Why This Matters Now
Regulatory attention on voice AI is intensifying across all major jurisdictions:
United States
- FCC 2024 Declaratory Ruling brought AI-generated voices explicitly under TCPA "artificial or prerecorded voice" requirements
- FCC One-to-One Consent Rule vacated by 11th Circuit January 2025, but compliance infrastructure still expected
- Illinois BIPA covers voiceprints explicitly with $1,000-$5,000 per violation; 2024 amendment caps per-person damages
- CIPA litigation wave expanding third-party liability for cloud AI vendors under capability test
European Union
- EU AI Act Article 5 banned emotion recognition in workplace/education as of February 2, 2025
Emerging Markets
- Brazil LGPD requires opt-in consent with extraterritorial reach; Anatel mandates 0303 prefix for high-volume callers
- China PIPL requires separate consent for cross-border transfers; first extraterritorial enforcement case in 2024
- India DPDP Act full compliance deadline May 2027 with consent-centric regime
- UAE PDPL requires two-party consent for call recording with AED 50,000 telemarketing fines
When regulators or auditors examine your voice AI deployment, they're not checking boxes—they're looking for evidence that your compliance program actually works. The difference between "program exists" and "program works" is where most companies fail.
Part 1: United States Federal Requirements
1.1 TCPA Consent Framework
The Telephone Consumer Protection Act establishes three distinct consent tiers. Each has specific requirements, and voice AI deployments must correctly identify which tier applies to each call type.
Prior Express Written Consent (PEWC)
When Required: Telemarketing/advertising calls using ATDS or artificial/prerecorded voice to cell phones OR residential lines.
The FCC requires PEWC to contain nine elements (47 CFR §64.1200(f)(9)):
- Written Agreement — Preserved copy of exact form/webpage at time of consent capture
- Signature — Electronic signature compliant with E-SIGN Act (timestamp, IP, session ID)
- Telephone Number — Specific number to be called, captured at consent time
- Seller Identification — Name of specific seller authorized to call (one seller per consent post-2025 best practice)
- Technology Disclosure — Clear statement that calls will use ATDS and/or artificial/prerecorded voice
- AI/Artificial Voice Disclosure — Post-FCC 2024: explicit mention of AI-generated voice if used
- Not-a-Condition Disclosure — Statement that consent is not required as condition of purchase
- Clear and Conspicuous — Disclosure not buried in fine print; font size/placement documentation
- Topically Related — Call content logically related to website/context where consent obtained
Conversational AI-Specific PEWC Risks:
- Scope Creep: Large language models (LLMs) free-associate across product catalog, exceeding "logically related" content boundary. Testing shows ~5% failure rate under adversarial prompting.
- Hallucinated Consent: AI fabricates permission claims ("You previously agreed to...") creating evidence of willful violation.
- RAG-Triggered Marketing: Informational call becomes telemarketing when retrieval-augmented generation pulls promotional content under customer pressure phrases.
Prior Express Consent (PEC)
When Required: Non-telemarketing calls using ATDS or artificial/prerecorded voice to cell phones.
PEC is less formal than PEWC but still requires demonstrable consent:
- Number Provision — Consumer provided phone number to caller for related purpose
- Purpose Relation — Calls closely related to purpose for which number was provided
- No Contrary Instructions — Consumer did not indicate they don't want calls at that number
- Oral or Written — Can be verbal or written (unlike PEWC which must be written)
Conversational AI-Specific PEC Risk: LLM + RAG access creates ~5% failure rate where AI proposes promotional content despite constraints. Edge case: customer mentions cost concerns → AI retrieves discount programs → informational call becomes telemarketing → PEC insufficient, PEWC required.
Prior Express Invitation or Permission (PEIP)
When Required: Calls to numbers on the National Do-Not-Call Registry.
PEIP allows calling DNC-registered numbers if consumer specifically invited or gave permission to call, permission is in writing (but doesn't require all 9 PEWC elements), and consumer specifically requested information about goods/services.
Established Business Relationship (EBR) Alternative:
- Transaction EBR: 18 months from last transaction
- Inquiry EBR: 3 months from inquiry (90 days)
Part 2: US State Recording Consent Requirements
2.1 Two-Party (All-Party) Consent States
The following 11 states require consent from ALL parties to record a telephone conversation. Voice AI systems processing calls to/from residents of these states must obtain affirmative consent before recording, transcription, or AI analysis.
