The episode walks through four forcing conditions — regulation, data residency, fine-tuned accuracy, and call volume economics — and builds an honest checklist operators can run before committing to GPU hardware:

• Regulated transcripts as the trigger: When callers voluntarily disclose PHI, PII, or privileged information (as they routinely do on medical practice phone lines or law firm intake calls), every vendor touching the transcript is in HIPAA's BAA scope — and self-hosting removes one counterparty from that chain entirely.
• Data residency as a binary question: Inference either occurred inside the required jurisdiction or it didn't. Most hosted model region lists shrink considerably once failover zones are accounted for, making self-hosting the only path to a hard compliance guarantee.
• The EU AI Act's enforcement window: High-risk use cases — employment screening, insurance intake, credit decisions — carry real obligations around risk management, data governance, and logging that are structurally easier to meet when the model is under your own roof.
• Fine-tuning economics on open-weight models: Domain-specific tuning on a smaller open model can outperform a larger generic hosted frontier model on actual business calls, with one cited study showing nearly 27% accuracy gains and a tuned 8B-parameter model hitting 92.5% domain accuracy.
• The utilization math: A continuously used H100 at current market rates runs roughly $2,500/month and can serve many concurrent sessions — but the per-minute rate only beats a hosted baseline at genuinely high call volumes. Two thousand minutes a month doesn't clear the bar; two hundred thousand probably does.
• The checklist before signing a GPU contract: The episode closes with four yes/no questions that quickly reveal whether self-hosting is a real option or just an interesting whiteboard exercise.

Whether you're building a voice agent for a regulated industry or simply scaling call volume to the point where per-minute costs become a line item worth attacking, understanding the LLM layer is no longer optional. For more on connecting that infrastructure to your existing phone setup, the episode Four Ways to Plug an AI Voice Agent Into Your Existing Phone System picks up where this one leaves off.

AI phone software

Before comparing patterns, the episode explains what every integration is actually moving — signaling, media, and routing — and why latency is the hidden cost that rarely makes it onto anyone's planning checklist. From there, each of the four patterns is unpacked with its real trade-offs:

• Forward-on-busy — the lightest-touch option, where the PBX stays completely intact and the AI agent handles overflow from a second number; best when a human receptionist answers first and the agent covers what gets missed.
• Full number port and cutover — the main number moves to the voice agent vendor, the PBX is retired, and hunt groups become agent-managed routing rules; the simplest cost structure, but every extension and handset has to be rebuilt or retired.
• SIP leg into the existing PBX — the agent joins as a SIP endpoint inside the current phone system, receiving calls just like any other extension; no forwarding overhead, lowest carrier costs, but requires a modern PBX and working knowledge of codecs and SIP addressing.
• Parallel number alongside the main line — a second number runs entirely through the agent while the main line stays untouched; ideal for piloting a single use case such as after-hours intake or campaign-specific routing, with no disruption to existing operations.
• Latency as a real budget — human conversation runs at roughly 200 ms; voice AI starts feeling off above 500 ms, and every extra forwarding hop or re-transcode erodes that window before the model speaks a word.
• Matching pattern to business type — a dental practice with a busy receptionist and a modern PBX typically fits the SIP leg; a solo law firm on a cell phone is usually a full cutover; a home services company testing after-hours coverage is almost always a parallel number.

The episode frames the decision around three practical facts about any existing setup: whether the PBX supports SIP endpoints, how many calls the business is currently missing, and whether porting the main number is on the table. If you're thinking about what your agent should actually do once it's connected, the human handoff configuration is often the first thing to get right after integration. For more on what surprise costs can show up once an agent is live, check out the earlier episode The Voice Agent Bill You Didn't See Coming.

