Spam Likely: how bad AMD wrecks your caller-ID reputation

You can buy the cleanest list, hire the best agents, and write a perfect script — and still watch your connect rate sink, because the prospect’s phone showed “Spam Likely” and they let it ring out. That label is decided before the call even completes, by systems you don’t control. But you feed those systems with every dial you place, and the quality of your answering machine detection quietly shapes what they see.

This is the part most operators miss: caller-ID reputation isn’t only a carrier problem or a list problem. It’s also an AMD problem. Here’s how the flagging works, where bad detection makes it worse, and what actually protects your DIDs.

How numbers get flagged

The “Spam Likely” and “Scam Likely” labels come from carrier and third-party analytics engines — the reputation layer riding on top of the phone network. They don’t read your intent or your script. They score behavior. Every number that dials gets a moving reputation score, and a handful of signals push it up or down:

• High dial volume from one number. A single DID blasting thousands of calls a day looks nothing like a normal subscriber. Volume alone is a red flag.
• Low answer ratio. If most of your calls go unanswered, the engine reads it as “people don’t want this number’s calls.”
• Low talk-time ratio. Calls that connect but end in seconds — no real conversation — look like robocalls.
• Short-duration calls. A pattern of sub-ten-second calls is one of the cleanest robocall fingerprints there is.
• Complaint rates. Subscribers reporting your number, or carrier-side “don’t connect this again” signals, hit your score directly.
• Honeypot / spam-trap hits. Calling a planted trap number is the loudest signal of all — it says your list is scraped or dirty, and it penalizes you fast.

None of these is a single switch. They combine. Cross enough thresholds and the label attaches algorithmically — no human reviews it, no one tells you, and it lifts far more slowly than it appeared.

The AMD connection

Now look back at that signal list and notice how many of them your answering machine detection touches. Three connections matter.

1. Stock AMD is blind to honeypots and spam-traps

Stock Asterisk AMD — app_amd, configured in amd.conf — has only two buckets. It sets AMDSTATUS to HUMAN or MACHINE (with a NOTSURE gray zone in between), and that’s the whole vocabulary. It cannot tell you a number is a trap, because “trap” isn’t a category it knows. So you keep dialing the exact numbers that train the flagging engines against you, pass after pass, with no way to see it happening. The single most damaging signal for caller-ID reputation is the one stock AMD is structurally incapable of catching.

2. Bad AMD wrecks your answer and talk-time ratios

Heuristic, timing-based AMD is wrong ~15–30% of the time — it runs around 70–85% accuracy and drops 5–15% of live humans by misreading a fast “Hello?” as a machine (see why “Hello?” breaks default AMD). Every dropped human is a connect that should have become talk time and instead became a sub-ten-second hangup. That’s a double hit: you lose the lead, and the call now looks exactly like a robocall to the reputation engines — short duration, no conversation. Multiply across thousands of daily dials and bad AMD is actively manufacturing the negative signals that get you flagged.

3. FAS and dead-air drag the same metrics down

Carrier false-answer supervision (FAS) is when the network marks a call “answered” that nobody actually picked up — across our network data it’s about 14% of answered calls. Those connections produce dead air, zero talk time, and short durations: more of the same robocall fingerprint. Stock AMD can’t distinguish a FAS connection from a real one, so it can’t protect you from it. (More on this in carrier false answers.) Between FAS, dead air, and dropped humans, the metrics that decide your reputation are being shaped by detection you can’t see into.

The compounding cost

Here’s how it spirals. A few DIDs get flagged. Answer rates on them fall, because labeled calls ring out. The floor notices connect rate dropping, so the operator rotates in fresh DIDs to replace the burned ones. Those fresh numbers get pushed into the same dialing pattern — same volume, same dirty list segments, same blind AMD — and within days or weeks they flag too. So you buy more. And more.

That’s the DID treadmill: a recurring cost in number provisioning, registration, and management overhead that never fixes the root cause. You’re treating the symptom (a flagged number) and re-creating the disease (dialing traps, manufacturing short-duration calls) on every replacement. On 10,000 daily dials, even a small honeypot-hit rate and a 10% human-drop rate compound into a steady stream of newly-flagged numbers — an expense that scales with your dialing, not with your revenue.

How detection helps

Better detection doesn’t scrub a label off a number. What it does is remove the causes that keep attaching new ones. AMDY is an AI/ML model that classifies the acoustic signature of the answer audio — the sound of the pickup, not a transcript — in under 200 ms at 99% accuracy, and it classifies into more than two buckets: human, voicemail, FAS, honeypot/spam-trap, fax, and silence.

Two of those buckets directly protect reputation:

• Honeypot / spam-trap detection. When AMDY flags a dial as a trap, you can suppress that number on the next pass — you stop feeding the engines that flag you. This is reputation protection at the source, and it’s the thing stock AMD simply cannot do.
• Accurate human-vs-machine classification. Keeping the live humans stock AMD would have dropped means more real conversations and longer talk time — the positive signals the reputation engines reward, instead of the short-duration ones they punish.

Be honest about the limits, though. Detection is one layer. It removes a major cause of flagging; it does not by itself launder an already-flagged DID or replace good list hygiene. Pair it with rotation, warming, and monitoring — the practices below — and you get a system where your numbers stay clean because you stopped doing the things that dirty them.

A reputation-protection checklist

• Rotate DIDs. Spread volume across a healthy pool. No single number should look like a one-DID call center.
• Warm new numbers. Ramp fresh DIDs with low, steady volume before scaling them. Cold numbers dropped straight into high volume flag fast.
• Use branded caller ID / CNAM. Where your carrier supports it, a recognizable name raises answer rates and counters the “unknown number” penalty.
• Monitor every DID. Track answer-rate per number so you catch a flagged DID early instead of after a week of wasted dials.
• Avoid traps with real detection. Use honeypot/spam-trap detection so you stop dialing the numbers that train the flagging engines — the cause-removal step everything else depends on.

AMDY runs as a telco-agnostic WebSocket gateway — you keep your carrier — with a native install on Vicidial, Asterisk, FreeSWITCH, and Issabel via one bash command in about five minutes. Its 23 analytics reports and queryable per-detection log give you the per-DID and per-trap visibility the monitoring step above actually requires. For the full picture on accuracy and why timing-based detection falls short, see the Vicidial AMD guide, the drawbacks of stock AMD, and the State of AMD 2026.

FAQ