California — Penal Code §632 requires all-party consent for confidential communications; CIPA adds AI-specific layers. Penalties: $5,000/violation (CIPA); criminal misdemeanor.
Connecticut — Gen. Stat. §52-570d. Criminal: one-party; Civil: all-party. Treat as all-party to be safe. Penalties: Class D felony (criminal); civil damages.
Delaware — 11 Del. C. §2402 requires all-party consent. Penalties: Class G felony; civil damages.
Florida — Fla. Stat. §934.03 requires all-party consent; mini-TCPA with broader autodialer definition. Penalties: 3rd degree felony; $1,000+ civil damages.
Illinois — 720 ILCS 5/14-2 requires all-party consent for private conversations; BIPA adds biometric layer for voiceprints. Penalties: Class 4 felony; BIPA: $1K-$5K/violation.
Maryland — Md. Code §10-402 requires all-party consent with strict enforcement. Penalties: Felony; up to 5 years imprisonment.
Massachusetts — M.G.L. c.272 §99 bans "secret" recordings; all parties must be aware. Penalties: Felony; up to 5 years; $10K fine.
Montana — Mont. Code §45-8-213 requires all parties know recording is occurring (notification, not explicit consent). Penalties: Misdemeanor; civil damages.
Nevada — Nev. Rev. Stat. §200.620 requires all-party consent for telephone calls (one-party for in-person). Penalties: Category D felony; civil damages.
New Hampshire — N.H. Rev. Stat. §570-A:2 requires all-party consent. Penalties: Class B felony; civil damages.
Pennsylvania — 18 Pa.C.S. §5704 strictly requires all-party consent for any oral or electronic communication. Penalties: 3rd degree felony; civil damages.
Washington — RCW 9.73.030 requires all-party consent for private conversations; most calls presumed private. Penalties: Gross misdemeanor; civil damages.
2.2 One-Party Consent States
The remaining 39 states and D.C. follow federal one-party consent rules. One party to the conversation (which can be your company) may consent to recording without notifying others.
2.3 Interstate Call Rules
Critical: When parties are in different states, the stricter law generally applies.
For national voice AI operations: default to all-party consent disclosure at call start. This satisfies both one-party and all-party jurisdictions.
Part 3: California CIPA Deep Dive
California Invasion of Privacy Act applies if the call recipient is in California, regardless of company location. With 12% of the US population, any national operation has significant CIPA exposure.
3.1 CIPA Sections Relevant to Voice AI
Section 631 (Wiretapping) — Prohibits reading contents while in transit. Voice AI processes real-time = interception by definition; Cloud processing = vendor receives contents in transit.
Section 632 (Recording) — Prohibits recording confidential communications without consent. Transcripts = recordings; Vector embeddings = recording in different format; Training use = indefinite retention.
Section 632.7 (Cellular) — Prohibits interception of cellular communications (no confidentiality requirement). Covers most calls (cell prevalence); Lower plaintiff burden than 632; Most common claim basis.
Section 632(d) (Emotion Analysis) — Examination of truthfulness/emotions requires express written consent. Sentiment analysis often runs by default; AI naturally infers emotional context; Feature creep: "caller intent detection."
3.2 Third-Party Question: Extension vs. Capability Test
Two competing legal tests determine whether your AI vendor is a "third party" under CIPA:
Extension Test (defendant-friendly) — Vendor processes data only for your benefit = your extension. Outcome: No third-party disclosure; lower consent burden.
Capability Test (plaintiff-friendly) — Vendor has capability to use data for own benefit = third party. Outcome: Third-party disclosure required; higher consent burden.
Recent Case Law Favoring Capability Test:
- Javier v. Assurance IQ (N.D. Cal. 2023): Capability to use data = third party status
- Ambriz v. Google (N.D. Cal. 2025): AI system itself can be "person" under CIPA; motion to dismiss denied
- Yockey v. Salesforce (N.D. Cal. 2024): Could use data to train AI = third party status
- Turner v. Nuance (N.D. Cal. 2024): Fraud detection database = own purposes = third party
3.3 Vendor Contract Risk Clauses
"Improve our services" — CRITICAL RISK. Capability for own benefit = third party status under capability test.
