My AI Read 400 Of My Own Sales Calls. The First Answer Was Wrong.
I published that 46% of stalled calls gave no reason and price was only 17%. Both were wrong, and not for the reason I first assumed.
Correction, 27 July 2026. An earlier version of this page said 46% of stalled deals gave no objection, that price was named in about 17% of stated blockers, and that most of these deals stayed “pending forever”. All three were wrong. I have rebuilt the analysis and rewritten the page around what the corrected numbers actually say. Figures published here before today should not be quoted.
Short answer: I pointed my own AI at 400 recorded calls to find out why deals stall. The first run gave me a clean story: it is not price, it is silence. Checking it found 3 separate bugs. About 30% of stalled calls still show no stated reason, which is real and is the headline. But price is not a small slice. It is roughly 30% of the reasons people did give, about level with timing. The tidy version of this story was an artefact of my own tooling.
I run my business on AI systems I build myself. One of them reads every recorded call and classifies what happened: won, lost, still pending, and the single biggest thing blocking the deal.
I pointed it at 400 of my own calls going back to August 2024. It came back with a clean, quotable story. Deals do not die on price. 46% of the stalled ones gave no reason at all, and price was only about 1 in 6.
I published that. Then I went looking for how it knew, and found 3 bugs.
The 3 bugs
Bug 1: it was only reading the start of each call. The classifier took the first 5,000 characters of each transcript and nothing else. My median transcript is about 36,000 characters, so it was seeing a median 14% of each call, and on 254 of the 400 it saw less than a quarter. That part is the opening: the hello, the rapport, the “so tell me about your business.” Objections do not happen there. Neither does the close. So when it reported “no objection”, what it often meant was “no objection yet, in the first few minutes”.
Bug 2: the price percentage had the wrong denominator. To get “17%” I had divided the price calls by every call that stalled, including the ones that named no reason at all, plus some that were won or unclassifiable. Dividing a share of stated reasons by a population that mostly did not state one guarantees a small number. It was never a measurement, it was a bad fraction.
Bug 3: price was split across 2 labels. The classifier was emitting “price” and “pricing” as separate categories, so the biggest money bucket was cut in half before anyone looked at it. Money objections now collapse into one label.
Only the first of those is the interesting one. The other 2 are ordinary arithmetic and ordinary sloppiness, and between them they did more damage than the AI bug did.
What the numbers actually are
I rebuilt the classifier to read every transcript end to end. All 400 calls, 100% of each one. To separate the reading bug from the arithmetic, here is the like-for-like comparison on the 292 calls that appear in both runs:
| Partial read, 14% of each call | Full read, 100% | |
|---|---|---|
| Stalled calls with no stated reason | 44% | 30% |
| Price, share of stated reasons | 26% | 30% |
So reading the whole call is what moved the silence number, from 44% down to 30%. It moved price by about 4 points. The gap between that 26% and the 17% I published is the denominator error and the split label, not the AI.
Across all 400 calls, the full ranking of stated reasons, over the 246 stalled calls where someone actually named something:
| Reason | Share of stated reasons |
|---|---|
| Timing | 33% |
| Price | 30% |
| Fit or scope | 15% |
| Partner or spouse decision | 11% |
| Trust or proof | 8% |
| Doing it themselves | 1% |
| Location | Under 1% |
| A competitor | Under 1% |
“Price” here means anything about money: too expensive, cannot afford it right now, cash flow, wanting a smaller package. “No stated reason” is its own warning label, because that bucket mixes calls where the person genuinely named nothing with calls where the tool could not tell. I cannot separate those 2, so treat it as an upper bound on real silence. The raw counts, if you want to check the rounding yourself: 108 of the 354 stalled calls named no reason, and 75 of the 246 stated reasons were about money.
What I got most wrong
I told people price was a small slice, about 1 in 6. It is about 3 in 10, and it is level with timing at the top of the list.
The lesson I originally drew from this page was “stop discounting, you are optimising the smallest slice.” That advice was built on a number I had computed incorrectly. Price is not the whole story, and silence is still the single biggest bucket, but price is not a rounding error either and I should not have told anyone it was.
What I removed completely
The old page said most of these deals were “pending forever”. Never won, never lost.
I cannot support that, so I have taken it out.
“Pending” in my data means the call ended without a decision on the call. That is all it means. What happened after the call is not something I measured. The only system that could have recorded it is my CRM pipeline, and I have not maintained it. It holds 262 opportunities with exactly 1 marked won and 1 marked lost, last touched in June 2025.
