AMDY.IO
Start free
← All articles
Product UpdatesAug 30, 2026 12 min read

Outbound Sales Call Performance: Why AMD Accuracy Is the Single Biggest Controllable Variable

Why One Detection Layer Decides Everything

If you run outbound sales calls, you already know the levers you can pull: script quality, list hygiene, agent training, call timing. You work through all of them and still watch live-connect rates stall. The culprit is usually sitting quietly in your dialer config, ignored because it has always been there — your answering machine detection layer.

AMD accuracy is the single biggest controllable variable in outbound call performance. Not because it's the most interesting setting to tune, but because every other optimization you make gets filtered through it first. A great agent can't save a call that was already dropped on a voicemail false positive. A clean list can't recover a prospect who was classified as MACHINE while they were actively watching their phone screen.

This article breaks down why AMD accuracy has such an outsized effect on outbound performance, what the accuracy gap between legacy and modern detection actually costs, and what changed in 2026 that makes this more urgent than it was a year ago.

What AMD Actually Does in an Outbound Dialer

When your predictive dialer places a call, it needs to know within the first few seconds whether a live person answered or whether it hit voicemail. That classification determines everything that follows: connect to an agent, skip it, drop a pre-recorded message, or move on.

Legacy AMD systems — including the one built into stock Asterisk — make that call by analyzing silence patterns. The logic dates to 2003: human speech has a certain cadence, voicemail greetings tend to run longer without interruption. The system listens, measures, and guesses.

The problem is that it's still guessing. The vicidial.org community has documented the ceiling for stock Asterisk AMD at 75 to 80% accuracy, even after manual tuning. That means somewhere between one in five and one in four classifications is wrong on any given batch of calls.

Modern ML-based AMD works at a different level. Instead of silence thresholds, it analyzes audio fingerprints in real time — comparing the acoustic signature of the answer against a trained model built on thousands of call patterns. Classification happens at the signal level, not the silence level, which is why the accuracy ceiling is fundamentally higher.

The Real Cost of a 20% Error Rate

A 20 to 25% misclassification rate sounds like a technical metric. In practice, it shows up in your revenue numbers in three distinct ways.

False positives on live answers. When AMD classifies a live human as MACHINE, the call disconnects before an agent connects. The prospect hears dead air or a click. They don't call back. You've burned a dial, a list record, and whatever goodwill came with it. At scale, this isn't an edge case — it's a predictable, recurring drain on your connect rate.

False negatives on voicemail. When AMD classifies a voicemail as HUMAN, an agent gets connected to a recording. They sit through the greeting, realize what happened, and hang up or fumble through a manual disconnect. Agent time is wasted, the dial counts against your daily capacity, and no message gets left because the voicemail drop logic never triggered.

Ghost calls. A ghost call is what a prospect experiences when AMD fires a false positive and the call drops before anyone speaks. They see a missed call from an unknown number. Sometimes they call back; usually they don't. Repeated ghost calls from the same number accelerate spam labeling, which compounds the damage across your entire outbound program.

Each of these failure modes is a direct function of AMD accuracy. Improve the accuracy, and all three shrink proportionally.

Why 2026 Introduced a New Classification Problem

The accuracy gap between legacy and modern AMD was already a real performance issue before this year. What changed in 2026 is that a third call state entered the picture — and legacy AMD has no way to handle it correctly.

iOS 26 introduced on-device Call Screening that intercepts outbound calls before a human voice is ever heard. When a prospect's iPhone runs Call Screening, the caller hears a prompt from the screening system, not the prospect. The prospect sees the call on their screen and decides whether to accept.

Legacy AMD hears that prompt and classifies it as MACHINE. The dialer hangs up. The prospect — who was actively watching the screen and considering whether to pick up — never gets the chance. From their perspective, the caller gave up immediately. From the call center's perspective, a live prospect was silently discarded.

This isn't a hypothetical edge case. iOS 26 Call Screening is estimated to affect around 150 million U.S. iPhone users. Google Call Screen, Samsung Bixby screening, T-Mobile Scam Shield, AT&T ActiveArmor, and Verizon Call Filter create the same interception behavior across Android and carrier-level filtering. The combined footprint of these systems means a meaningful share of your outbound list is now unreachable by any AMD that only classifies HUMAN or MACHINE.