"Aggregated data for analytics" — HIGH RISK. Depends on whether aggregation can be traced; likely third party.
"Train models" — CRITICAL RISK. Clear independent use; definite third party status.
"Act as service provider only" — LOW RISK. Agency positioning; supports extension test argument.
3.4 CIPA Penalty Stacking
Theoretical maximum exposure per call:
- CIPA 631 (wiretapping): $5,000
- CIPA 632.7 (cellular interception): $5,000
- CIPA 632(d) (emotion analysis): $5,000
- TCPA (if applicable): $1,500
Total per-call exposure: Up to $16,500
Part 4: Illinois BIPA Biometric Requirements
The Illinois Biometric Information Privacy Act explicitly covers voiceprints. Any voice AI system that creates voiceprints for speaker identification must comply with BIPA for Illinois residents.
4.1 BIPA Definition of Voiceprint
740 ILCS 14/10: "Biometric identifier" means a retina or iris scan, fingerprint, voiceprint, or scan of hand or face geometry.
A voiceprint is created when voice data is processed to identify a specific individual. This includes:
- Speaker verification ("Is this the account holder?")
- Speaker identification ("Who is speaking?")
- Voice enrollment ("Say these phrases to train the system")
4.2 BIPA Requirements
Section 15(a) — Written policy on retention and destruction, made publicly available. Evidence needed: Published policy with specific retention periods and destruction procedures.
Section 15(b) — Written notice of collection + purpose + retention period; Written release (consent). Evidence needed: Consent form with all required disclosures; Electronic signature with timestamp.
Section 15(c) — Prohibition on selling, leasing, or profiting from biometric data. Evidence needed: Vendor contracts prohibiting commercialization; No revenue from voice data.
Section 15(d) — No disclosure without consent (or narrow exceptions). Evidence needed: Data flow documentation; Third-party agreements.
Section 15(e) — Reasonable security measures. Evidence needed: Security documentation; Industry-standard protections.
4.3 2024 BIPA Amendment (SB 2979)
Key Change: Same biometric identifier from same person = single violation (not per-scan).
Penalties Post-Amendment:
- $1,000 per negligent violation
- $5,000 per intentional/reckless violation
- Electronic signatures now explicitly valid for consent
Part 5: European Union Requirements
5.1 GDPR Call Recording Framework
Under GDPR, voice recordings and transcripts are personal data. Voice can also be biometric data if used for identification.
Lawful Bases for Processing (Article 6)
Consent (Art. 6(1)(a)) — When applicable: Training, quality assurance, marketing analysis. Documentation required: Freely given, specific, informed, unambiguous; Affirmative action; Withdrawal mechanism.
Contract (Art. 6(1)(b)) — When applicable: Recording necessary to perform contract with caller. Documentation required: Contract documentation; Necessity analysis.
Legal Obligation (Art. 6(1)(c)) — When applicable: Financial services (MiFID II), regulated industries. Documentation required: Regulatory requirement documentation.
Legitimate Interest (Art. 6(1)(f)) — When applicable: Fraud prevention, security, limited training. Documentation required: LIA (Legitimate Interest Assessment); Balancing test; Data subject rights don't override.
Critical: Implied consent (continuing call after disclosure) is no longer sufficient under GDPR. Affirmative consent (press 1, say "I agree") required.
5.2 EU AI Act Requirements
Article 5 Prohibited Practices (Effective February 2, 2025)
Emotion Recognition in Workplace — AI inferring emotions of workers based on biometric data (including voice) is prohibited. Exceptions: Medical reasons; Safety reasons (e.g., driver fatigue detection).
Emotion Recognition in Education — AI inferring emotions of students based on biometric data is prohibited. Exceptions: Medical reasons; Safety reasons.
Biometric Categorization — Categorizing individuals based on biometrics to infer race, political opinions, religion, sexual orientation is prohibited. Exceptions: Lawfully acquired data labeling/filtering (e.g., law enforcement).
Voice AI Impact: Employee call monitoring with sentiment/emotion analysis is now prohibited in the EU unless for medical/safety purposes. Customer emotion analysis remains permitted but subject to Article 50 transparency requirements.