Phony.ai

The episode works through each layer of the bill in sequence, covering:

• Billed vs. talk minutes: Why carriers round up, why agents run before and after the human speaks, and how failed calls still generate charges — a combination that compounds quickly at volume.
• The transfer tax: In deployments where human handoff is common, the overlap window where both the agent and a live rep are on the line can add four to eight percent to the bill before anything else is counted.
• Voice and model licensing: A per-minute rate is actually a stack — speech-to-text, LLM, text-to-speech, telephony, and platform fee. A premium voice at the top model tier can cost roughly double a standard voice on identical traffic, a difference vendors rarely surface unprompted.
• Minimums and pooling rules: Whether unused committed minutes expire monthly or pool annually can shift the real cost by ten to twenty percent on businesses with uneven call volume — a detail buried in the master service agreement, not the sales deck.
• Overage multipliers: Market overage rates typically run one-and-a-half to two times the base rate, but some platforms go significantly higher. A worked example in the episode shows how the same traffic commitment and the same overage volume can produce a monthly bill that differs by nearly seven times depending solely on which overage clause applies.
• Renewal uplift: Year-two pricing is where long-term cost diverges sharply. Without a cap, documented increases of fifteen percent or more are common — and a contract with an uncapped fifteen percent annual uplift reaches nearly seventy-five percent more in total cost by year four compared to one capped at four percent.

The episode closes with a practical one-page modeling exercise any buyer can run using a vendor's order form and master service agreement side by side — including how to test call reporting data against a vendor's billed-minute methodology before committing. For more on the economics of building with AI voice, also check out the companion episode The Real Cost of Cloning Your Voice for an AI Phone Agent.

Phony.ai

Here's what the episode covers:

• Instant vs. professional cloning: The length of your recording determines which mode you're in — instant cloning works from as little as ten seconds, while a near-indistinguishable professional clone needs at least thirty minutes of clean, consistent audio, and often two to three hours.
• What "clean audio" actually means: One speaker, no music, no room echo, consistent mic distance, and a delivery tone that matches how you want the agent to sound — a parked car often beats a kitchen with a dishwasher running.
• More isn't always better: ElevenLabs explicitly states that recording beyond three minutes for an instant clone can be detrimental — handing the model a long Zoom recording introduces noise, not signal.
• Consent documentation that holds up: A checkbox at upload is a contractual attestation from you, not a verification of the speaker. Professional clones from ElevenLabs and Cartesia require a spoken consent statement in the speaker's own voice. Either way, you need a signed written release, the consent audio file, the provider's upload timestamp and clone ID, and a clear revocation path — especially if the speaker ever leaves the business.
• The legal landscape is moving fast: The NO FAKES Act, which passed unanimously out of the Senate Judiciary Committee in June 2026, would create a federal intellectual property right in a person's voice — turning a departing employee's cloned voice from an HR footnote into an active rights question. The FCC's 2024 declaratory ruling already brings AI-generated voices under existing rules on artificial or prerecorded messages, meaning outbound calls require prior express consent regardless of clone status. Businesses with a deployed voice agent should review their disclosure scripts now — the episode suggests an eighteen-word opener that satisfies both regimes in about two seconds of airtime.
• What cloning actually costs: A cloned voice consumes characters at the same rate as a stock voice — the real cost is the plan tier it sits on. Instant cloning starts on lower-tier plans; professional cloning requires a step up, and multiple brand voices across multiple locations compounds that further.

The episode also flags the scenarios where the honest answer is to skip cloning altogether: if the voice belongs to a public figure, a celebrity, or a staff member without a signed release, the clone should not go live. For teams thinking about compliance from the start, it's worth reviewing AI disclosure requirements before a voice is ever recorded. More from the show: if you're working through how your agent connects to your phone infrastructure, the episode How to Connect an AI Phone Agent to Your Existing SIP Trunk covers the routing side of the setup.

