Two things are true about AI outbound calling at the same time. Placing a call has never been cheaper. And on a badly run program, every call you place makes the next one less likely to be answered.
Answer rates for AI outbound agents depend on the network side of the call as much as on the agent itself. The main factors are the caller ID the recipient sees, the number’s STIR/SHAKEN attestation, how call-analytics engines have scored that number, and what happens in the first second after the recipient picks up. A better model won’t fix any of those. Most teams discover this after they’ve spent a quarter tuning prompts while their numbers were quietly being labeled “Spam Likely.”
This post takes the operations view: problem, cause, fix. It’s written for voice AI platforms running outbound at volume. If you’re still designing the connectivity layer, our SIP trunking for AI agents hub covers the architecture; this post picks up once calls are flowing.
The problem: agent quality doesn’t show up in answer rates
The symptom is a program where conversations improve and results don’t. Completion rates on answered calls go up. Answer rates go down, or stay flat while volume grows.
That’s because an answer rate measures a decision made before the agent speaks. The recipient sees a number, possibly a label, and decides. The agent’s quality only matters for the calls that survive that decision. So the ceiling on your program is set by the network and the number, and the agent works inside it.
There are five causes worth checking, roughly in the order they tend to bite.
Cause 1: weak attestation on US calls
In the US, calls carry a STIR/SHAKEN attestation that tells the terminating carrier how much the originating provider vouches for the caller’s right to use the number. Lower attestation doesn’t block a call by itself, but it feeds the systems that decide how to label it.
The levels, briefly: “A” (full attestation) means the originating provider knows the customer and knows they’re authorized to use that number. “B” means the provider knows the customer but can’t confirm the number. “C” means the provider is passing the call along without knowing either. Our STIR/SHAKEN compliance guide goes through the framework in full.
AI platforms often end up with B or C without realizing it. A common cause is presenting numbers that weren’t provisioned through the provider signing the call: a client’s own number, a number ported elsewhere, or numbers bought from one carrier and sent out through another.
Fix: present numbers that are provisioned through, or verifiably authorized with, the provider originating your calls. If clients want their own numbers displayed, document the authorization and confirm with your carrier how those calls will be attested. Check attestation on sample calls rather than assuming it.
Cause 2: call-analytics labeling
Most US mobile users see caller ID filtered through call-analytics engines run by or for their carrier. Those engines score numbers on behavior, and an AI outbound program is close to a textbook example of the behavior they flag.
The signals are the obvious ones. High call volume from a single number. A large share of calls lasting a few seconds. Low answer rates. Recipients blocking or reporting the number. An AI agent campaign that pushes thousands of calls a day through a small pool of numbers hits all of them at once.
Labels are sticky. Once a number is flagged, answer rates drop, which produces more short and unanswered calls, which reinforces the flag. The loop runs in the wrong direction on its own.
Fix: register your outbound numbers with the analytics providers. The Free Caller Registry lets you register numbers once for the major US analytics engines. Then manage behavior at the number level: cap daily calls per number, spread volume over more numbers in proportion to your traffic, and monitor labels on a sample of numbers weekly. And don’t rotate numbers to escape labels. That trades one flagged number for a pool of them and trains the engines that your whole range is suspect.
Cause 3: the silent first second
This is the cause most specific to AI, and the one teams least expect. The recipient answers, says “Hello?”, hears nothing, and hangs up.
Silence at the start of a call happens for mechanical reasons. Answering-machine detection holds the audio while it classifies the answer. The media path takes a moment to establish. The agent pipeline waits for speech-end detection on the recipient’s greeting, then runs speech recognition, the model, and speech synthesis before saying anything. Each piece is small. Together they can add up to a gap that feels, to a human, like a robocall that didn’t connect.
The damage is double. You lose the conversation. And the call ends in a couple of seconds, which is one more short-duration data point for the analytics engines in Cause 2.
Fix: treat time-to-first-word as a core metric, measured from the 200 OK, not from when your pipeline receives audio. Place media servers close to where your calls terminate. Prepare the opening line before the call is answered so there’s no model latency on the first turn. Decide whether you need AMD on every call; if you do, test how long it holds audio and whether a brief spoken opener during classification performs better than silence. Watch the share of answered calls lasting under six seconds; if it’s high, silence is a likely culprit.
Cause 4: caller ID that doesn’t arrive
Outside the US, or on international campaigns, the attestation framework doesn’t apply the same way, but caller ID still decides the answer. If your route strips or replaces your number, recipients see “private” or an unfamiliar foreign number and ignore it.
