How often does a real person actually pick up?
Across 2,740,075,503 detections in the window, 12.2% resolved to a live human and the rest to a machine — voicemail, an IVR, or a carrier recording. That ratio is the whole economic case for detection: about 8.8 in every 10 answered calls are answered by something that is not a person.
Is false answer supervision getting worse?
In this window it rose. FAS moved from 7.4% of calls on the first day of the window to 9.8% on the last, while the human rate stayed inside a 11.1–13.3% band on full-volume days. The saw-tooth is the weekly cycle, not a data gap: a typical busy day runs 144,321,290 detections against 19,802,143 or fewer on the quietest days.
Which line types reach humans, and which burn calls?
Landlines still answer. RBOC and independent-telco numbers reach a live person at more than double the rate of PCS wireless, and carry a fraction of the false answer supervision. Only line types with over 100 million calls in the window are shown; the long tail is too thin to read.
Does the hour you dial change who answers?
Barely. Between 09:00 and 20:00 US Eastern — the hours that hold 98.8% of all volume — the human rate sits in a narrow band. The excluded overnight buckets show higher percentages, but the largest of them holds 12,476,706 calls against 362,679,277 in the busiest hour, so those are small-sample artifacts rather than an argument for dialing at 4am.
Is false answer supervision spread out or concentrated?
Concentrated. Of 2,876 carriers seen in the window, the largest three account for 21.5% of every false answer, and the largest ten for 41.8%. We do not publish the names. The data shows where FAS is observed, not who is responsible for it, and naming third parties would be a claim we cannot settle with data we are able to open.
What do the machines actually say?
Detection is not one signal. These are the greeting types the classifier separates most often in the window. Two buckets are honest admissions rather than findings: “unclassified greeting” is audio the classifier could not place, and “no audio captured” is a call that ended before anything usable arrived.
The key figures in one table
| Figure | Value | Derived from |
|---|---|---|
| Detections analyzed (30-day window) | 2,740,075,503 | SUM(count), amd_carrier_daily |
| Live human answers | 334,821,032 (12.2%) | SUM(human_count) / SUM(count) |
| Machine answers | 2,405,254,103 (87.8%) | SUM(amd_count) / SUM(count) |
| False answer supervision | 368,990,180 (13.5%) | SUM(fas_count) / SUM(count) |
| Spam-trap (honeypot) hits | 2,418,874 (0.088%) | SUM(honeypot_count) / SUM(count) |
| Distinct outcome classes | 471 | COUNT(DISTINCT classification), amd_hourly_classification |
| Classified events | 3,219,070,300 | SUM(count), amd_hourly_classification |
| Average audio per verdict | 3.98s | volume-weighted avg_duration, amd_hourly_stats |
| Carriers observed | 2,876 | COUNT of distinct company, amd_carrier_daily |
| FAS from the top 3 carriers | 21.5% | top-3 fas_count / total fas_count |
| FAS from the top 10 carriers | 41.8% | top-10 fas_count / total fas_count |
| Window covered | Sun Sep 13 to Sun Oct 11 | MIN(day) / MAX(day), excluding the partial current day |
What this data cannot tell you
These are our detector’s own verdicts, not externally labelled ground truth. That rules out a whole class of claim, and we would rather say so than fill the gap:
- Accuracy, precision, recall, or a false-positive rate. There is no labelled ground truth here, so any percentage would be invented.
- Time to identify a human specifically. The hourly aggregate carries one average duration for all calls, with no classification split.
- Latency percentiles. The aggregates store averages only.
- Connect-rate lift versus dialing without AMD. There is no control group and no pre-AMD baseline.
- Revenue, ROI, or dollars recovered. Nothing financial exists in this database.
- Customer counts or “trusted by N companies”. These aggregates hold IPs, not identities, and IPs are not customers.
Methodology
All figures on this page come from AMDY’s own detection logs, queried live from our analytics database when the page is rendered. Unless stated otherwise, each number covers a trailing 30-day window ending the day before the page was last built, and the window is printed alongside the charts. Because it is a rolling window, the numbers move: a figure published in an earlier post or a previous build reflects a different 30 days and will not match. That is expected, not an error — our State of AMD 2026 post reports 12.5% human on its own earlier window, against 12.2% here.
Everything is aggregated across the entire platform. No individual customer, account, IP address, phone number, or campaign is identified, and no figure is broken out in a way that would isolate one account. Carrier-level analysis is published as concentration and by regulatory line type, never by carrier name — the finding holds either way.
Two limits worth stating plainly. First, these are our detector’s own verdicts, not externally labelled ground truth, so nothing here measures detection accuracy and we make no accuracy claim from this data. Second, the carrier and line-type breakdowns cover only calls whose number maps to a known North American NPA-NXX in LERG (2,740,075,503 calls), which is a smaller set than total classified events (3,219,070,300); the two denominators are never mixed inside a single chart. Every figure drawn from the daily table excludes the current day, which is always partial; the hour-of-day, greeting and classification figures come from hourly aggregates over a rolling 30×24-hour clock instead.
Run it against your own traffic
Sandbox is free at 50,000 detections a month, no card. Your dashboard shows the same breakdowns for your own calls.
Start free