Three Classifications Instead of Two

The right response to call screening interception isn't to classify it as MACHINE and move on. It's to recognize it as a distinct state that requires a different action.

AMDY.IO built a third classification called CALLGUARD specifically for this. When a call is intercepted by a screening system, CALLGUARD flags it separately from both HUMAN and MACHINE — giving your dialer logic a third path: retry the call, reroute it to a different number or time slot, or leave a message designed for screening prompts.

Without that third path, you're not just losing the call. You're losing a prospect who was engaged enough to look at their phone. That's a qualitatively different loss than a voicemail false positive.

CALLGUARD is a live production feature, not a roadmap item. It currently detects iOS 26 Call Screening, Google Call Screen, Samsung Bixby screening, T-Mobile Scam Shield, AT&T ActiveArmor, and Verizon Call Filter. No documented competitor has published a named, production-ready equivalent for this classification type.

The Accuracy Gap in Numbers

The performance difference between 75 to 80% accuracy and 99% accuracy isn't linear. Because false positives compound across a dialer session, the gap in live-connect rates is wider than the raw numbers suggest.

Consider a straightforward scenario: 1,000 dials in a session, 400 reaching a live person and 600 hitting voicemail. A legacy AMD at 78% accuracy will misclassify roughly 88 of those 400 live answers as MACHINE, hanging up before an agent connects. It will also misclassify roughly 132 voicemails as HUMAN, wasting agent time on dead connections.

At 99% accuracy, the same session produces 4 false positives on live answers and 6 false negatives on voicemail. The difference is 84 recovered live connections per 1,000 dials, plus the agent time reclaimed from false negatives.

At any reasonable conversion rate, those 84 recovered connections represent real pipeline. The ROI calculator at amdy.io lets you run this math against your actual call volume and conversion numbers to see what the accuracy gap costs your operation specifically.

Why the Fix Doesn't Require a Platform Migration

The obvious objection to upgrading AMD accuracy is that better detection tools usually come bundled inside platforms that require abandoning your existing stack. For most of the market, that's true.

MightyCall claims 97% AMD accuracy but requires full platform adoption. For a team of 10 or more agents, its per-user pricing at $25 to $45 per user per month exceeds a flat-rate alternative — and it doesn't support existing ViciDial or Asterisk installations. NobelBiz OMNI+ bundles AMD inside an enterprise omnichannel platform with custom pricing and multi-year commitments. Genesys Cloud CX runs $75 to $240 per user per month and requires a full migration.

AMDY.IO is structured differently. It's a standalone detection layer that drops into your existing dialer via a one-line install, working with ViciDial, Asterisk, FreePBX, Issabel, GoAutoDial, and 3CX. For any SIP-based dialer not on that list, a WebSocket API handles the integration. Your agents, your IVR, your routing logic, your scripts — none of that changes. Only the AMD layer is replaced.

Pricing is $79 per month flat with unlimited concurrent channels and no per-seat or per-minute fees. A 14-day free trial is available with no setup fees.

What Outbound Teams Actually Experience After Upgrading AMD

The operational changes that follow an AMD accuracy improvement tend to cluster around three areas.

Agent utilization goes up. When false negatives drop, agents stop sitting through voicemail greetings on calls that should have been skipped. That time flows back into the session as available capacity for live conversations.

Ghost calls drop. False positives on live answers are the direct source of ghost calls. Fewer false positives means fewer unexplained hang-ups from your number, which slows the accumulation of spam labels and keeps your caller ID cleaner over time.

Screening interceptions become recoverable. With CALLGUARD flagging screened calls instead of discarding them as MACHINE, your team has options. A retry at a different time, a reroute to a different number, or a message crafted for the screening prompt all produce better outcomes than a silent hang-up.

None of this requires retraining agents or rebuilding workflows. These are downstream effects of a more accurate detection layer operating at the same point in the call flow where the old one was.

Choosing the Right AMD Layer for Your Stack

If your operation runs on ViciDial, Asterisk, FreePBX, Issabel, GoAutoDial, or 3CX, the integration path is a one-line install. If you're running a custom SIP-based dialer or building AMD into a cloud contact center product, the WebSocket API handles the connection without requiring a plugin.

