Win loss analysis from the conversations you already recorded

The published method for this is a program, and it is the same program on every page of the results. Define the scope, pick a sample, write the interview guide, bring somebody in from outside so the seller is not in the room, run twenty to thirty minute calls, keep going to roughly twenty-five or thirty interviews per segment, publish a report each quarter. None of it is wrong. It also describes a research project you have to fund and staff before it answers anything, and the first answer lands a quarter after the deals it describes.

The sharper guides open by attacking the CRM field, and they are right to. Clozd, which sells win loss programs, puts the agreement between what a buyer says and what the seller recorded at about 15%, so roughly 85% of closed lost reasons in a CRM are wrong or incomplete. The figure appears in its win loss guide and is attributed to its own 2025 report, with no sample size, period or method published alongside it. A number about bad data arriving without its own denominator is worth noticing. The direction is almost certainly right. The magnitude is not something anyone outside Clozd can check.

What none of these guides do is count past two. The CRM field is one record of the reason. The interview is a second. Both are reconstructions, produced later, by people with something at stake in the wording. There is a third record, it was made at the time, nobody had to be asked for it, and it is sitting in a folder of audio files.

A menu of loss reasons returns the answers that cost nothing to pick

The loss reason field is a form with a list attached, filled in by the person whose quarter the deal just left. Too expensive and no budget dominate those lists everywhere, for a structural reason: they are the two entries on the menu that nobody follows up on. An entry that invites a follow up question gets picked less often than one that ends the conversation.

Two other things happen to that field. The list was written by somebody who was not on the call, usually once, usually years ago, and it has to cover every deal the company will ever lose. And the field is almost always single select, so a deal that died of three things gets filed under one of them, which is the same flattening that turns every mention of a rival into one kind of event. Neither problem is fixed by training people to fill it in more honestly, because neither problem is about honesty.

The field is also not failing at its job. Its job is to make a pipeline report add up, and it does that. It was never built to carry an explanation, and treating it as the baseline that win loss analysis corrects gives it a status it never claimed. The question is not whether the field is accurate. The question is what you would compare it against. Which of your closed lost rows still has a recording behind it?

An interview three months later returns an account, not a memory

A win loss interview collects what the buyer now believes happened. That is a real thing to collect, and it is not the same thing as what happened. The gap is not about honesty either. People produce fluent and confident explanations for their choices, and the explanation gets assembled at the moment of asking, out of whatever seems like a sensible reason to give.

The classic demonstration is nearly fifty years old and still the cleanest. In Telling More Than We Can Know, published in Psychological Review in 1977, Nisbett and Wilson describe a study run in a shop under the cover of a consumer survey. Fifty-two passersby were asked which of four pairs of nylon stockings was the best quality. The four pairs were identical. The rightmost pair was chosen over the leftmost by a factor of almost four to one, and when asked why, people talked about sheerness, weave and knit. Not one of them mentioned position on the table. Asked about position directly, virtually all of them denied it mattered.

The modern version is harsher. Johansson, Hall, Sikström and Olsson had people choose which of two faces they found more attractive, then handed back the face they had rejected and asked them to explain the choice. Across all conditions, no more than 26% of the switched trials were caught. For the rest, people explained a preference they had not expressed, and blind raters found the explanations no less emotional, no less specific and no less certain than the genuine ones. A confident, detailed answer in an interview is not evidence that the answer is a recollection.

The win loss field already concedes this quietly. Several of the better guides tell you to interview within four weeks of the close. That instruction only makes sense if recall decays, and once you accept that it decays, you have accepted that the interview is a race against it.

The recording is the only version made while the reason was live

Each of the three records something the other two do not, so the useful move is to hold them apart instead of ranking them. The CRM entry tells you what a team decided to file. The interview tells you what the buyer has since settled on. The recording tells you what was said, in what order, and what the response to it was.

Read down the last column and the case for keeping all three gets stronger, not weaker. Interviews reach the people who never had a call with you, the part of the decision that happened in a room you were not in, and the comparison with the vendor who won. Recordings reach the moment the buyer named the thing in their own words, before anybody had a story about it. Neither replaces the other. What is odd about current practice is that the cheapest of the three, the one already paid for and already collected, is the one most programs skip.