So I know how the calls ended. I do not know how the deals ended. Saying otherwise was me describing a gap in my own record keeping as if it were a finding about buyers.
What held up, and what did not
The silence finding held up. It shrank from 46% to about 30%, but it is still the biggest single bucket, and it is the thing worth acting on.
My close rate did not hold up as a claim. On 88 calls the transcript shows a clear yes or no on the call itself, and 29 were a yes. I am not calling that a close rate. A yes on a call is not a collected payment, and the tag is noisy: several of those 29 are check-ins or support calls with people who were already clients, and a few are the same conversation recorded twice. Read it as “somewhere between a quarter and a third of calls that reached a decision reached a yes”, and do not read it as a business conversion rate.
819 bookings on my main call calendar between June 2024 and July 2026, of which 248 were cancelled before they happened. That calendar carries coaching and catch-up calls too, so it is not 819 sales calls.
And one number the fixed tool could measure that the old one could not:
On 45% of calls, nobody asked for a decision or set a dated next step. That is not the buyer’s fault. It is mine. It is also visible in the transcript, which is exactly why I trust it more than “pending forever”.
What I changed in how I sell
Ask the 2 questions early: what is the biggest thing in the way, and what do you want the next 6 to 12 months to look like. Then stop talking.
Never end a call without a decision or a dated next step. 45% of mine ended without either. This is the one I would fix first if I were you.
Treat price as a real objection again. That was the mistake in my first read. When price comes up it is often standing in for “I am not sure this works for me specifically”, but it comes up far more than I claimed.
Make it easy to say no. A clean no frees the hour.
The lesson that is bigger than my sales calls
Ask any AI system what percentage of the source it actually read. Then check the arithmetic you wrapped around its answer.
Mine gave me a confident, well formatted, completely quotable answer built on 14% of the evidence, and I then divided it by the wrong number and split its biggest category in half. Nothing about the output looked wrong at any stage. There is no warning label on a number that came from a partial read, and none on a clean-looking fraction either.
Every row now logs its own coverage: how many characters existed, how many were read, what fraction that is. If a number cannot say what it looked at, it does not go on a page.
How I ran this analysis
400 recorded calls from my sales calendar, August 2024 to July 2026. Some turned out to be coaching or check-in calls rather than sales calls. Transcribed, then classified by a language model against a fixed set of categories, 1 call at a time, full transcript, no sampling. Every row records its own coverage, and every row in this run read 100% of its transcript. Earlier versions of the classifier read a window rather than the whole call: first 5,000 characters, then a 12,000-character head and tail, then all of it.
These are calls, not distinct deals. The 400 calls came from 318 people, so repeat conversations with the same person are counted more than once. Of the 400, 59 ended lost and 295 ended without a decision, which is the 354 I am calling stalled. 29 reached a yes on the call and 17 were too unclear to classify. Booking counts come from my scheduling calendar, pulled directly from it.
This is 1 operator’s book of calls, not a market study. Treat it as a pattern worth testing against your own calls.
Common questions
Why trust the second set of numbers more than the first? Because I can tell you exactly what each one read and exactly how each one was calculated. The first saw a median 14% of each call and then got divided by the wrong denominator. The second read every call in full, and the arithmetic is on this page.
Does this mean price matters most? No. Silence is still the biggest single bucket, at about 30% of stalled calls. Among people who named a reason, price is about 30% too, roughly level with timing. Those 2 numbers sit on different denominators, so do not read them against each other.
What counts as a stalled call? A call that ended lost, or ended without a decision. 354 of the 400. A further 17 were too unclear to classify and are in neither group.
Is 400 calls enough to trust? It is 1 practice’s book, not an industry study. The useful part is the method, not my percentages.
Why leave the mistake up? Because the correction is the more useful post. If I quietly swapped the numbers, you would have no way to judge either one.
For journalists and analysts
The findings on this page can be quoted with attribution to Sam Eye Am, sameyeam.info. They come from 400 recorded calls from a single consulting practice, taken from its sales calendar and classified in July 2026, each read in full. These are calls rather than distinct deals, from 318 people. Figures published on this page before 27 July 2026 were produced by a classifier reading a median 14% of each transcript, with a denominator error in the price share, and should not be quoted.
See what an AI system would find in your business
This whole thing exists because the system was already reading everything. Including, it turns out, reading it badly for a while, and then me doing the arithmetic badly on top. That is what building in the open actually looks like.
That is what I build for experts: the machine that does the work you do not have hours for.
Not ready for that? Take the free course and see the method I use.