Episodes

  • Card Payments on AI Phone Calls and PCI Scope
    Nov 20 2026

    Taking a credit card payment over an AI phone call seems straightforward — until you map out every component the audio touches. This episode of Phony.ai works through the PCI DSS 4.0.1 scope problem in detail, examining how cardholder data can silently flow through a voice stack and what operators must do to contain it. The discussion draws on the Phony.ai deep-dive on card payments and PCI scope to give builders a practical compliance framework.

    The episode covers the full landscape of risk and mitigation, including:

    • What "scope" actually means for a voice call — a typical AI phone call routes audio through at least five components (carrier, speech-to-text, language model, text-to-speech, and recording storage), and a spoken card number brings all of them into the cardholder data environment.
    • The CVV storage trap — PCI DSS 4.0.1 prohibits storing card verification codes after authorisation, and a call recording or transcript that captures a spoken CVV counts as storage, regardless of how briefly the file was intended to be kept.
    • DTMF masking — intercepting keypad tones at the carrier level and replacing them with a neutral beep before they reach downstream systems; the most effective single step for reducing scope, keeping card digits confined to the carrier-to-processor path.
    • Pause-and-resume recording — a fallback layer that closes the storage window during the payment moment but does not address the transmission path; best treated as a complement to tone masking, not a substitute.
    • Tokenised payment link handoff — the agent sends the caller a text or email link to the processor's own hosted payment page, so the card number never crosses the operator's carrier; this approach can reduce a compliance self-assessment from roughly 300 controls (SAQ D) down to 13 (SAQ A).
    • Fraud risk and agent guardrails — with fraud attempts hitting approximately one in every 599 contact-centre calls and knowledge-based authentication failing more than half the time, payment flows on AI agents need bounded, structured transaction logic rather than open-ended model improvisation.

    The episode closes with a practical diagnostic: track the ratio of payments completed through the tokenised handoff against those that end up being finished by a human reading digits off a callback note. A rising second number is an early warning that the payment flow is routing around the design — and the fix usually lives in SMS deliverability, the handoff page, or the resume logic inside the call data security layer, not in the agent prompt. For teams thinking about how actions an agent can take intersect with regulated data, this episode offers a concrete starting point.

    For more on building compliant AI voice experiences, listen to What Your AI Voice Agent Must Say Before the Second Ring — a companion episode on disclosure obligations.

    AI phone software

    Show More Show Less
    5 mins
  • What Your AI Voice Agent Must Say Before the Second Ring
    Nov 17 2026

    AI voice agents have quietly stepped into a third legal layer that most businesses never see coming. Beyond call-recording consent and TCPA dialing rules sits a separate and fast-growing body of law: the right of a caller to know they are speaking with a machine. This episode of Phony.ai walks through the current state-by-state disclosure landscape, drawing on this in-depth guide to bot disclosure laws for AI phone calls, and shows exactly how to craft an opening line that satisfies all of them at once.

    Here's what the episode covers:

    • California's two-track framework: The 2019 BOT Act set the original commercial disclosure floor, while AB 2905 — effective January 2025 — specifically targets synthetic and AI-altered voice in outbound calls, with civil penalties up to $2,500 per violation.
    • Utah's on-request trap: Utah doesn't require upfront disclosure for most callers — but the moment anyone asks "am I talking to a real person?", an agent that deflects or answers in the first person triggers a per-call fine. Regulated professions like law firm intake calls and healthcare face an even stricter proactive duty.
    • Colorado and Texas heading for proactive defaults: Colorado's law (effective June 2026) requires disclosure unless it would be "obvious to a reasonable person" — a carve-out that doesn't protect a natural-sounding voice agent. Texas healthcare providers face a similar proactive requirement from January 2026.
    • The EU AI Act's Article 50: Enforceable from August 2026, it mandates that any user interacting with an AI system be informed of that fact — a global consideration for any business taking calls from EU numbers.
    • What a compliant opening line actually sounds like: A single first-turn phrase using a recognizable non-human noun ("AI assistant" or "automated assistant") satisfies California, Utah, Colorado, and Texas healthcare rules simultaneously — no lengthy script required.
    • The four compliance fundamentals: Disclose in the first turn, use a noun callers recognize, ensure the agent can handle the follow-up question gracefully, and log the exact opening line per call for every conversation.