There is a reason it gets skipped, and it is not mysterious. Listening through a month of recordings to find the sentence that mattered is a bad use of a week, so nobody does it, so the material stays unread. That is a tooling problem wearing the costume of a methodology problem, which is the better of the two to have.

Deal status read from conversations is not a funnel outcome

A status derived from conversations and a status derived from a CRM are two different measurements that happen to share a vocabulary, and mixing them produces nonsense quietly. In Spellit ClientTruth the status on a client sits on four values, winning, needs work, losing and dead, and the screen carries a plaque saying that the status there is not a funnel outcome and comes from the conversations rather than from a CRM. The plaque is not modesty. It stops somebody exporting the number into a forecast.

We learned how much the framing of a report matters by building the wrong one first. The weekly customer voice screen originally opened with what was going well: the healthy shares, the things that had improved. It was judged harmful before it reached customers, on exactly one argument, that a reader would finish it thinking everything was fine and nothing needed doing. The screen was rebuilt the other way round, opening with what blocked clients that week and what got worse against the week before. The second version is less pleasant to read and it is the only one anybody acts on.

That applies directly to a win loss report. A deck that opens with the win rate and puts the losses in an appendix has made the same mistake we made, and it will get the same result, which is agreement followed by nothing. Put the losing side first and make the first slide a group of real sentences, not a bar chart. We can put the form entry, the interview and the recording for one lost deal side by side if you want to see how far apart they land before committing to a process around them.

A loss reason is finished when it reaches a sentence somebody said

Put a group in a report with no buyer sentence under it and you have published a hypothesis. It can still be useful, as long as it is labeled as one, because the distance between a tag that says budget and a buyer saying they had already committed the money to something else is the distance between a chart and a decision. In ClientTruth every finding is tied to a verbatim buyer line with two turns either side and a timecode, and the proof link opens the call itself inside the Actions side of the product.

Our own uncomfortable finding here is about what a screen is allowed to say. The client card reads the whole client row, 155 columns, and renders a small part of it. Fourteen fields arrive in the response and never make it onto the screen, among them the flags for lost, recoverable and warming, and the reasoning the model produced behind the intent score. Nothing is broken, nothing needs computing, and the layout work is a day.

What held it up was the question the plaque is about. A flag that says a client is recoverable, derived from conversations, shown next to a status derived from conversations, will be read by somebody as a forecast within a week of shipping. We have been slower to settle what those words are permitted to claim than we were to calculate them, and I think that order is correct even though it looks like a backlog.

The grouping work itself has its own rule, which is to group the wording first and name the category afterwards, and we set it out in customer pain points analysis from calls.

Where the standard program is right, and what recordings cannot do

None of this is an argument against interviews. The win loss program literature is right about several things a recording cannot touch, and a team that drops its interview program because it has call recordings will lose real coverage of the decision.

The second column is the honest limit. Recordings cover the deals where somebody talked to you, which is not all of them, and they stop at the last call, which is often before the decision. The arrangement that works is boring: read the calls first, because they are already there and they cost nothing to read, then spend the interview budget on the questions only an interview can reach. That is close to the opposite of the usual order, where the interview program goes first and the recordings never get opened.

Counting loss reasons runs into the same denominator trap as counting anything else out of a month of calls, and that trap is set out in voice of customer analysis without surveys. The second thing that bites is vocabulary: how many distinct things your list of labels actually tells apart. We learned the size of that one counting objections, in what 424 objection labels look like in practice.

What to do next

Take one deal your team still disagrees about. Pull three things: the entry on the form, whatever anybody remembers from the debrief, and the last two calls before it went quiet. Read the calls last, and write down the first sentence where the buyer names a problem in their own words. Compare that sentence to what is on the form.

Do it for five deals, not one. One deal is an anecdote and will prove whatever you already think. Five is enough to see whether the gap is random or points one direction, and if it points one direction, that direction is usually the most informative thing you will learn this quarter.

Then decide what your interview budget is for. If the five deals show that the calls already contain the reason, the interviews should be reaching for what the calls cannot: the buying group members who never appeared, and the weeks after the last call. Pick the deal your team still argues about and let the recording answer first.