Article 50 Transparency Obligations
- AI Disclosure: Must inform person they are interacting with AI (unless obvious from context)
- Emotion Recognition Disclosure: Must inform exposed persons about operation of emotion recognition system
- Synthetic Voice Disclosure: AI-generated voice content must be disclosed as artificially generated
5.3 EU Penalties
- GDPR (standard violations): Up to €20 million or 4% global annual turnover
- EU AI Act (prohibited practices): Up to €35 million or 7% global annual turnover
Part 6: United Kingdom Requirements
6.1 UK GDPR + Data Protection Act 2018
Post-Brexit, UK maintains GDPR-equivalent requirements under UK GDPR:
- ICO (Information Commissioner's Office) is the supervisory authority
- UK GDPR fines: Up to £17.5 million or 4% of global turnover
- Same consent standards as EU GDPR
6.2 Privacy and Electronic Communications Regulations (PECR)
PECR governs direct marketing calls in the UK:
- TPS/CTPS: Telephone Preference Service and Corporate TPS must be checked before marketing calls
- Automated Calls: Specific consent required before making automated/prerecorded marketing calls
- AI Voice Calls: AI-generated voice calls treated as automated calls requiring specific consent
PECR Penalties: Up to £500,000 (increasing to GDPR-level penalties under pending legislation). Directors can be personally liable.
Part 7: Canada Requirements
7.1 Federal: PIPEDA
Personal Information Protection and Electronic Documents Act applies to private sector organizations across Canada.
PIPEDA Call Recording Requirements:
- Inform at call start with automated message or agent disclosure that call is being recorded
- State purpose clearly with specific purposes stated; cannot say "quality assurance" if used for marketing
- Obtain consent; implied consent acceptable if notification given and call continues; express consent recommended
- Limit use to stated purpose; recording can only be used for purposes disclosed at time of collection
- Provide access on request; customers have right to access their call recordings
7.2 Quebec Law 25 (Loi 25)
Quebec has stricter requirements than federal PIPEDA:
- Privacy Impact Assessments required before launching tech projects involving personal data
- Clear opt-in consent for sensitive data
- Notice required when automated decision-making affects rights
- Effective: Fully rolled out September 2024
7.3 Criminal Code One-Party Consent
Section 184 of the Criminal Code: Recording is legal if one party consents. However, this only addresses criminal liability—PIPEDA/provincial laws add additional requirements for business use.
Part 8: Latin America Requirements
8.1 Brazil: LGPD (Lei Geral de Proteção de Dados)
Brazil's LGPD is one of the most comprehensive data protection frameworks in Latin America, with GDPR-like requirements and extraterritorial reach.
Key LGPD Requirements for Voice AI
- Consent Model: Opt-in consent required; must be free, informed, unambiguous, and specific
- Written Consent: Must be in writing for specified purpose; general authorizations void
- Withdrawal Rights: Data subject can revoke consent at any time; must be as easy as giving consent
- Burden of Proof: Controller must prove consent was obtained in compliance with LGPD
- DPO Appointment: Data Protection Officer required regardless of processing volume
Brazil Call Center Regulations (SAC Decree + Anatel)
- Call Recording Mandatory: All customer service contacts must be recorded
- Recording Availability: Interaction history must be available to consumers on request
- 0303 Prefix Requirement: Companies making >10,000 daily calls must use 0303 prefix (effective March 2025)
- Maximum Wait Time: 60 seconds maximum for customer to reach representative
- 24/7 Availability: Customer service must be available around the clock
LGPD Penalties
- Warning with deadline for corrective measures
- Simple fine up to 2% of revenue (capped at R$50 million per violation)
- Daily fine
- Publicization of infraction
- Data deletion
8.2 Mexico: Ley Federal de Protección de Datos Personales
Mexico's federal data protection law applies to private sector entities processing personal data of Mexican residents:
- Privacy Notice: Must be provided before or at time of data collection
- Consent: Express consent required for sensitive data; tacit consent may be sufficient for non-sensitive
- Revocation: Data subjects can revoke consent at any time
- ARCO Rights: Access, Rectification, Cancellation, Opposition rights must be honored
Part 9: Middle East Requirements
9.1 UAE: PDPL (Federal Decree-Law No. 45/2021)
The UAE has implemented comprehensive data protection through the PDPL, with strict call recording requirements under the Cybercrime Law.