    The pace of legislation is accelerating — 99 AI-related laws passed in 2024, 144 in 2025, with active disclosure bills moving through Illinois, New York, and New Jersey. Getting the opening line right today is far cheaper than retrofitting it after a complaint. If you're building or refining a voice agent, it's worth reviewing how call handling and transcripts can help you log and audit exactly what your agent says on every call. For more on recording obligations that often travel alongside disclosure rules, listen to What Your AI Voice Agent Must Say Before It Records the Call.

    Phony.ai

    Show More Show Less
    5 mins
  • What Your AI Voice Agent Must Say Before It Records the Call
    Nov 15 2026

    Most businesses deploying AI phone agents have heard of the TCPA. Far fewer have mapped the parallel legal regime that governs whether a call can be recorded at all — and that gap is where serious liability quietly accumulates. This episode uses a 2025 federal court ruling against an AI voice assistant handling Domino's franchise calls as a jumping-off point to unpack what AI agents must disclose before recording begins, and why getting it wrong carries consequences far heavier than most operators realize.

    The episode works through the full landscape of recording consent law — from the federal baseline to the state statutes that override it — and lands on concrete, implementation-ready guidance for anyone building or deploying a voice agent today. Key ground covered includes:

    • One-party vs. all-party consent: Federal wiretap law lets a party to the call consent on its own behalf, but roughly thirteen states reject that standard entirely — requiring every participant's consent before a recording is lawful.
    • Criminal exposure, not just civil fines: Illinois treats a first-offense business recording violation as a Class 4 felony; Pennsylvania classifies it as a third-degree felony with fines up to $15,000; California's statutory floor is $5,000 per call.
    • Extraterritorial reach: The caller's state controls, not the business's. A dental practice in Texas answering a call from a California number must deliver the California disclosure — the business's physical location is irrelevant.
    • The disclosure script itself: Say "recorded," not "monitored." Combine the AI identity disclosure and the recording notice in a single breath. And critically — any offered opt-out must be technically real; an opt-out the voice agent cannot actually honor is legally worse than no opt-out at all.
    • Washington's recording-order rule: The recording track must start before the disclosure plays — not after. If the opening greeting is trimmed for storage efficiency, the consent evidence is gone with it, and call logs alone are not a substitute for the audio record.
    • State-aware routing: The caller's state is knowable from the number before the agent says a word. That signal should drive which disclosure script fires. Unknown or spoofed numbers should default to the strictest applicable standard, not a guess — and calls where consent cannot be evidenced end-to-end simply should not be recorded. Reviewing how your platform handles call handling and transcripts is a sensible first audit step.

    For more on the regulatory side of AI-powered outbound calling, the episode TCPA and AI Calls in 2026: What Every Outbound Campaign Must Know covers the companion framework that governs whether an AI agent can dial in the first place. Together, the two episodes map both sides of the legal picture every voice-agent operator needs to understand.

    Phony.ai

    Show More Show Less
    5 mins
  • TCPA and AI Calls in 2026: What Every Outbound Campaign Must Know
    Nov 13 2026

    AI-generated voices changed the TCPA compliance landscape overnight — and most outbound teams haven't caught up. This episode of Phony.ai unpacks the full legal picture for AI voice campaigns in 2026, using real enforcement numbers to show exactly what's at stake. If you're placing outbound calls with any kind of AI voice, this episode explains the rules your campaign has to meet before the first dial goes out. The discussion draws directly from the show's deep-dive guide on TCPA compliance for AI outbound calls.