Phony.ai

Here's what the episode covers:

• Bring Your Own Carrier (BYOC) explained: Calls still enter through your existing trunk, authenticate against your carrier, and land on an AI agent endpoint — no porting, no new contracts, no change to your 911 posture or per-minute rate.
• The cost math: Bundled platforms layer their margin on top of telephony costs. Under BYOC, wholesale origination rates (roughly a third of a cent per minute for US local on Twilio, around half a cent outbound on Telnyx) are passed through separately from the agent runtime, so every line on the invoice is independently negotiable.
• Authentication — credentials vs. IP ACL: SIP digest credentials are the safer default; IP allowlisting breaks down when the platform's egress IP shifts or when tenants share a SIP host. A 2024 FCC rule also requires carriers to sign outbound calls with their own STIR/SHAKEN certificate when using third-party caller ID authentication — worth confirming before provisioning.
• Codec selection: G.711 is the universal floor; G.722's wider frequency range meaningfully improves speech-to-text accuracy on digits and spelled names. Mismatched codecs force per-packet transcoding that stacks directly onto the AI agent's existing think-time latency — a problem the episode explains how to avoid.
• DID routing patterns: The episode walks through three common configurations — one DID to one agent, many DIDs to one agent with personalized greetings via the SIP To header, and DID-per-campaign for isolated outbound compliance. Phony.ai's numbers and call routing capabilities map directly to these patterns.
• Failover architecture: A single trunk is a single point of failure. The episode covers secondary SIP destinations at the DID level, keeping a warm backup trunk at a second carrier, and wiring human handoff as the terminal fallback — plus why failover should be tested on a schedule, not just at setup.

For more on how AI stacks up against human agents across cost and resolution rates, the episode AI vs Human Call Centers: Cost, Resolution, and What Actually Belongs Where is a useful companion listen.

AI phone software

Drawing on the line-by-line cost and resolution breakdown published by Phony.ai, the episode walks through:

• The true cost of a human agent — the $7.20 industry average per inbound call includes loaded labor, facilities, technology, and a 40–65% annual turnover problem that compounds every other cost line.
• What AI voice actually costs — voice agents price on minutes, not hours, converging on roughly 7–15 cents per minute all-in, which translates to $0.28–$0.60 on a typical four-minute call versus $3–$7 for a human-handled equivalent.
• Why the 70% resolution figure is almost meaningless — survey data across nearly 200 AI deployments shows a range from 15% to 98.3%, making the median a poor guide for any single business's queue.
• Which calls AI reliably wins — structured conversations with a defined success state (scheduling, order status, hours, intake triage) are where AI often outperforms humans on consistency, not just cost, making them strong candidates for a blended queue with clean human handoff reserved for escalations.
• Which calls still belong to humans — complaints, negotiation, clinical judgment, and compliance-sensitive conversations carry real risk when AI can't gracefully exit; a poor transfer costs more than the call itself.
• A practical three-number diagnostic — before changing anything, pull two weeks of call logs, tag calls by type, and check your routine-call percentage; if more than half the volume is routine, the cost case for AI is already strong.

The episode also works through a concrete blended-queue example at 10,000 calls per month, showing how routing roughly 70% of routine volume to AI and reserving the rest for human agents can reduce blended cost per call by around two-thirds — without exposing the complaint queue to a voice agent that isn't ready for it. A warm transfer that puts a structured transcript on the human agent's screen before the call arrives matters; a cold transfer that forces the caller to repeat everything should be counted as its own cost line. For more from the show on the financial case, listen to AI Voice Agent ROI vs Human Call Centers: The Real Math for 2026.

Phony.ai

Here's what the episode covers:

• True cost of a human call center minute. ContactBabel's 2026 US data puts the average inbound call at $7.20; Gartner's median runs $9–$16. The cost climbs sharply by vertical — healthcare can reach $60 per call when clinical time is included — and the often-overlooked in-house receptionist seat loads out at roughly $50,000 per year before you factor in weekend and evening coverage.
• What a voice AI minute is actually made of. Carrier passthrough, speech-to-text, model inference, text-to-speech, and platform orchestration are each a discrete line item. A three-minute booking call lands in the $0.36–$0.60 range all-in — but only counts as cheap if the call is actually resolved, not quietly escalated to a human anyway.
• Containment rate: the metric most responsible for inflated ROI claims. Vendor demos quote 90%+; production deployments in early tuning typically land at 20–40%, maturing to 40–70% on well-scoped intent sets. Appointment booking and status checks sit near the top of that range; emotionally charged calls and clinical judgment sit near the bottom. Designing for that split is the real work — and configuring a reliable human handoff for the calls AI shouldn't own is just as important as the automation itself.
• Missed-call recovery: the most underrated ROI lever. Across 85 businesses and 58 industries, only 37.8% of inbound calls were answered by a live person. Of callers who didn't get through, 85% never called back and 62% dialed a competitor. With home services leads worth $300–$900 each and legal intake worth $500–$2,500+, recovering even one in five previously missed calls often outweighs headcount savings on the model entirely.
• Five-year total cost of ownership at 10,000 minutes per month. In-house receptionists: ~$320,000. US onshore BPO: ~$420,000. Offshore BPO: ~$180,000 (before QA costs, attrition, and vendor churn erode the gap). Voice AI with human escalation for the remaining 30% of calls: ~$72,000. The offshore comparison — the one most often used to argue against AI — collapses fastest under scrutiny.
• Realistic payback periods by vertical. Healthcare and dental: 2–5 months. Home services: 1–3 months. Real estate: 3–6 months. Financial services: 6–12 months. These are planning bands, not promises, and the episode is explicit about the conditions that move a deployment toward the faster or slower end of each range.

For more on how AI-handled calls are logged and reviewed once a deployment is live, the Phony.ai platform's call reporting tools surface the containment and resolution data that makes ongoing ROI tracking possible. Also worth your time: the previous episode, Seven Steps, One Booked Slot: How Voice AI Actually Schedules Appointments, which walks through the end-to-end booking flow that drives some of the strongest containment numbers discussed here.

Phony.ai

Here's what the episode covers across the seven steps from intent to confirmed slot:

• Intent detection before availability: The agent must classify the type of appointment — not just find open time. Getting this wrong means cheerfully booking a 30-minute slot for a 90-minute procedure.
• Free-busy queries vs. event lists: Pulling a full event list from a calendar leaks meeting titles and attendee names into logs — a security red flag. The right approach reads gaps only, not details.
• System-of-record integration: Most service and healthcare businesses don't run on Google Calendar. Dentrix, ServiceTitan, Jane, and similar platforms each define "available" differently — and a generic calendar primitive can't express that a dental operatory needs a provider, a chair, and a procedure code free simultaneously. The actions an agent can take must map natively to each business's actual system.
• Business-rule enforcement: Free time is not bookable time. Minimum notice windows, provider matching, resource constraints, sterilization buffers, new-patient caps, and insurance filters all sit on top of raw availability — and none of them are defaults in any scheduling platform.
• Concurrency and the double-booking window: Two callers can request the same slot within seconds. Without a short-lived tentative hold between reading availability and writing the booking, a race condition creates a double-booking that surfaces at billing time.
• Write verification: The agent should read the record back immediately after writing it. A mismatch means the booking is treated as failed — and the caller is told on that call, not in a follow-up complaint. Silent write failures are the leading complaint pattern against voice scheduling vendors.
• Reminders and attendance economics: A Cochrane review found text reminders lifted attendance rates from ~68% to ~79%. With industry no-show rates averaging 23% and unfilled healthcare slots costing individual practices over $150,000 a year, the confirmation artifact isn't optional — it's the close of the deal.

The episode also situates all of this against the baseline problem: 62% of inbound calls to small businesses go unanswered, and 85% of callers who reach voicemail hang up without leaving a message. The pipeline only matters if the agent picks up — and getting every subsequent step right is what separates a booked slot from a lost one. For a closer look at how a well-structured voice agent handles these conversations end to end, the episode Why Hybrid AI and Human Voice Is Winning Callers Back is a natural companion listen.