UAE Call Recording Requirements
- Two-Party Consent: All parties must provide explicit permission before recording
- Consent Type: Verbal consent (and recorded) acceptable, but written preferred
- Criminal Liability: Recording without consent violates Cybercrime Law (Federal Law No. 5/2012)
- Corporate Exception: Call centers may record with automated disclaimer if caller continues
UAE Telemarketing Regulations (Cabinet Resolutions 56 & 57 of 2024)
- Prior Approval: Explicit approval from authorities required before telemarketing
- Recording Disclosure: Callers must be informed at outset that call is recorded
- Do Not Call Registry: Must respect consumers listed in DNC registry
- Operational Hours: Telemarketing calls permitted only 9 AM - 6 PM
- Local Numbers: All marketing calls must originate from locally registered numbers
UAE Penalties
- Administrative fines (amounts to be specified in Executive Regulations)
- 2024 enforcement: 159 companies fined AED 50,000 each for telemarketing violations
- Total fines reached AED 3.8 million for DNC violations
9.2 Saudi Arabia: PDPL
Saudi Arabia's Personal Data Protection Law came into effect September 2023:
- Consent: Express consent required before processing personal data
- Purpose Limitation: Processing limited to purposes disclosed at collection
- Data Minimization: Collect only necessary data
- Cross-Border Transfer: Restrictions on transferring data outside Saudi Arabia
Part 10: Africa Requirements
10.1 South Africa: POPIA + RICA
South Africa has a dual framework: POPIA governs data protection while RICA governs interception of communications.
Recording Consent Framework
RICA allows one-party consent — Recording legal if one party consents; exceptions for business purposes.
POPIA requires processing consent — Personal information cannot be processed (recorded) without consent.
Combined Effect: While RICA allows one-party, POPIA's processing requirements effectively require disclosure. Treat as all-party for compliance.
2025 POPIA Amendments (Effective April 17, 2025)
- Opt-Out ≠ Consent: Providing opt-out mechanism alone does not constitute valid consent
- Telephonic Consent Recording: If consent obtained by phone, must be electronically recorded and available to data subject
- Response Timeline: 30 days to respond to correction/deletion requests
- Compliance Framework: Must be "continuously improved"
South Africa Penalties
- Fine up to R10 million
- Imprisonment up to 10 years
- Civil liability for privacy infringement
10.2 Nigeria: NDPR (Nigeria Data Protection Regulation)
Nigeria's NDPR applies to all transactions involving Nigerian data subjects:
- Consent: Data subjects must consent to processing of personal data
- Purpose Specification: Purpose must be specified at time of collection
- Data Minimization: Only collect data necessary for purpose
- Security: Appropriate technical and organizational measures required
Part 11: Asia-Pacific Requirements (Expanded)
11.1 Japan: APPI
Act on the Protection of Personal Information applies to any business handling personal information of Japanese individuals:
- Purpose Specification: Specify and notify purpose of use before collection
- Consent for Sensitive Data: Explicit consent required for sensitive personal information
- Third-Party Transfer: Consent required before providing personal data to third parties
- Cross-Border Transfer: Special requirements for transfers outside Japan
Japan AI Act (May 2025): Japan's first comprehensive AI law takes a voluntary, best-practices approach (no strict prescriptive rules like EU AI Act).
11.2 Australia: Privacy Act 1988 + Telecommunications (Interception) Act
Call Recording Rules
- General Rule: Calls may not be recorded (Section 7 prohibition)
- Exception: Recording permitted with knowledge of person making communication
- State Laws: Listening device laws may also apply
- APPs: Notice of collection required; purpose limitation applies
11.3 Singapore: PDPA (Personal Data Protection Act)
Singapore's PDPA is one of the most developed frameworks in Southeast Asia.
PDPA Call Recording Requirements
- Consent: Clear, unambiguous, voluntary consent required before recording
- Notification: Inform callers that calls may be recorded with specific purposes
- Opt-In: Pre-ticked boxes or assumed consent not valid
- Withdrawal: Must inform of right to withdraw consent; mechanism required
- Do Not Call Registry: Must check DNC before marketing calls
PDPA Penalties
- Individuals: Up to S$200,000 or 5% of annual Singapore turnover (whichever higher)
- Organizations: Up to S$1 million or 10% of annual Singapore turnover (whichever higher)
- Private right of action for individuals who suffered loss
11.4 South Korea: PIPA (Personal Information Protection Act)
South Korea's PIPA is one of the world's strictest data protection regimes, with opt-in consent requirements.
