Call Screening Statistics: What Actually Answered 15,066 Live Outbound Calls
Every outbound calling operation on earth is being screened, and nobody has published how much. The industry talks about answering-machine detection as a two-class problem — machine or human — and every dashboard, every disposition report and every vendor accuracy claim inherits that assumption. We labelled 15,066 live answered calls to see whether the assumption survives contact with real traffic.
It does not. A third class is sitting inside the machine bucket, it is roughly one call in seven, and it is not an answering machine. It is a call-screening service: a carrier or handset feature that answers on the subscriber’s behalf, asks who is calling and why, and connects the call if the answer satisfies it.
This page is the measurement, the method, and — at length, because it is the part that decides whether the number means anything — what the sample cannot tell you.
The finding
Of 15,066 answered outbound calls that an answering-machine detection engine had already classified as non-human, 1,889 (12.54%) were a call-screening service, not an answering machine. Counted against the 14,228 calls whose audio was decidable at all, screeners are 13.28% (95% CI 12.73-13.84) — about 1 in 7.5.
| What answered | Calls | Share of 15,066 |
|---|---|---|
| Answering machine or voicemail | 12,101 | 80.32% |
| Call-screening service | 1,889 | 12.54% |
| Undecidable audio | 838 | 5.56% |
| Carrier failure or intercept | 130 | 0.86% |
| Live person | 108 | 0.72% |
The window is two days, 2026-07-23 and 2026-07-24. The two days agree with each other: screeners were 14.11% of decidable calls on the first and 12.87% on the second.
Screening is hiding inside the machine number
The class is invisible because the engine has no cell for it. Cross-tabulating the engine’s own verdict against the adjudicated label, 1,879 of the 12,959 calls it called MACHINE were screeners — 14.50% (95% CI 13.90-15.12), roughly 1 in 7 of its answering-machine verdicts. The remaining screeners are almost nowhere else: 10 of the 2,077 calls the engine recorded as having no audio, and none of the 30 it called a carrier failure.
Strictly, the engine is not wrong on its own taxonomy. A screener is genuinely not a live person, so hanging up is a defensible action given a two-class world. The problem is what the two-class world does to everything downstream. A screened call and a voicemail land in the same disposition bucket, which means:
- Your machine rate is not a machine rate. About one seventh of it is people who are home.
- Your list-quality analysis is wrong in a specific direction. Numbers that screen look exactly like numbers that never answer, so a retry policy tuned on the blended bucket is tuned on two different populations averaged together.
- Vendor accuracy claims are unfalsifiable on this axis. An engine scored as machine-versus-human on a corpus with no screener class will report a high score whether or not it can tell a screener from a voicemail, because both count as machine.
Screeners are the easiest class to detect, not the hardest
The most useful result for anyone building or buying detection is that this class is not a hard problem. It is the most reliably detectable class in the corpus.
| Class | First cue, median | First cue, p90 | Decidable within 2s |
|---|---|---|---|
| Call-screening service | 1,144 ms | 1,560 ms | 92.54% |
| Answering machine | 720 ms | 3,096 ms | 70.04% |
“First cue” is the start of the first word of the first decisive phrase, taken from speech-recognition word timings. Answering machines have the earlier median and a far heavier tail, because the class mixes an instant carrier number-readback against a personal greeting that rambles for seconds before it says anything decisive. Screeners are tightly clustered — their entire purpose is to interrupt the caller immediately — so by two seconds after answer, 92.54% of them have said something that identifies them, against 70.04% of answering machines.
Nothing about the audio makes this class hard. It is missing from detection output because it is missing from detection taxonomies.
What a screener actually says
Screener transcripts are formulaic, which is what makes the class separable in the first place. Of the 1,889 screener calls:
- 85.71% asked the caller to state or record a name.
- 85.28% asked the reason for the call.
- 10.59% named themselves as a screening service.
- 7.57% offered to connect the call.
These families overlap — most screeners ask for a name and a reason in the same breath — so they do not sum to 100% and are not a partition. They are also floors rather than exact prevalences: a screener whose phrasing our patterns missed is counted as not asking.
The shape of that script is the whole operational argument. A screener is a reachable subscriber behind a gate, not a dead number. It is asking for information. A dialler that hangs up on it is declining to answer a question that was asked, on a line where somebody is standing by, and then filing the call next to a voicemail nobody was home for.
Method
Capture. Audio was written straight off the dialler’s answering-machine detection path — for each captured call, the first few seconds after answer plus a sidecar recording the engine’s own verdict and exactly how many bytes of audio it decided on. No production configuration was changed to produce this data, and nothing was written back to the dialler.
Transcription. Every call was transcribed with a single speech-recognition engine on 8 kHz telephony audio, with per-word start and end timings. Those word timings are what the cue-latency figures above are computed from.
Labelling. Transcripts were labelled against a fixed five-class taxonomy defined before counting: answering machine or voicemail (including carrier number-readbacks, auto-attendant menus and hold announcements), call-screening service (asks the caller to identify themselves before connecting), carrier failure or intercept, live person, and undecidable. Every one of the 15,066 transcripts was judged, and every call proposed as a live person went through a second adversarial pass instructed to assume the first labeller had been careless; disagreements were resolved precision-first, so a doubtful live-person call became a machine or an undecidable rather than a human.