Before choosing any AMD solution, these are the questions worth asking:

  • Does it classify call screening interceptions as a distinct state, or does it lump them with MACHINE?
  • Can it integrate with your existing dialer without a platform migration?
  • What is the published accuracy figure, and how was it measured?
  • Does the pricing model scale with call volume or stay flat?

For most small-to-mid outbound operations running open-source dialers, the answers point toward a standalone ML-based layer rather than a bundled enterprise platform. The performance difference is measurable, the integration is low-friction, and the cost of the accuracy gap compounds every day the legacy system stays in place.

Frequently Asked Questions

What is answering machine detection and why does it affect outbound sales calls?
Answering machine detection (AMD) is the system your dialer uses to classify whether a call was answered by a live person or a voicemail. In outbound sales calls, AMD accuracy directly controls how many live prospects reach an agent versus how many are incorrectly hung up on or misrouted. A 20 to 25% error rate on a legacy system translates directly into lost pipeline.

Why does stock Asterisk AMD have a 75 to 80% accuracy ceiling?
Stock Asterisk AMD uses silence pattern analysis to distinguish human speech from voicemail greetings. This approach was designed in 2003 and doesn't account for modern voicemail systems, call screening prompts, or the acoustic variety of current telephony. The vicidial.org community has confirmed the 75 to 80% ceiling even after manual tuning, because the underlying detection method has structural limits that tuning can't overcome.

What is CALLGUARD and why does it matter in 2026?
CALLGUARD is a third AMD classification that identifies calls intercepted by call screening systems such as iOS 26 Call Screening, Google Call Screen, Samsung Bixby, T-Mobile Scam Shield, AT&T ActiveArmor, and Verizon Call Filter. Legacy AMD classifies these interceptions as MACHINE and hangs up, losing a prospect who was actively looking at their phone. CALLGUARD flags them separately so your dialer can retry, reroute, or leave a targeted message instead.

Can I upgrade my AMD accuracy without replacing my existing dialer?
Yes. AMDY.IO integrates into ViciDial, Asterisk, FreePBX, Issabel, GoAutoDial, and 3CX via a one-line install. It replaces only the detection layer, leaving your existing dialer configuration, routing logic, and agent workflows intact. A WebSocket API is available for any SIP-based dialer not on that list.

How does flat-rate AMD pricing compare to per-user dialer platforms?
Per-user pricing models from platforms like MightyCall at $25 to $45 per user per month scale with headcount. For a team of 10 or more agents, that model exceeds a flat-rate alternative. AMDY.IO charges $79 per month regardless of concurrent channels or agent count, keeping costs predictable as call volume grows.

What is a ghost call and how does AMD accuracy affect it?
A ghost call is what a prospect experiences when AMD fires a false positive on a live answer and the call drops before anyone speaks. The prospect sees a missed call from an unknown number with no explanation. Repeated ghost calls from the same caller ID accelerate spam labeling. Improving AMD accuracy reduces false positives, which reduces ghost calls and slows the degradation of your caller ID reputation.

How quickly can I see the effect of improved AMD accuracy on live-connect rates?
The effect is immediate at the detection level. From the first session after integration, false positives and false negatives drop in proportion to the accuracy improvement. Live-connect rate changes are visible within the first few days of call volume at normal operating pace. The ROI calculator at amdy.io lets you estimate the revenue impact before committing to a full deployment.

The Variable Worth Fixing First

Outbound sales call performance has a long list of moving parts. Most of them are slow to change: list quality takes time to build, agent skill takes time to develop, and market conditions are outside your control entirely.

AMD accuracy is different. It sits at the front of every call, it affects every dial, and it can be upgraded without touching anything else in your stack. The gap between a 78% legacy detector and a 99% ML-based one isn't an abstract percentage — it's live prospects recovered, agent time reclaimed, and screening interceptions handled instead of discarded.

If you want to see what that gap costs your operation in concrete numbers, the ROI calculator at amdy.io is a practical starting point. A 14-day free trial is available with no setup fees if you'd rather measure the difference directly in your own call volume.