Phony.ai

The episode walks through the economic case, the psychology of caller behavior, and the operational rules that make a hybrid model actually work in practice:

• The cost math: A mid-sized operation running an 80/20 AI-to-human split lands around $300/month — roughly four to six times the cost of pure AI, but still about a quarter of full live-answering rates.
• What callers are really doing: Research across 37 vocal emotion studies shows listeners detect 25 distinct emotions from tone alone — callers aren't just hearing words, they're reading the speaker, and a hollow synthetic voice registers as worse than silence.
• The four call types that always need a human: Grief, injury, and crisis situations; high-value sales where emotional connection drives recall and conversion; complaints and disputes where the caller already feels dismissed; and anything requiring genuine human judgment about a person.
• What AI handles best: Scheduling, rescheduling, after-hours intake, insurance verification, status checks, and routine quote requests — faster, more consistent, and without end-of-day fatigue.
• Disclosure is non-negotiable: The EU AI Act's Article 50 is enforceable as of August 2026, with fines up to €15 million. But the commercial argument mirrors the legal one — 81% of callers find undisclosed AI unethical. Early, clear identification earns the AI the right to handle the next 500 calls. (See Phony.ai's guidance on AI disclosure requirements.)
• The handoff is where hybrid models get judged: Conversational psychology puts the natural gap between speakers at around 200 milliseconds — anything over a second reads as a dropped call. A warm, context-rich transfer to a human within that window is what separates a functional hybrid from a frustrating one. Phony.ai's human handoff tooling is built around exactly this problem.

For more on the mechanics of routing callers to the right destination — human or AI — the companion episode When to Hand the Call to a Human: Routing Rules That Actually Work goes deep on the decision logic behind smart call routing.

AI phone software

The episode walks through four categories of routing triggers and the baseline metrics every deployment should establish before writing a single rule:

• Set a baseline first: Industry guidance puts a healthy escalation rate at 15–30% of AI-handled conversations — sustained rates above that usually mean the agent's scope is too wide. Before any rule is written, teams should lock in an expected escalation rate, real human availability hours, and a maximum wait time before offering a callback.
• Hard categories that never self-serve: Active emergencies, complaints and refund demands, legal or medical advice, and any call involving press or regulators should route to a human immediately — regardless of how capable the agent sounds. Phony.ai's human handoff tools are built to handle these transfers cleanly, with context preserved.
• Caller signals mid-conversation: Four signals carry most of the weight — an explicit request for a person (transfer on first mention, no retry), two consecutive turns of repeated misunderstanding, detectable emotional intensity, and prolonged dead air after a direct question.
• System-side signals the caller can't see: Tool failures like a downed calendar API, low speech-to-text confidence on a critical field after two attempts, and missing or expired consent all warrant routing rather than guessing.
• Time and patience thresholds: Research shows more than half of callers hang up within eight minutes — the equivalent for a voice agent is time spent on a single intent without resolution. Setting per-intent ceilings (e.g., 90 seconds for a reschedule) and transferring when hit prevents silent abandonment.
• Business-hours and repeat-caller rules: The same intent can route differently depending on time of day, and a second call from the same number within two hours almost always signals the first interaction failed — route it before the agent retries.

The episode also touches on how these rules apply differently across industries — a dental scheduling line will sit at the low end of the escalation band, while a law firm intake line fielding distressed callers should expect to sit higher and should plan accordingly. The underlying principle throughout is that a routing rule pointing to a human who isn't there isn't a routing rule — it's a dead end. For more on how call routing and numbers can be structured to support these decisions, see Phony.ai's numbers and call routing platform features.

For more on the business case for keeping AI agents in their lane, listen to AI Voice Agents Are Now a Commodity — The Value Has Moved.

Phony.ai

Here's what the episode covers:

• The fifty-vendor signal: When more than fifty competitors enter the same space in a single year, it's evidence that the technical barriers have fallen — not that the market is thriving at the frontier. Meanwhile, the infrastructure layer underneath has consolidated to a handful of providers, the classic shape of a commodity shakeout.
• Pricing convergence: All-in per-minute rates across developer and managed platforms have compressed dramatically. The spread between the cheapest and most expensive honest quote is now roughly five times — compared to fifty times just a few years ago — and the cost of running a GPT-3.5-class model has fallen more than 280x in under two years.
• Latency as a non-differentiator: The gap between the fastest and slowest platforms now sits at around 150 milliseconds — a margin no caller would reliably notice. What once separated a "live" agent from a phone tree is no longer a meaningful competitive edge.
• Where the value actually lives — integrations: A voice agent that can't read a live calendar, write to a CRM, open a job ticket, or transfer a call with context intact is a sophisticated answering machine. The operational wiring is what determines whether a deployment earns its place in a business's workflow.
• Smarter evaluation questions: Instead of leading with a voice demo, buyers should ask vendors to name their integrations (not middleware bridges), demonstrate the agent's graceful failure path, itemise per-minute costs by component, and state scope and knowledge base limits explicitly before contract.
• Commoditisation as progress: The episode reframes the commodity shift as a natural and healthy milestone — the interesting engineering problems have simply moved up a layer, into scheduling logic, dispatch systems, billing flows, and handoff rules.