PIPA Requirements for Voice AI
- Opt-In Consent: Explicit, informed, opt-in consent required (not opt-out)
- Consent Specificity: Must specify: purpose, data items collected, retention period, right to refuse
- Separate Consent: Sensitive data requires separate, specific consent
- 2024 Amendment: Mandatory consent for contract performance no longer permitted; bundled consent prohibited
- Call Recordings: Phone lists and recordings = personal data; consent required
PIPA Penalties
- Administrative fines up to 3% of relevant revenue
- Corrective orders
- Criminal sanctions possible
11.5 India: DPDP Act (Digital Personal Data Protection Act)
India's DPDP Act establishes a consent-centric regime with phased implementation through May 2027.
DPDP Act Requirements
- Consent Standard: Free, specific, informed, unconditional, unambiguous through clear affirmative action
- Privacy Notice: Must accompany or precede consent request; must include purposes, data categories, grievance mechanism
- Language: Notice must be in English or any of 22 scheduled Indian languages
- Consent Manager: Novel concept: intermediaries help data principals manage consent
- Children's Data: Verifiable parental consent required for under-18s
DPDP Implementation Timeline
- November 13, 2025: Data Protection Board established
- November 13, 2026: Consent Manager registration opens
- May 13, 2027: Full compliance required (consent, notice, security, breach notification)
India Call Recording Rules
- Self-Recording: Legal if you are party to conversation (similar to US one-party)
- Interception: Prohibited except with government authorization (Telegraph Act Section 5)
- Business Recording: Since August 2023, businesses recording calls are "Data Fiduciaries"
- DPDP Requirements: Clear notice, free/specific/informed consent, secure storage, delete when purpose served
DPDP Penalties
Up to ₹250 crore (approximately $30 million) per violation
11.6 China: PIPL (Personal Information Protection Law)
China's PIPL imposes GDPR-style obligations with additional requirements for cross-border transfers.
PIPL Consent Requirements
- General Consent: Free, voluntary, explicit, based on full information
- Separate Consent: Required for: third-party transfers, public disclosure, overseas transfers, sensitive data
- Sensitive Data: Includes: biometrics, financial accounts, location, health, minors' data
- Written Consent: May be required for biometrics and certain financial data
- New Consent: Required if processing purposes, means, or data categories change
PIPL Voice AI Specific Rules
- Call Recordings: Voice recordings may constitute biometric data (sensitive PI)
- Cross-Border Transfers: Separate consent + security assessment/certification/standard contract required
- 2024 Enforcement: First extraterritorial case: French hotel group found non-compliant for cross-border data sharing
PIPL Penalties
- Confiscation of illegal gains
- Fines up to RMB 50 million or 5% of previous year's revenue
- Individuals: Up to RMB 1 million; may be prohibited from senior management roles
Part 12: Unified Consent Architecture for Voice AI
For multinational voice AI operations, a unified consent architecture must satisfy the strictest requirements across all applicable jurisdictions.
12.1 Pre-Call Consent (Written)
Your written consent form must include:
- TCPA PEWC Elements: Autodialer/artificial voice disclosure; Specific seller; Not-a-condition; Phone number; Signature
- Recording Consent: Calls will be recorded; Transcripts created; Retention period
- AI Processing Consent: AI/automated systems will analyze call; Purposes specified
- Third-Party Disclosure: Data shared with [named technology providers] for processing
- Training Consent (CIPA 632(d)): Separate consent if data may be used to train AI models
- Biometric Notice (BIPA): Voiceprint collection notice; Purpose; Retention; Destruction policy
- Cross-Border Transfer (PIPL/LGPD): Separate consent for international data transfers
12.2 Call-Start Disclosure (Scripted, Not AI-Generated)
Critical: Disclosure must be delivered by deterministic system (IVR, scripted audio) before conversational AI takes control. AI cannot skip, modify, or delay required disclosures.