That’s a route problem, not an agent problem. AI platforms that buy termination on price alone often land on routes that don’t pass CLI, especially to mobile networks abroad. Our explainer on CLI versus non-CLI routes explains what happens to caller ID on cheaper paths and how to test for it.
Fix: put outbound agent traffic on CLI routes for any destination where the recipient’s answer depends on the displayed number, which for outbound is nearly all of them. Present local numbers in your calling countries where you can; provisioning DID numbers in-market gives recipients a familiar format and a number they can call back.
Cause 5: consent and calling discipline
Answer rates also fall when the list is wrong: people who didn’t expect a call, called at the wrong hour, too many times. That’s partly a compliance question and partly plain operations.
In the US, the FCC ruled in February 2024 that AI-generated voices count as “artificial” voices under the Telephone Consumer Protection Act, so the consent rules for artificial or prerecorded voice calls apply to AI agents. Other countries have their own rules. We’re not giving legal advice here, and you should get it for your markets. Operationally, though, the same disciplines that keep you compliant also keep answer rates up: call people who agreed to be called, respect local calling hours, cap attempts per contact, and honor opt-outs immediately.
Fix: build consent status, calling-hour windows and attempt caps into the dialer layer so the agent can’t place a call that breaks them. Treat complaints as an operational alarm, because they’re also feeding the labeling engines.
An answer-rate dashboard that finds problems early
The five causes above are easier to spot with metrics at the right level. Campaign-level averages hide number-level trouble.
Track, per outbound number per day: calls placed, answer rate, share of answered calls under six seconds, average duration, and complaint or block reports where you have them. Track, per destination and route: answer rate, post-dial delay and duration patterns, using the approach in our guide to reading ASR, ACD and PDD. Add time-to-first-word from your own pipeline.
Then set a few simple alerts. A number whose answer rate falls sharply week over week. A rise in the short-call share on any route. Time-to-first-word crossing whatever threshold your testing shows starts losing callers. Each of those points to one of the causes above.
Callbacks are part of the answer-rate story
A missed call isn’t always a lost contact. Some recipients call the number back, and those callbacks are often the most willing conversations a campaign produces. Programs that ignore them leave that value on the table.
Make sure every number you present can receive calls, and that an inbound call to it reaches something useful: the same agent with context about the missed outbound attempt, a human queue, or at minimum a clear message about who called and why. A number that rings out or plays a generic error on callback teaches recipients that your calls aren’t worth returning, and some of them will report it instead.
Track callback volume per number alongside answer rate. A number with a modest answer rate but steady callbacks is performing better than the dashboard suggests. A number with no callbacks at all may be displaying incorrectly on the recipient’s side, which sends you back to Cause 4.
Where the carrier fits
Most of the fixes above are yours to make, but two depend on your carrier: whether routes deliver your caller ID unchanged, and how your US calls are attested. Ask both questions explicitly before you commit traffic.
For platforms building on IDT Express, voice termination covers 160+ countries on 1,000+ CLI-certified routes, which lets you put agent traffic on routes that carry your number through while keeping lower tiers for traffic where caller ID doesn’t matter. Pair that with numbers provisioned through the same carrier and the attestation and CLI questions get simpler to answer. For the broader picture of how termination works, our voice termination guide is the place to start.
Frequently asked questions
Why are my AI agent’s calls being labeled as spam?
Call-analytics engines label numbers based on behavior such as high volume per number, many short calls, low answer rates and complaints, all common in AI outbound programs. Registering numbers with analytics providers, capping calls per number and fixing short-call causes like opening silence all reduce labeling risk.
Does STIR/SHAKEN affect AI outbound calls?
Yes. In the US, calls with weaker attestation are more likely to be labeled or treated with suspicion. Presenting numbers that are provisioned through, or verifiably authorized with, the provider originating the call helps you get full attestation.
Do TCPA rules apply to AI voice agents?
In the US, the FCC ruled in February 2024 that AI-generated voices are “artificial” voices under the TCPA, so the consent rules for artificial or prerecorded voice calls apply. Get legal advice for your specific use case and markets.
Why do people hang up on AI calls in the first few seconds?
A common cause is silence after the call is answered, caused by answering-machine detection, media setup and the agent’s processing time. Measuring time-to-first-word from answer and preparing the opening line before the call connects both reduce early hang-ups.
Should I rotate phone numbers to avoid spam labels?
No. Rotating numbers to escape labels tends to spread the problem across your whole number range. Managing volume per number, registering numbers and fixing call behavior addresses the cause instead.
Build on CLI routes
Building an outbound voice agent? See how IDT Express SIP trunking connects AI calling platforms to CLI routes and in-market numbers.