For more on the cost side of AI phone deployments, listen to What a Dental AI Phone Agent Actually Costs, Line by Line, which walks through a real budget breakdown for a specific practice type.

Phony.ai

Here's what the episode covers:

• Why flat-rate pricing is a packaging decision, not a cost: Subscription tiers between $300 and $800/month bundle costs that can — and should — be evaluated line by line.
• How patient call behavior shapes the bill: Call volume isn't smooth — it spikes at open, surges at lunch, and keeps coming after close. Roughly 35–38% of inbound calls go unanswered, and that's exactly where a voice agent earns its keep.
• The actual invoice for a two-chair practice: At ~1,500 agent-handled minutes, the all-in monthly total comes to just over $205 — covering phone number rental, carrier minutes, agent operating costs, call recording, and SMS confirmations.
• Scaling to a four-chair practice: ~2,600 minutes works out to roughly $355/month — still well under the entry tier of most flat-rate competitors once setup fees and practice-management surcharges are factored in.
• The four variables that actually move the total: After-hours share, voice vendor tier, speech-to-text accuracy, and human handoff rate. Everything else is noise. Notably, a higher-accuracy speech model that recognizes dental terminology can cut average call length by 10–20 seconds — often paying for itself several times over. For more on how human handoff affects both cost and patient experience, the episode breaks down what a healthy handoff rate (5–15%) looks like in practice.
• Costs that live outside the per-minute bill: Three additional expense categories are flagged that operators should account for separately when building a complete picture of total cost.

If you're evaluating AI phone agents for a dental practice — or scrutinizing any vendor's pricing — this episode gives you a reusable framework for reading between the lines of any proposal. For more on the contracts side of this decision, check out the related episode The Math Most Buyers Skip Before Signing a Voice Agent Contract.

Phony.ai

Here's what the episode covers:

• Start with 90 days of call data. Pull four columns from whatever system the front desk already uses — timestamp, direction, duration, and outcome — then split the file into three buckets: business-hours inbound, after-hours inbound, and outbound confirmations. Never blend them into a single average.
• Average handle time is the biggest lever. A dental front desk averaging 3.2 minutes per call across 420 monthly inbound calls bills 34% more minutes than the same volume at 2.4 minutes. At twelve cents a minute, that's a $40 monthly swing driven entirely by how promptly the agent wraps up after a booking is confirmed.
• Four inputs do most of the work. Monthly answered call count (not offered — answered), per-bucket handle time, handoff rate, and retry/callback behaviour. Each is weighted separately; collapsing them inflates or deflates the estimate in ways that only become visible on the invoice.
• After-hours share is the most underweighted input. A business with zero after-hours coverage today has zero after-hours minutes in its current log — but a 24/7 AI system will generate hundreds of new billable minutes there, alongside the new revenue those calls can capture. Both sides of that equation belong in the budget. For businesses like roofing companies or dental practices managing appointment scheduling around the clock, this line item can be the largest single variable in the estimate.
• A worked example ties it together. A two-location dental practice with 1,180 answered inbound calls, 240 abandoned calls, and 95 manual outbound confirmations lands at roughly 3,896 billable minutes per month — about $467 for the AI layer before carrier cost. Sensitivity testing shows how a tighter prompt, a high handoff rate, or adding after-hours coverage each shift that figure.
• Billing increment math matters. Per-minute rounding on a workload with short calls can inflate invoices by 30–50% at an identical headline rate compared with per-second billing. The rate card number and the effective rate are not the same thing.