Sample compliant disclosure:
"This call is being recorded and may be analyzed by AI systems for [stated purposes]. By continuing this call, you consent to recording and AI analysis. To speak with a human agent at any time, say 'agent' or press zero. Do you consent to proceed?"
12.3 Affirmative Consent Capture
- Press 1: DTMF tone capture logged with timestamp
- Say "I agree": Speech recognition logged with timestamp and audio segment
- Decline: Route to non-recorded continuation or human agent
Timing Problem Solution: CIPA requires consent before interception, but AI needs to listen to hear consent. Solution: Consent capture runs on separate non-AI system; only after consent captured does AI processing begin.
Part 13: Audit Evidence Matrix
The following sections map each compliance area to specific evidence requirements, distinguishing between evidence that passes regulatory review and evidence gaps that trigger findings.
13.1 Consent Documentation
Evidence That Passes:
- Database entry linking phone number to consent with timestamp, source, version
- Exact screenshot/HTML of form at consent time
- Explicit mention of autodialer, AI/artificial voice, specific seller, not-a-condition
- E-SIGN compliant signature: timestamp, IP address, session ID
- Audit trail showing which form version was active for each consent
- System that verifies consent exists before call proceeds
- Separate consent for international transfers (PIPL, LGPD)
Red Flags That Trigger Findings:
- Consent stored only in CRM notes, not structured database
- No preserved copy of actual form user signed
- Generic "receive calls" without AI/artificial voice mention
- No signature or non-compliant capture method
- Form changed but old consents not flagged for re-consent
- No pre-call consent verification in call flow
- Bundled consent without specific cross-border disclosure
13.2 Disclosure Verification
Evidence That Passes:
- Documented, version-controlled script with all required elements
- System logs confirming disclosure played before AI conversation
- Call recordings showing disclosure in first 10 seconds
- Tested opt-out mechanism that actually functions
Red Flags That Trigger Findings:
- AI generates disclosure dynamically (can omit/modify required elements)
- "We have a script" without verification it was delivered
- No recording samples available for review
- "Press 1 to opt out" without testing that pressing 1 works
13.3 Revocation Handling
Evidence That Passes:
- Database with timestamp, source, phone number, request language
- Metrics showing time from request to system update (≤10 days, better if real-time)
- Logs showing DNC API actually called with correct parameters
- AI waits for API success before confirming to customer
- Kill switch in recording pipeline; processing stops when customer says "stop recording"
Red Flags That Trigger Findings:
- Revocations logged in conversation notes but not systematically processed
- Calls placed during "processing window" after revocation
- AI claims success but API never called (hallucinated tool execution)
- AI confirms removal before async API call completes
- AI acknowledges but has no connection to recording pipeline
13.4 Vendor Management
Evidence That Passes:
- List of all third parties processing voice data with data flows
- Contract with agent relationship; Data use restrictions; No training use; Sub-processor restrictions
- Visual showing where voice data goes, who processes it, what happens to it
- Consent language covers third-party technology provider processing
- Documentation of transfer mechanism (SCCs, adequacy, consent)
Red Flags That Trigger Findings:
- "We use an AI provider" without documentation of which one
- Contract allows vendor to "improve services" or "train models"
- No data flow documentation
- Consent says "recorded for quality assurance" with no AI/third-party mention
- Data transferred internationally without legal basis
13.5 AI Output Controls
Evidence That Passes:
- Architectural controls (not just prompts) preventing compliance violations
- Pricing/terms/eligibility retrieved from verified database, not generated
- 500+ adversarial prompt test results documented
- Continuous sampling dashboard with alert thresholds
Red Flags That Trigger Findings:
- "We told the AI not to do that in the prompt"
- AI generates pricing dynamically (hallucination risk)
- Testing only covers happy path scenarios
- "Passed compliance at launch" with no ongoing monitoring
Part 14: The Difference Between "Program Exists" and "Program Works"
"Program exists" (triggers findings):
- We have a consent form
- We have a disclosure script
- We have a DNC policy
- We have an AI vendor contract
- We told the AI what to do
"Program works" (passes review):
- We verify consent before every call and can prove it
- We verify disclosure was delivered and can prove it
- We verify revocations are processed and can show the tool call logs
- We verify vendors are restricted and can show the contract terms
- We verify AI outputs are controlled and can show the testing results and monitoring data
Auditors and regulators distinguish between documentation that says you comply and evidence that proves you comply. Voice AI deployments fail this test more often than traditional systems because the AI introduces variability that paper policies don't control.