    Here's what the episode covers:

    • The enforcement wake-up call: The FCC's $6 million forfeiture case tied to deepfake Biden robocalls — and why the carrier that only transmitted the traffic still paid a separate $1 million — frames what's really at risk for teams without proper consent records.
    • The February 2024 FCC ruling: A single declaratory ruling folded AI-generated and cloned voices into TCPA's "artificial or prerecorded voice" definition, making every AI outbound campaign subject to the same framework as robocalls.
    • Two consent tiers, two standards: Informational calls (reminders, delivery updates, 2FA) require prior express consent; marketing calls require prior express written consent with four specific elements — all on the same form — including the dialed number and a clear statement that consent isn't a purchase condition.
    • State law stacks on top: Florida, Oklahoma, Washington, Maryland, and others each carry their own per-violation damages. The rule that applies is determined by where the dialed number is registered — not where the business is located.
    • What a compliant per-call record looks like: Timestamp, source URL, IP address, disclosure text as shown, consent scope and named sellers, a reassigned-numbers check, a DNC scrub result from dial time, the AI disclosure script, and a logged opt-out path — every field, every call. Teams using Phony.ai can capture and review call-level detail through the platform's call handling and transcripts tools, and pair that with built-in AI disclosure requirements to keep the agent's opening seconds compliant.
    • The one-to-one consent reversal: The Eleventh Circuit vacated the FCC's one-to-one consent rule in January 2025, and the FCC formally removed it in September 2025 — a single form can again cover multiple logically related sellers, but the form must still name them, and it must stay current as that list changes.

    The episode closes with a clear-eyed look at the litigation math: $500 per negligent violation, $1,500 per willful violation, a four-year statute of limitations, and class-action exposure that requires no showing of actual harm. A campaign that miscategorizes consent across 10,000 numbers carries that tail risk through 2030. For more on the economics of running AI-powered phone operations, check out the episode The Real Math Behind Replacing Your Receptionist and Phone System.

    Phony.ai

    Show More Show Less
    5 mins
  • The Real Math Behind Replacing Your Receptionist and Phone System
    Nov 10 2026

    Running a business phone line feels like a single expense — until you actually add it up. This episode of Phony.ai works through the complete, fully-loaded cost of the typical small-business phone stack: the receptionist, the hosted phone system, and the overflow answering service. If you've never seen all three bills on the same page, the total tends to be a surprise. The source article on replacing your receptionist and phone system goes even deeper, and this episode turns those numbers into an honest, line-by-line comparison.

    Here's what the episode covers:

    • The true cost of a receptionist seat. A $43,000 salary becomes closer to $56,000 once payroll taxes, benefits, and paid time off are factored in — and that's before overtime for practices open six days a week.
    • What your PBX invoice actually contains. Per-seat monthly fees are only the starting point; handsets, provisioning, hold-music licences, and support contracts routinely push a small-team cloud PBX to $50–$100 per line per month — none of which answers a single call.
    • The hidden double-billing inside overflow answering services. Live coverage around the clock typically runs $800–$1,500 a month, and the callback the front desk makes the next morning is a second labour cost on top of it.
    • A real per-minute comparison. Nine hundred inbound minutes at twelve cents per minute lands around $117 for the month — versus more than $1,000 through a live answering service at prevailing rates. Beyond the headline number, AI cost is variable while the existing stack is almost entirely fixed.
    • The calls an AI agent shouldn't take. Situations requiring unscripted human judgment, emotionally distressed callers, and anything outside documented consent should route to a person. Seamless human handoff isn't a fallback — it's a design principle that makes the whole system work.
    • Three numbers to pull before you run the comparison. Total inbound minutes, average call duration, and the percentage of calls hitting voicemail or a hold longer than 60 seconds. Industry benchmarks call for answering 80% of calls within 20 seconds and keeping abandonment below 5% — if you're far from either, the current stack may already be costing you business.

    The episode also walks through a simple DIY calculation: multiply last month's minutes by $0.12, add $0.01 for carrier passthrough, and compare the result to what you actually paid across all three bills. The delta tells you whether a voice agent is worth exploring further. For businesses where every call needs to be logged, transcribed, and acted on, the efficiency case is especially clear. For more on what happens after you make the switch, listen to Voice Agent First 90 Days: What Breaks and What the Fix Actually Costs.

    AI phone software

    Show More Show Less
    5 mins
  • Voice Agent First 90 Days: What Breaks and What the Fix Actually Costs
    Nov 8 2026

    A demo environment is designed to succeed. Production is not. This episode of Phony.ai works through the operational reality of deploying a voice agent and surviving the first ninety days — drawing on the companion article on first-90-days voice agent failures and fix costs to map out what goes wrong, in what order, and what distinguishes a one-hour configuration fix from a full re-implementation.