The label column is the load-bearing decision. The corpus carries two labellings and they disagree. An earlier phrase-list judge, ported from a different telephony stack, puts screeners at 2,333; the adjudicated labelling puts them at 1,889. Publishing from the retired column would have overstated this page’s headline by 444 calls, so the generator hard-codes the adjudicated column instead of exposing it as an option, and ships a check that proves the two files are what we say they are.
How far apart are they? Of the 1,718 calls the phrase-list judge called a live person, 1,622 were overturned — 94.41%. Only 96 survived, and 1,518 of the overturned calls were answering machines. Across all 15,066 calls the two labellings agree 84.63% of the time. That is the evidence for retiring that judge on this audio, and it is a directional measurement rather than a symmetric one: it counts overturns of one judge by another, and treats the adjudicated column as the reference because it is the one that went through an adversarial second pass, not because it is infallible.
Intervals. All confidence intervals are 95% Wilson, chosen because the smallest class sits below 1% of the sample, where the textbook normal approximation misbehaves.
Reproducibility. The aggregate table and the script that produces it are both in our repository. The script emits aggregates only and refuses to write at all if any phone-, email- or identifier-shaped string reaches its output. Point it at a fresh capture and every number on this page recomputes.
Limitations
This is the section that decides whether the number above is worth citing, so it is longer than the finding.
1. It is a stratum, not live traffic. This is the big one. Audio was captured only when the engine had already decided the call was not a live person. Every share on this page is therefore conditioned on “the engine hung up on this call”. It follows directly that the rate at which call screening answers an outbound call is not measurable from this data — a screener the engine passed through as a human was never recorded, so the denominator required for that rate does not exist here. If you cite one number from this page, cite it with that condition attached.
2. Two days is a snapshot, not a rate. The window is 2026-07-23 to 2026-07-24. Do not annualise it, do not multiply it out, and do not read a trend into two points. The two days agreeing (14.11% and 12.87% of decidable calls) is evidence of stability across those two days and nothing more. Screening adoption is a moving target driven by handset and carrier defaults, so this figure has a shelf life.
3. One dialler, one destination profile. These are the destinations one operation happened to be calling in one window. Screening prevalence plausibly varies by geography, by carrier mix, by handset mix, by time of day and by how a number was sourced. Nothing here establishes that 13.28% transfers to your list.
4. The unit is a call, not a person. A number dialled twice contributes two calls. Screening prevalence per unique destination is a different quantity from screening prevalence per call attempt, and this page reports the second.
5. The days are not equally weighted. One of the two is a complete capture day and the other is a random subsample of its day, which is why the per-day figures are quoted rather than only a pooled rate.
6. Labels come from transcripts, so speech-recognition error is inside every number. It is one engine, so its errors are correlated across all classes rather than averaging out, and a phrase it systematically mangles is mislabelled consistently. The screener class is comparatively robust to this because its scripts are formulaic and repetitive, but “comparatively” is not “immune”.
7. The undecidable class is real and is not zero. 838 calls (5.56%) could not be labelled from audio. We report them rather than dropping them, because a benchmark corpus that reports zero undecidable calls has been cleaned, and comparing a live share against a cleaned one flatters the cleaned corpus.
8. The live-person cell is not a general false-positive rate. 108 live people (0.72%) is the share of live people inside the stratum this engine had already discarded, on this dialler, in this window. It is not a claim about answering-machine detection generally, and it should not be quoted as one.
What we would want measured next
The obvious missing arm is a capture on the human side of the decision. Until a sample of calls the engine passes through as human is labelled with the same taxonomy, nobody can state an overall screening rate or an engine’s true recall — both quantities need the cell this capture structurally cannot see. That is one line of capture configuration and a labelling pass, and it is the difference between “1 in 7.5 of what we hung up on” and “1 in N of everything we dial”.
The second thing worth measuring is what happens after the gate. A screener asks for a name and a reason; nothing here establishes what share of callers who answer it get connected, because our dialler hangs up before finding out.
Why we ran this
We run outbound phone floors, so a class that is roughly one seventh of our machine bucket is not an academic curiosity — it is people we could be talking to, filed under a label that says do not bother. We published twenty real storm-damage roofing appointment calls in full and a study of 10,794 transcribed roofing appointment-setting calls for the same reason: the claims in this industry are almost never checkable, and the fix for that is to show the evidence rather than describe it.
The two studies also answer different halves of the same question. That one is about what happens on the calls that connect. This one is about a class of call that never got the chance to.
If you are choosing a dialler or a detection engine, the practical question this raises is short: ask what your vendor does with a call screener, and ask what corpus their accuracy number was measured on. If the answer to the first is “treats it as a machine” and the answer to the second is a two-class corpus, the accuracy number does not describe roughly one in seven of the calls it is being quoted about. That question sits alongside the rest of how a dialler and CRM stack actually fit together, and it is the kind of thing calling data tells you when you look at it properly.
Cite as: The Call Center Doctors, Call Screening Statistics: What Actually Answered 15,066 Live Outbound Calls, first-party study of 15,066 labelled answered outbound calls, 2026-07-23 to 2026-07-24. Figures are shares within calls an answering-machine detection engine classified as non-human.