The full estimation framework, including the bucket breakdown and sensitivity table, is in the source article linked above. If you're weighing a platform switch alongside this financial exercise, the episode How to Switch Voice Agent Vendors Without Losing a Number or a Booking picks up where this one leaves off.

AI phone software

Here's what the episode covers:

• Port the number first, before anything else. The phone number is the one asset the losing vendor controls, and the porting clock is set by the losing carrier — not by you. Local numbers can move in as little as one business day once accepted, but rejections are common; the account name on your letter of authorization must match the billing record character for character. Toll-free numbers move through the Somos registry via a RespOrg change and can take up to 21 business days, making them the longest dependency in any migration.
• Rebuild the agent prompt — don't paste it. The routing rules, persona, business hours, and after-hours behavior are a half-day of writing on any modern platform. The old prompt carries the failure modes of the old vendor; rewriting from scratch is faster and cleaner than inheriting those assumptions. Building a voice agent from a clean brief produces a better result than porting a legacy configuration.
• Calendar and CRM reconnections are manual and can't be skipped. Every OAuth grant is scoped to the old vendor's client ID. The new vendor has to re-request consent with its own credentials, one account at a time. Skip this step and bookings simply don't land.
• Run a parallel week on a secondary line, not a hard cutover. Port a spare marketing or department number to the new agent first. Main line stays put. Measure booking completion rate, human handoff rate, and appropriate deflection rate across both agents against the same call mix before any traffic shifts.
• Export recordings and transcripts before cancelling. Most vendors delete everything within 30–90 days of account closure. Keep raw audio and structured outcome fields separately — a transcript has no value until the booking, callback request, or complaint has been extracted from it.
• Know what actually breaks on cutover day. Rarely the port itself. What breaks is downstream: calendar tokens go silent without alerting anyone, recording consent disclosures drop if the new agent isn't yet configured, and outbound caller ID shows "Unknown" for up to 72 hours while carrier name databases update. Each of these has a concrete fix — but only if you know to check.

If you're still evaluating vendors and haven't committed to one yet, the episode 14 Questions to Ask a Voice AI Vendor Before You Sign Anything is the right place to start before returning to this one. Both episodes are available wherever you listen to Phony.ai.

Phony.ai

The episode covers five critical evaluation areas every buyer should probe:

• Telephony ownership and number portability — Whose carrier account does your traffic run on, and can you port your numbers out within 48 hours if you leave? STIR/SHAKEN A-level attestation isn't optional; without it, outbound calls can be blocked before a phone ever rings.
• Consent enforcement as a hard control — Following the FCC's February 2024 ruling, AI-generated voice calls fall under the TCPA's robocall consent standard. The right platform blocks the call when a call recording consent record is missing — not a warning, a hard stop — and propagates mid-call opt-outs to every downstream campaign automatically.
• Model transparency and hallucination rates — A vendor who can't name the model powering your calls can't explain your risk. Stanford HAI's 2024 data puts hallucination rates for smaller models as high as 27.8% on specialised benchmarks, versus 1.5% for frontier models. Knowing which model runs your calls — and whether the agent is grounded in a structured knowledge base — is non-negotiable.
• Human handoff triggers and warm transfer quality — Production agents break not in demos but in real calls with frustrated callers. A warm transfer with a context summary is the baseline; specific, codified triggers (explicit request, detected distress, three failed intents, out-of-scope topic) signal that a vendor has actually learned from real-world failures.
• Pricing legibility and bring-your-own flexibility — A bundled per-minute rate hides cost structure. Buyers should ask for a layer-by-layer breakdown and ask explicitly what happens to pricing if they bring their own model API key or carrier account. A platform that treats bring-your-own as first-class will discount; one that doesn't will deflect.

The episode closes with what the source article calls the best question to save for last: ask the vendor for the last three failure modes they fixed in production. Real bugs, real customers, real fixes. A vendor with production experience will have a specific list. One without will change the subject. With Gartner projecting that over 40% of agentic AI projects will be cancelled by 2027 due to inadequate risk controls, the buyers who survive will be the ones who asked the boring questions first. For more on how platform architecture shapes these trade-offs, check out the companion episode All-In-One vs. Build-Your-Own: The Real Cost of a Voice Agent Stack.

