TCPA Voicemail Detection for Dialers: AMD Benchmarks 2026

Written by: Matt Beucler, CEO, Plura AI

Updated September 2026

Key Takeaways for Contact Center Leaders

• TCPA-compliant voicemail detection depends on AMD accuracy, agent availability controls, and carrier infrastructure working together to keep abandonment under 3% and avoid dead air.
• AI-powered AMD typically reaches 95-99% overall accuracy with 1-3% false positives, compared to 60-75% accuracy and 18-25% false positives in legacy rule-based systems.³
• The 2-second rule and 3% abandonment cap sit at the center of day-to-day compliance. AMD false positives count as abandoned calls even when the dialer labels them as machines.
• Conservative handling of uncertain AMD results and continuous abandonment monitoring at the carrier layer are essential to staying within the safe harbor.
• Plura AI’s AI Predictive Dialer uses FCC-licensed carrier infrastructure, real-time DNC scrubbing, and immutable consent logging to support compliance while maintaining throughput.

The Regulatory Framework for Abandoned Calls and Voicemail Detection

Two federal frameworks govern abandoned calls and voicemail detection for outbound dialers. Operators and their counsel should consult the primary regulations directly.²

The FTC Telemarketing Sales Rule (TSR), codified at 16 CFR 310.4, limits abandoned calls to no more than 3% of calls answered by a live person, measured per campaign over each successive 30-day period.² A call is abandoned if no live representative connects within 2 seconds of the person’s completed greeting. Because these two conditions define the cap, the safe harbor also requires a minimum ring time, a prompt prerecorded identification message when no agent is available, and documented recordkeeping. Missing any one condition forfeits the safe harbor entirely.

The FCC’s TCPA rules, codified at 47 CFR § 64.1200, define an abandoned call as one where a person answers and no live agent connects within 2 seconds of the completed greeting. Machine answers do not count toward the 3% cap, so AMD accuracy becomes a direct compliance variable. Statutory damages under 47 U.S.C. § 227 run $500 to $1,500 per violation. TCPA class action settlements averaged $6.6M in 2023, and a campaign of 10,000 non-compliant calls can generate $5M to $15M in potential exposure.³

The FCC’s February 2024 declaratory ruling classified AI-generated voices as artificial voices under the TCPA, applying the same consent standard as traditional robocalls. A related ruling from November 2022, FCC 22-85, held that ringless voicemail deposits to wireless numbers are “calls” under the TCPA, subject to the same consent requirements. Operators should consult qualified counsel on how these rulings apply to their specific programs.

Accuracy Benchmarks: What Good AMD Looks Like

AMD performance is measured across four metrics, and each one maps to a specific compliance or operational outcome.

Metric	Benchmark	Compliance Impact
Overall accuracy	95-99% (AI-driven) vs. 60-75% (legacy rule-based)	Higher accuracy means fewer misclassified live humans and fewer hidden abandoned calls
False positives (live human classified as machine)	Default Asterisk: 18-25%; manually optimized: 8-15%; AI-powered: 1-3%	Each false positive is a dropped live call and a potential abandoned-call violation
False negatives (machine classified as live human)	Target below 2% for AI-powered systems	Voicemails routed to agents waste talk time and distort pacing
Detection latency	500ms-2s; total_analysis_time should stay under 4000ms on predictive dialers	Latency beyond 2 seconds creates dead air that consumers interpret as a robocall

The structural ceiling for rules-based AMD is well documented. On Asterisk’s stock AMD application, false positive rates of 15-25% are typical in production. AI classifiers trained on greeting audio now run at 1-3% false positives, an order of magnitude better than the silence-and-duration heuristics used by rules-based systems.

The compliance blind spot created by false positives is significant. A mid-size operation dialing 15,000 numbers daily with a 38% live-answer rate and a 20% AMD false positive rate would misroute over 1,100 live contacts daily as machines, and those calls would not appear in the abandonment report. Plaintiff’s firms increasingly request raw call logs in TCPA discovery, and a high machine-classification rate on a list with a normal live-answer profile is a recognized red flag.