    The episode is anchored by a striking benchmark: pre-launch evaluations typically show 90–95% of calls succeeding, while production can land 20–30 points lower — and 88% of AI agent pilots never reach production at all, largely due to operational shortfalls rather than model limitations. Voice agents fail in specific, nameable ways, and the show names them in sequence:

    • Routing failures (weeks 1–2): Agents do exactly what they're configured to do — but routing tables are wrong. After-hours calls drop into unmonitored inboxes; emergency transfers hit full voicemail boxes. The fix is a disciplined audit of every hour boundary, holiday, and edge case, validated with off-hours test calls that cost less than $12 in platform time.
    • Speech-to-text degradation (weeks 2–4): The best STT models score 95–98% accuracy on clean audio — and lose 7–10 points when background noise enters the picture. Acoustic mismatches from job sites, drive-thrus, and care facilities account for 63% more voice AI failures than background noise alone. The episode distinguishes between a configuration fix (switching STT vendor or tuning endpointing) and a stack swap — which may require a full re-implementation if the platform only supports one provider.
    • The half-booking problem (weeks 3–6): The agent confirms a booking; the calendar has no record. Date format errors, timezone mismatches on API write-backs, and silently swallowed 503 errors all produce the same result: a confirmation SMS goes out for an appointment that was never written. The rule the episode lands on — a booking isn't complete until the calendar confirms it — is a morning's worth of configuration work once you know to look for it.
    • Configuration vs. stack changes: Misrouting, calendar write-back logic, voice selection, and human handoff state handling are all fixable on platforms that expose them as choices — and require re-signing on platforms that don't. The episode frames the single most important vendor question: which of these failures can you fix yourself?
    • The cost of not knowing in advance: Operators who are still happy at day 90 are the ones who priced the operational work before launch and knew which lever to pull when each failure arrived. The episode closes by reframing the entire 90-day arc as a planning problem, not a troubleshooting one.

    For more on how outbound calling infrastructure compounds these same failure modes, the episode Six Layers, Four Stacks: What Your Outbound AI Calling Setup Is Hiding covers the hidden complexity on the dial-out side. Full show notes, sources, and benchmarks referenced in this episode are in the companion article — and if you want to see how testing a voice agent before launch can shorten this failure timeline, that's a good next stop.

    Phony.ai

    Show More Show Less
    5 mins
  • Six Layers, Four Stacks: What Your Outbound AI Calling Setup Is Hiding
    Nov 6 2026

    Building an outbound AI calling system is surprisingly easy to start and surprisingly hard to understand. The gap between a working demo and a production-grade stack is usually invisible — hidden inside vendor pricing tiers, opaque orchestration layers, and a compliance surface that most tutorials skip entirely. This episode of Phony.ai tears the whole thing apart, layer by layer, using the deep-dive article on outbound AI calling stacks as its technical foundation.

    The episode walks through all six infrastructure layers every outbound call touches — and the four realistic ways teams are assembling them in 2026:

    • The six layers defined: campaign and dialer logic, carrier and phone number, media transport, speech-to-text, text-to-speech, and the language model — plus the layer most demos skip: consent evidence and suppression-list scrubbing before the first dial.
    • Twilio + your own orchestrator: Solid SDK coverage and managed PSTN routing, but the agent layer is entirely your responsibility. All-in costs land between 10 and 18 cents per minute once STT, LLM, and TTS are added, and the published median turn latency of 491 ms doesn't include carrier variance or orchestration overhead.
    • Telnyx + your own orchestrator: Roughly half the carrier rate of Twilio's long-code pricing, with deeper SIP controls including SIPREC and codec selection — but a narrower SDK surface. At volume, the default 20-calls-per-second-per-IP limit can produce silent SIP 503 failures if pacing isn't handled deliberately.
    • All-in-one platforms (Retell, Vapi, Bland, Synthflow): Fast to launch — days, not weeks — with everything collapsed into a single billing line around 7–10 cents per minute. The tradeoff is layer opacity: STT and TTS choices are vendor-curated, LLM execution is off your infrastructure, and a quiet TTS tier upgrade can triple voice costs without changing the invoice format.
    • Fully custom SIP stack: Floor costs can drop below 3 cents per minute at steady volume using tools like Jambonz or FreeSWITCH alongside self-hosted models, but the engineering burden — SBC config, codec negotiation, DTMF handling, 3 a.m. packet captures — is entirely owned in-house.
    • Compliance as infrastructure: The FCC's updated TCPA revocation rules require opt-out keywords to be honored within 10 business days, STIR/SHAKEN A-level attestation affects whether calls survive carrier scrutiny, and any stack that can't produce an immutable consent receipt per call attempt is a liability, not a product. Teams handling sensitive outbound use cases — from number provisioning and call routing through to call recording consent — need this layer designed in from the start, not bolted on later.

    The episode also covers the latency math that determines whether a call feels human: the full round-trip budget from end of speech to start of reply is under 800 milliseconds, and carrier media paths alone contribute 80–150 ms before the model loop is even counted. For more on how AI voice capabilities can expand responsibly beyond initial deployment, check out From Receptionist to AI COO: Expanding Your Voice Agent Safely. More from the Phony.ai blog and platform documentation is linked throughout the episode.

    Phony.ai

    Show More Show Less
    5 mins
  • From Receptionist to AI COO: Expanding Your Voice Agent Safely
    Nov 3 2026

    Most voice agent deployments reach a natural inflection point: the basics are running smoothly, and the temptation to hand the agent more responsibility is real. But expanding from inbound receptionist to something closer to an AI operations manager is a fundamentally different undertaking — one with distinct legal exposure, failure modes, and architectural requirements. This episode of Phony.ai digs into the guide on expanding voice agent scope and permissions safely, laying out a principled framework for growth that doesn't trade short-term convenience for long-term liability.

    Here's what the episode covers:

    • Inbound vs. outbound are different jobs. An agent that answers calls operates within a narrow, bounded brief. An outbound agent initiates contact — choosing the number, the moment, and the opening — and carries fundamentally different legal and operational risk as a result.
    • The TCPA exposure is real and uncapped. The FCC's February 2024 ruling placed AI-generated voices under TCPA rules for artificial or prerecorded calls. Statutory damages of $500–$1,500 per violation, with no aggregate cap, can turn a careless outbound campaign into a business-defining mistake. Expansion should be built against a written permission model, not an owner's enthusiasm.
    • Scope creep is already widespread. A 2026 Enterprise Management Associates survey found AI agents had acted outside their intended scope at 65% of surveyed enterprises, with 29% reporting measurable organizational impact — and only 34% evaluating authorization at execution time. Smaller businesses typically have fewer controls, not more.
    • The right outbound tasks share three traits. The counterparty is known or has given consent, the action has a verifiable outcome, and the cost of a wrong call is bounded. Good fits include appointment confirmations to existing patients, supplier status checks, shift confirmations with staff, and warm follow-ups to leads who requested a callback — not cold outbound, debt collection, or anything irreversible. Phony.ai's agent actions platform is built around exactly this kind of scoped execution.
    • Benchmark performance sets the operational ceiling. Top models hit around 61% success on a single agentic attempt, dropping below 25% across eight consecutive dependent steps — with roughly half of failures traced to wrong arguments passed to function calls. Planning around this baseline is essential.
    • Role-based permissions are the architecture. A role bundles allowed actions, readable and writable systems, operating hours, permitted counterparties, and the precise handoff threshold. Roles are versioned, reviewed, and evaluated at execution time — not set once at setup. Every action generates an itemized log keyed to the role that authorized it, and anything outside the role's scope routes to a human through the same human handoff discipline used on inbound calls.

    For more on the underlying model and infrastructure considerations, the episode When Self-Hosting Your Voice Agent's LLM Actually Makes Sense is a natural companion listen. Full source material, including the permission model framework, is in the blog post linked above.

    Phony.ai

    Show More Show Less
    5 mins