by who produced it, on four levels. Measured by the client: a number they observe or read off a system. Estimated by the client: their own recollection or round figure. Claimed by us. Or a projection, a target, a success criterion. The rule written at the top is to never blend them, because a claim is only as good as its weakest input.

Each of those levels is a question of whose mouth the sentence came out of. Strip the name off a number and you cannot grade it, and what you cannot grade you cannot publish.

Three examples, all harmless when attributed and dangerous when not.

A number a client quoted to us in passing turned out to be another vendor's marketing claim about a different firm's advisors. Read without the attribution, it looks like a result. It is somebody else's brochure.

A figure about pricing decks reads in the transcript like a measurement of the job today. Go back to the question that produced it and it is the answer to "what does success look like six months from now." It is a target, and the post that uses it says so.

And a closing-ratio figure for a retirement planning firm came up on a call as an off-the-cuff estimate, in a file where the speaker labels are scrambled, about a result that moves with sales coaching as well as with software. It is the most quotable number in the archive. It is not going on the website.

What we changed

Four things, and none of them is a model upgrade.

Every quote carries an attribution note. Not a flag on the bad ones. A line on all of them, saying how confident we are that this person said this. The ones that read "clean label, corroborated in a second file" are the ones that reach a page.

Written sources outrank spoken ones. Email attribution is not ambiguous: a sender is a sender. Once we treated the mailbox as a first source rather than a backstop, it held more usable material than the transcripts had suggested.

A number has to survive a second telling. A figure that appears once, in one file, under a label we are unsure of, stays in the working file and off the site. The best-corroborated number we have was restated three times on one call, with clean labels each time, and again in a meeting note the same day, and that is why it is the one we lead with.

Every section ends with a do-not-use list. The figures that did not survive sit next to the ones that did, with the reason. That list is the part of the file I reach for most, because the temptation is never to invent a number. It is to reuse a good one that has quietly lost its paperwork.

If an agent is about to read your recordings

Plenty of people are pointing models at call recordings now, and most of those uses are fine. If the output is a summary and action items for the people who were in the room, attribution barely matters. They remember who said what.

It matters completely when the recording is evidence: a commitment, an approval, a number, a quote for a slide, anything a decision will rest on later. Before you build on that, three things are worth half a day.

Pull ten transcripts and check two attributed lines in each against the recording or against what you remember. You will know within an hour whether your labels hold up, and the answer is specific to your meeting tool and how much your people talk over each other.

Then require citations. Anything that reports what somebody said should hand back the file and the timestamp with it, so a person can check in seconds rather than hunt. Every quote in our file carries both, and that is the only reason the checking was possible at all.

And grade before you aggregate. Decide what the source makes each claim, then keep those piles apart to the end. Nothing undoes careful reading faster than averaging it.

The file that came out of that reading is shorter than the pile that went in. That is the normal result of grading your own evidence honestly, and it is the only version worth putting a name on.