See Plura in a live dialing walkthrough to review carrier-level AMD enforcement and real-time compliance monitoring.

The 2-Second Rule and Abandonment Rate

The 2-second rule drives how predictive dialing must behave in production. Under 47 CFR § 64.1200, a call is abandoned when a person answers and no live agent connects within 2 seconds of the completed greeting. AMD latency and AMD accuracy both influence whether that 2-second window is met.

The relationship between AMD false positives and abandonment rate is direct. When AMD misclassifies a live person as a machine, the dialer drops the call. That dropped call meets the FCC’s abandoned-call definition regardless of how the dialer dispositioned it internally. AMD false positives count as abandoned calls under FCC rules when a live person answers and no agent connects.

The operational target for well-run predictive dialing campaigns is 1% to 2% abandonment. Many compliance teams use a 2.5% operational governor as a buffer against intra-day variability that could push a campaign over the legal ceiling. Abandonment should be tracked continuously at the carrier and platform layer, with alerting as the rate approaches 3%. Teams cannot reliably reconstruct this metric from logs after the fact.

Conservative Handling of Uncertain AMD Results

AMD algorithms produce three result types: LIVE, MACHINE, and UNCERTAIN (or NOTSURE in Asterisk terminology). The way a platform handles uncertain results often separates compliant operations from risky ones.

The recommended decision sequence for AMD result handling is:

1. LIVE: Route to an agent immediately.
2. MACHINE: Log the disposition and proceed to the next call.
3. UNCERTAIN: Treat as human and route to an agent.

Preserving live connections is safer than misclassifying humans as machines. Every false positive is a potential abandoned call and a wasted lead you paid to acquire. When NOTSURE classifications are too high, raising total_analysis_time to 5500ms can help; when AMD is taking too long, lowering total_analysis_time to 4000ms improves pacing. On a predictive dialer, every extra millisecond of AMD analysis time affects pacing and abandonment rates, so total_analysis_time becomes a critical parameter.

Dialer Configuration for Predictive and Power Modes

AMD settings differ materially by dialing mode. The configuration that supports compliance on a predictive dialer differs from the configuration appropriate for a power dialer.

Predictive dialer configuration (for platforms including AI Predictive Dialer deployments and VICIdial-based operations):

• Agent availability gating: set Available Only Tally = Y so the system dials only when agents are actually idle.
• Auto Dial Level: start at 1.5 for adaptive modes and adjust based on real performance data from logs.
• ADAPT_HARD_LIMIT: set the drop rate ceiling to 3% for TCPA compliance support.
• AMD total_analysis_time: keep under 4000ms to protect pacing, and raise to 5500ms if NOTSURE classifications are too high.
• Conservative uncertainty handling: route all NOTSURE results to agents instead of voicemail disposition.
• Codec: use G.711 pass-through; G.729 compression discards waveform detail and raises false positive rates.

Power dialer configuration

• 1:1 dialing keeps a live rep on the line at pickup, which structurally eliminates abandonment.
• The absence of a silence gap at pickup removes the dead-air signature that carriers often flag as spam.
• AMD becomes optional and can be disabled entirely for high-value leads without creating abandonment risk.
• Progressive dialing without AMD removes both misclassification risk and abandoned-call risk, at the cost of throughput.

The choice between dialing modes is an economics and compliance trade-off. Predictive dialers maximize throughput but require careful AMD tuning and agent-availability gating to stay within the 3% safe harbor. Power dialers structurally avoid abandonment but deliver lower throughput per agent hour.

Consent, Prerecorded Messages, and AMD

TCPA consent requirements sit alongside AMD configuration. Prior express written consent is described as required before using an autodialer or prerecorded voice to contact a mobile number for marketing purposes; prior means before the call, express means affirmative, and written includes electronic records but involves specific language disclosures.