Quick Audit Readiness Scorecard
Use this checklist to assess your current state:
Consent Infrastructure
- ☐ Consent records with timestamps, form preservation, version control
- ☐ Pre-call verification system confirms consent before dialing
- ☐ Consent language explicitly covers AI/artificial voice technology
- ☐ Separate consent for cross-border transfers where required (PIPL, LGPD)
Disclosure Delivery
- ☐ Deterministic disclosure delivery (not AI-generated)
- ☐ System logs proving disclosure played before AI conversation
- ☐ Sample recordings available showing disclosure in first 10 seconds
Revocation Processing
- ☐ Tool call logs proving DNC/revocation requests actually processed
- ☐ Processing time metrics (request → system update)
- ☐ Mid-call withdrawal mechanism connected to recording pipeline
Vendor Compliance
- ☐ Vendor contracts prohibit training use and "improve services" clauses
- ☐ Data flow documentation showing all third-party processing
- ☐ Customer consent language covers third-party AI providers
- ☐ Cross-border transfer mechanisms documented
AI Output Controls
- ☐ Architectural guardrails (not prompt-only)
- ☐ Edge case testing documentation from last 90 days
- ☐ Production monitoring with alert thresholds
Jurisdictional Compliance
- ☐ EU AI Act: No workplace emotion recognition
- ☐ Brazil LGPD: Opt-in consent with withdrawal mechanism
- ☐ India DPDP: Verifiable consent with grievance mechanism
- ☐ China PIPL: Separate consent for sensitive data and cross-border transfers
Scoring:
- 15-18 boxes checked: Audit-ready for global operations
- 11-14 boxes checked: Gaps exist—prioritize remediation
- Below 11: Significant exposure—conduct full assessment
Self-Assessment Questions
- Can you produce, right now, the consent record for a specific phone number you called last week—including the exact form they signed?
- Can you produce call recordings proving disclosure was delivered in the first 10 seconds?
- Can you produce tool call logs showing a DNC request from last month was actually sent to your suppression API?
- Can you produce your AI vendor contract and show the clause prohibiting training use?
- Can you produce edge case test results for your voice AI from the last quarter?
- For calls to California, can you prove third-party AI processing was disclosed and consented to?
- For calls involving Illinois residents, do you have BIPA-compliant voiceprint consent with retention/destruction policy?
- For EU operations, have you audited for prohibited emotion recognition in workplace contexts?
- For Brazil operations, do you have opt-in consent with documented withdrawal mechanism?
- For China operations, do you have separate consent for cross-border data transfers?
- For India operations, are you prepared for May 2027 DPDP compliance deadline?
If any answer is "no" or "I'd have to check," you have audit readiness gaps.
VoiceLint Audit Readiness Assessment
We evaluate voice AI deployments against regulatory evidence requirements across all jurisdictions covered in this guide—not just policy existence, but proof of operational compliance.
Assessment Scope:
- Consent Architecture Review: Can you prove consent existed before each call? Form preservation, version control, call-to-consent matching
- Disclosure Verification Audit: Can you prove disclosures were delivered? Timing verification, deterministic delivery, sample recordings
- Revocation Handling Test: Do tool calls actually execute? Correct parameters? Confirmation handling? Mid-call withdrawal?
- Vendor Contract Review: Do terms create third-party exposure under CIPA capability test?
- AI Output Testing: Edge case battery (500+ prompts), hallucination detection, production monitoring gaps
- Jurisdictional Compliance: State-by-state consent requirements, EU AI Act prohibited practices, BIPA voiceprint requirements, LGPD opt-in, PIPL cross-border, DPDP readiness
Deliverables:
- Gap Assessment Report: Exactly what evidence you're missing, mapped to specific regulatory requirements across 40+ jurisdictions
- Risk Exposure Analysis: Per-jurisdiction penalty exposure based on current gaps
- Remediation Roadmap: Prioritized by risk exposure, with specific implementation guidance
- Audit-Ready Documentation: Templates for consent forms, vendor contracts, testing protocols
