The Analyst's Approach to Win/Loss Interview Analysis

The PMM who runs win/loss has a cabinet of forty-six transcripts and three weeks to turn them into a battle card update. They will read the first six carefully, skim the next ten, and pattern-match the rest from memory. The output will be three bullet points the sales team has heard before. This is the actual failure mode of win/loss programs — not that the interviews aren't happening, but that the analysis stops at the point a tired human runs out of patience.

An analyst — meaning an LLM with a structured prompt, the full transcript corpus, and a defined output schema — doesn't get tired at transcript twenty-two. That's the whole pitch. It's not smarter than the PMM. It's a second pair of eyes that will read every line, count every mention, and surface the quotes the PMM would have caught on a fresh Monday morning.

62%

This guide covers how to do that handoff well. What the analyst is good at, what it isn't, and how to structure the prompts so you don't end up with confident-sounding nonsense.

What the analyst is actually doing

It's helpful to be precise about the work. An LLM analyzing win/loss transcripts is doing four things, and only four:

• Extracting structured data from unstructured text. Pulling the moment a buyer named a competitor, the price they were quoted, the integration they asked about.
• Counting and clustering. Grouping similar objections across forty interviews so "the API is rate-limited" and "we hit throttling" land in the same bucket.
• Surfacing quotes. Returning the exact sentence a buyer said, with timestamp, so the PMM can verify it wasn't paraphrased into existence.
• Comparing wins to losses on the same dimension. What did won deals say about pricing that lost deals didn't?

It's not doing causal inference. It's not telling you why you lost. It's telling you what was said, by whom, how often, and in what context. The interpretive work — the part where you decide which of the seven mentioned objections is actually the one that killed the deal — stays with the human.

Step 1 · Get the transcripts into a usable shape

Before the analyst sees anything, the transcripts need three things: speaker labels, timestamps, and a consistent file format. If your recordings come from Gong or Chorus, the export already includes these. If they're Zoom auto-transcripts, you'll spend an hour cleaning speaker attribution before any analysis is worth running.

Strip the filler. The model has a context budget — usually generous, but not infinite — and feeding it forty-five minutes of small talk per interview burns tokens that should go to actual signal. A clean transcript is roughly 60% the length of the raw one.

Tag each transcript with three pieces of metadata in the filename or header: deal outcome (won/lost/no-decision), competitor (if known), and ICP segment. The analyst will use these as filters.

Step 2 · Prompt the analyst with a schema, not a question

This is where most attempts go sideways. The PMM opens a chat window, pastes a transcript, and asks "what did this buyer care about?" The model returns a fluent paragraph that sounds insightful and contains roughly nothing the PMM didn't already know.

The fix is to prompt with structured output requirements. Tell the analyst exactly what fields you want, in what format, with what evidence requirements.

The verbatim-quote rule is the single most important discipline. It forces the model to ground its claims in the source text and gives the PMM a fast verification path. If the model "extracts" an objection but can't produce the line where the buyer said it, that objection didn't happen.

Step 3 · Run interviews in batches, not one at a time

Single-transcript analysis is useful for spot-checking. The actual value shows up at batch scale, where the analyst can count across the corpus.

A typical batch prompt: "Across these twelve lost deals to Competitor X, list every distinct objection raised, the count of interviews mentioning it, and one verbatim quote per objection. Sort by frequency."

What you get back is the rough draft of a battle card update. Eight objections, ranked. The top three account for two-thirds of the mentions. The bottom three are noise. You — the human — decide which of the top three is the real one and which is the cover story.

LP

We were running win/loss for two years with a spreadsheet and a lot of highlighting. The first batch we ran through the analyst surfaced an objection that had been mentioned in fourteen of nineteen lost deals. We'd somehow missed it for eighteen months because no individual interview made it sound urgent.

Step 4 · Compare wins to losses on the same dimensions

The diagnostic that earns its keep: take the same prompt schema, run it across won deals and lost deals separately, and compare the distributions.

If "integration with Salesforce" shows up in 70% of won deals as a driver and in 60% of lost deals as a blocker, you have a real story about who the product fits. If pricing shows up as the named driver in 40% of losses but only 5% of wins mention price as a reason they chose you, pricing is probably a proxy for something else.

The analyst can run this comparison in minutes. Doing it by hand across forty interviews takes two days and the PMM tends to weight whichever transcripts they read most recently.

Step 5 · Sanity-check the patterns before they leave the room

The temptation, once you have a clean output, is to ship it. The discipline is to spot-check.

Before the output becomes a battle card

This step takes thirty minutes. It's the difference between a battle card that survives contact with sales and one that gets quietly ignored after week two.

What this doesn't replace

The analyst doesn't replace the interview itself. It doesn't replace the PMM's judgment about which patterns matter. It doesn't replace the conversation with the AE who lost the deal and has context the buyer didn't share.

What it replaces is the part of win/loss that nobody likes: reading transcript twenty-two on a Friday afternoon and pretending you're as sharp as you were on transcript four. That's the part where insights get missed, and that's the part the analyst is genuinely good at.

A win/loss program lives or dies on whether the analysis stays as honest as the interviews.

The prompt pack

The prompts that produce the outputs above are not magic — they're just specific. We've packaged the working set we use in client work: extraction prompts, batch-comparison prompts, the verbatim-quote rule baked into each, and the output schemas formatted for paste-into-spreadsheet.

What to do Monday

Pull the last six lost-deal transcripts. Run one extraction prompt against each, with the verbatim-quote rule. Compare the objections list to your current battle card. If they don't overlap, the battle card is older than the buyer's reality — and you found that out in an afternoon instead of a quarter.