The FCC’s November 2022 Declaratory Ruling (FCC 22-85) held that ringless voicemail deposits are “calls” under the TCPA, subject to the same consent framework as other prerecorded-voice calls. The FCC’s February 2024 declaratory ruling classified AI-generated voices as artificial voices under the TCPA.

On February 25, 2026, the Fifth Circuit ruled in Bradford v. Sovereign Pest Control of TX, Inc. that prior express consent does not need to be written under the TCPA’s statutory text, which conflicts with FCC implementing regulations at 47 CFR § 64.1200(a)(1). This ruling is currently circuit-specific. Operators should consult qualified counsel on consent strategy for their jurisdiction and program type.

Plura Platform Capabilities That Support Compliance Operations

To operationalize these controls, the dialing platform itself must support compliance at the infrastructure level. Plura AI is an FCC-licensed carrier. Voice originates on Plura’s domestic infrastructure, not a third-party CPaaS, so branded caller ID is issued at the carrier level and enforcement occurs at origination. Plura provides infrastructure to support compliance; customers remain responsible for their own regulatory obligations and consent programs.

Features of Plura’s AI Predictive Dialer that support compliance operations include:

• Real-time DNC scrubbing against federal and state registries before every dial
• Immutable consent logging with timestamps and IP addresses
• Automated quiet-hours enforcement via time-zone detection
• Audit-ready exports in one click
• SOC 2, HIPAA, ISO certification, GDPR, and SHAKEN/STIR caller ID verification¹
• FCC-licensed carrier with branded caller ID issued at the carrier level
• Dynamic pacing, list management, timezone logic, and answer rate management built into the dialer

Integrations with CRM platforms, validation services, and attribution tools allow compliance data to flow into the systems operators already use. Abandonment tracking runs continuously at the carrier and platform layer, with alerting as rates approach the 3% threshold.

Walk through Plura’s compliance engine in a live session to see how AMD configuration and carrier-level enforcement work together.

Common Compliance Pitfalls in Dialer Operations

1. Using AMD to justify prerecorded messages. AMD does not replace consent. Prior express written consent remains a separate requirement for prerecorded marketing content regardless of AMD accuracy or classification result.
2. Ignoring state laws. State mini-TCPAs in Florida, Oklahoma, Washington, and Maryland add stricter requirements on top of federal rules, including broader autodialer definitions and explicit consent expectations. Counsel can advise on state-specific obligations.
3. Setting too-aggressive AMD thresholds. Aggressive AMD maximizes agent talk time, but every false positive throws away an answered call from a lead you paid to acquire and creates TCPA exposure.
4. Not monitoring abandonment rate continuously. Abandonment tracking should be calculated continuously at the carrier and platform layer, with alerting or automatic throttling as the rate approaches the 3% safe harbor.
5. Trusting summary reports over raw logs. Plaintiff’s firms request raw call logs in TCPA discovery. Sampling AMD-dispositioned call recordings weekly helps verify classification accuracy independently of summary metrics.

FAQ

What is the 2-second rule in TCPA?

Under 47 CFR § 64.1200, an abandoned call occurs when a person answers and no live agent connects within 2 seconds of the completed greeting. AMD latency and accuracy both affect whether that window is met. If AMD takes too long to classify a pickup, the 2-second window can expire before an agent connects, which creates an abandoned call even when an agent was available. Consult the regulation and qualified counsel for how this applies to your specific dialing program.

How does AMD affect abandonment rate?

AMD false positives, where a live human is misclassified as a machine, cause the dialer to drop a live call. That dropped call meets the FCC’s abandoned-call definition regardless of how the dialer logged the disposition. Because machine answers do not count toward the 3% cap, accurate AMD is critical. A high false positive rate creates hidden abandoned calls that do not appear in standard abandonment reports but remain visible in raw call logs.

What is the legal abandonment cap and operational target?

The legal cap under both the FTC TSR and FCC TCPA rules is 3% per campaign over a 30-day period. The operational target for well-run predictive dialing campaigns is 1% to 2%, with many compliance teams using a 2.5% operational governor as a buffer against intra-day variability. The safe harbor requires four simultaneous conditions: the 3% cap, a minimum ring time of 15 seconds, a prerecorded identification message when no agent connects, and documented proof of compliance.

