Before your agent says a single word, the carrier has already made a decision.
Three analytics companies — Hiya, TNS/Call Guardian, and First Orion — power the spam-detection systems baked into AT&T, T-Mobile, and Verizon. When a call arrives, they score the originating number in real time and push a label to the recipient's phone before the first ring. That label might say "Spam Risk," "Scam Likely," or nothing at all — but if it says anything, the answer rate drops by 60–80%.
Your AI agent may have beautiful conversation design. It may disclose itself perfectly, handle escalations smoothly, and stay within scope. None of that matters for the calls that never get answered because the number was already tagged.
The scoring engine you're invisible to
The analytics layer runs entirely outside your infrastructure. You have no API access to it. You can't query a number's current score. You can't submit a support ticket when your number gets flagged. You find out the same way your users do: the call rings, no one picks up, and eventually you notice your answer rate is 12%.
The systems work by building behavioral profiles on every number that places calls. A number with a long history of bidirectional traffic, moderate volume, and good answer rates looks like a legitimate business line. A number that appeared two weeks ago, has placed 4,000 calls, received zero, and has a 15% answer rate looks like an outbound dialer running a scam campaign.
AI agents score exactly like scam campaigns across every dimension.
The five signals your agent trips
How Carrier Analytics Score a "Spam" Number
estimated weight of each signal in reputation scoring — hover to inspect
Volume spike (28% of score): Scammers provision a number, blast calls for 48–72 hours until it gets flagged, then provision another. This is also exactly what an AI agent does on launch day when a team spins up a new workflow. The number is cold. Call volume goes from zero to hundreds overnight. The scoring algorithm treats this as a near-certain indicator of fraud.
Low answer rate (24%): A legitimate business number that places 1,000 calls over a month and has a 45% answer rate looks healthy. An agent that places 1,000 calls in two days and has a 14% answer rate looks like a robocall campaign hitting numbers that mostly don't pick up. The timeline matters as much as the absolute numbers.
No inbound call history (21%): Real businesses receive calls. Customers call back. Support lines have inbound volume. A number that has placed thousands of outbound calls and received zero inbound calls is a statistical anomaly that correlates almost exclusively with automated outbound campaigns. Most AI agents never receive an inbound call on their outbound number.
Short average call duration (16%): Scam calls are brief — a few seconds of silence, a recorded message, then hang-up when the call connects or goes to voicemail. AI agents that hit voicemail frequently (which is most of them) produce the exact same duration signature. A portfolio of calls averaging 18 seconds looks like a robocall farm.
Repeated audio fingerprint (11%): TTS engines produce consistent acoustic signatures — same voice model, same speech cadence, same EQ profile on every call. Human voices vary. Carrier analytics systems can cluster calls by audio fingerprint, and a number that places calls with an identical synthetic voice across thousands of calls is a reliable indicator of automation.
Hit all five, and your number is flagged within days of launch. Sometimes within hours.
What your logs tell you vs. what's actually happening
When a call reaches a phone labeled "Spam Risk," most recipients decline it or let it ring to voicemail. Your call log records this as "No Answer" or "Voicemail." Nothing in the standard telephony data tells you the call was presented with a spam label.
This is why so many teams debug answer rates by A/B testing conversation scripts, adjusting call timing, and rewriting opening lines — and see no improvement. The conversation never starts. The problem is upstream.
The only reliable way to diagnose carrier flagging is to test your own number on a separate device with a consumer carrier — place a call from the number you're using for your agent and see what label appears on the receiving end. If you see a spam indicator, the number is already in the scoring system's bad category.
What you can actually do about it
Warm new numbers before deploying agents on them. A number with 30 days of modest, bidirectional traffic — a few dozen calls per day, some inbound, reasonable answer rates — starts from a much better baseline than a cold number hit with a spike on day one. Treat number provisioning as infrastructure with a lead time, not a same-day operation.
Vary calling patterns. Strictly uniform call volume (e.g. exactly 200 calls every hour, every day) is a signature the scoring systems recognize. Human calling patterns have variance. Add randomness to call schedules, distribute volume across time zones, and pause campaigns during off-hours. This is both better for answer rates and harder for analytics to fingerprint.
Use separate numbers for different workflows. A number used for appointment reminders has different call patterns than one used for outbound lead qualification. Mixing them inflates the total volume on a single number and muddies the behavioral profile. Segment by use case and size the volume to each number appropriately.
Monitor for degradation. Number reputation erodes gradually, then suddenly. Track answer rate by number over time. A number whose answer rate was 38% three months ago and is now 14% has almost certainly been flagged. Rotate it out before the whole workflow degrades.
Register for CNAM and ensure STIR/SHAKEN A-level attestation. CNAM (Caller Name Display) shows a business name rather than the raw number, which increases answer rates independent of spam scoring. STIR/SHAKEN A-level attestation — meaning your carrier can verify the call originated from your authenticated account — is weighted positively by some scoring systems and required by FCC regulations for all US carriers. A call with full attestation from a reputable carrier is treated differently than an unattested call from a wholesale VoIP provider.
The infrastructure tax on every AI calling deployment
Every AI voice agent has an invisible cost that doesn't appear in your telephony bill: the ongoing operational work of maintaining number reputation. It's not glamorous, and it's not in any LLM provider's documentation, but it determines whether your agent's conversation ever gets heard.
The gap between "my agent works in testing" and "my agent reliably reaches people at scale" is mostly this. Test environments use the same number a handful of times and never hit a carrier threshold. Production environments place thousands of calls on numbers that have no established history, and the scoring systems react accordingly.
Building this infrastructure in-house — number warm-up workflows, reputation monitoring, carrier relationship management, rotation automation — is typically several months of engineering work that has nothing to do with what your agent actually does. It's the plumbing that has to exist before the agent can operate.
The teams that figure this out ship reliable AI calling. The ones that don't spend months debugging answer rates and never find the real cause.