How do you reduce false positives in AMD?

The highest-impact tuning levers are initial_silence and min_word_length. Raising initial_silence tolerates hesitant human responses, and lowering min_word_length reduces the chance that a short human response is classified as a machine. Shortening the analysis window reduces the dead-air duration that causes humans to hang up. However, below an 8% false positive rate, the error becomes structural in rules-based systems, which is why AI-powered AMD trained on greeting audio is needed to reach 1-3% false positives compared to 18-25% for default Asterisk configurations. Tune one parameter at a time, run each change for approximately 24 hours, and sample machine-classified recordings before and after each adjustment.

How should uncertain AMD results be handled?

Teams should treat uncertain results as live humans and route those calls to an agent. Preserving live connections is safer than misclassifying humans as machines. Every false positive is a potential abandoned call and a wasted lead. If NOTSURE classifications are too frequent, raising total_analysis_time is a better response than defaulting uncertain results to voicemail disposition.

What dialing mode supports TCPA compliance goals?

Predictive dialers maximize throughput but require careful AMD tuning, agent-availability gating, and continuous abandonment monitoring to stay within the 3% safe harbor. Power dialers operate at 1:1 and keep a live rep on the line at pickup, which structurally eliminates abandonment but lowers throughput per agent hour. For high-value leads, progressive dialing without AMD removes both misclassification risk and abandoned-call risk entirely. The right mode depends on list quality, agent count, and throughput requirements.

Does AMD replace TCPA consent requirements?

AMD does not replace TCPA consent requirements. Prior express written consent is described as required for prerecorded marketing messages regardless of AMD accuracy or how the call is classified. The FCC’s November 2022 Declaratory Ruling (FCC 22-85) confirmed that ringless voicemail deposits are “calls” under the TCPA, subject to the same consent framework. Qualified counsel can advise on consent programs.

How accurate is voicemail detection in 2026?

AI-driven AMD typically targets 95-99% overall accuracy with false positive rates of 1-3% and false negative rates below 2%. Legacy rule-based systems using energy thresholds and silence heuristics, such as default Asterisk AMD, often stall at 60-75% overall accuracy with false positive rates of 18-25% in production. Manually optimized rule-based settings can reach 8-15% false positives but do not reach AI-powered accuracy levels because human and machine greetings overlap at the acoustic decision boundary.

Conclusion and Next Steps for Dialer Teams

TCPA-compliant voicemail detection functions as a property of the entire dialing system. AMD accuracy, agent availability gating, carrier infrastructure, abandonment tracking, and consent logging must work together. A false positive rate of 20% on a 15,000-dial-per-day operation creates over 1,100 hidden abandoned calls daily, and those calls do not appear in standard reports even though they are discoverable in litigation. The 2-second rule, the 3% safe harbor, and conservative handling of uncertain AMD results form the core operational controls for a defensible predictive dialing program.

Plura’s AI Predictive Dialer enforces controls at the carrier level. As an FCC-licensed carrier, Plura issues branded caller ID at origination, runs real-time DNC scrubbing before every dial, logs consent records with timestamps and IP addresses, and exports audit-ready reports in one click. The platform supports SOC 2, HIPAA, ISO certification, GDPR, and SHAKEN/STIR caller ID verification.

Compare plans and rates side by side on Plura’s pricing page. Run your numbers with Plura’s ROI calculator to estimate savings in real time.

Schedule a live Plura dialer review to see carrier-level AMD enforcement, abandonment monitoring, and audit-ready infrastructure in action.

---

Read Next

• Voicemail Detection Best Practices for AI Predictive Dialers
• Voicemail Detection Accuracy Benchmarks for Outbound Teams
• Predictive Dialer Voicemail Detection: How AMD Works
• Voicemail Detection Pricing: AMD Costs Compared
• Voicemail Detection Alternatives for Outbound Leaders