Phony.ai

The episode works through the full decision, covering:

• What's inside a flat per-minute rate — all-in-one vendors typically charge 25–50¢/min by averaging four separate cost layers: carrier, speech-to-text, AI model, and voice synthesis, which can total as little as 5–15¢/min on a self-built stack, as explored in the source article on all-in-one vs. build-your-own voice stacks.
• The hidden subsidy problem — on high-volume, well-scoped calls, a flat rate means short, efficient calls quietly cross-subsidize a vendor's longest, most expensive edge cases.
• The real cost of assembly — enterprise teams running their own pipelines often land at $40K–$70K/year once engineering time is properly counted; the all-in-one vendor is selling the removal of that work.
• Four places lock-in actually lives — phone number ownership, prompt and flow portability, transcript and event log export, and the integration rebuild cost that never appears in either quote; platforms that make call transcripts and event logs fully exportable give you meaningful control over your own training data.
• Compliance as a product feature, not a policy PDF — since the FCC's February 2024 TCPA ruling on AI-generated voice, the real question is whether the platform enforces consent and disclosure rules in the product itself, not whether a document exists.
• Latency benchmarks that actually matter — p50 latency (~680ms) is the number vendors want to show; p95 (~1,200ms) is the one that covers your worst calls, and callers notice awkward pauses above 800ms.

The episode also walks through how to read two competing quotes side by side — not just for the per-minute number, but for contract exit costs, data portability, and what each vendor's performance claims look like under real production load. For more on what goes into a minute of AI call cost specifically, the episode What a Minute of AI Phone Call Actually Costs in 2026 is a natural companion listen.

Phony.ai

The episode walks through each cost layer in turn and explains why the gap between advertised and actual rates is almost never accidental:

• Telephony is cheap — until it isn't. The raw carrier leg costs roughly a penny per minute, but platforms routing calls through Twilio's Conversation Relay add up to seven cents per minute on top of that, a charge most bundled bills never surface.
• Speech-to-text is close to a rounding error. Streaming transcription from providers like Deepgram runs well under a cent per call minute and is rarely where budgets blow out.
• Text-to-speech is where costs climb fast. Premium voice synthesis can run around eight cents per minute at standard concurrency — and double to sixteen cents when simultaneous-call limits are breached during busy periods. Silence billing (metering hold time and pauses) can quietly add another 15–30% on top.
• The model layer carries the biggest swing. Depending on whether a call uses a flagship real-time LLM or a lighter text-mediated pipeline, the model cost alone can range from two cents to well over fifteen cents per minute — a variation that bundled pricing makes completely invisible to buyers.
• Stacked together, an honest mid-range minute costs more than most ads suggest. Combining all four layers realistically lands a delivered minute somewhere between fourteen and twenty-two cents, with model choice explaining most of the spread. Rates below that range are typically a promotional subsidy, a concurrency trap, or a thinner pipeline; rates above it usually reflect compliance add-ons such as HIPAA requirements.
• Five questions to ask before signing anything. The episode closes with the specific written questions buyers should demand answers to — including which carrier ran the leg, which underlying model providers are in use, and whether you can swap to a lighter model for simpler call intents.

Whether you're evaluating a bundled platform or pricing out your own stack, understanding what each layer actually costs is the only way to hold a vendor accountable — or to know when a low quote is a gift and when it's a clause waiting to activate. For more on a related dimension of AI call design, check out the earlier episode Barge-In Is Not a Feature — It's a Promise You Have to Keep.

AI phone software

Here's what the episode covers:

• What barge-in actually requires — detection, playback interruption, and speech recognition all have to work in concert, not just independently.
• The false-positive problem — poor echo cancellation on speaker-mode calls can cause the agent's own voice to trigger the barge-in detector, making the agent seem fragmented and erratic.
• The buffer-drain failure — even when detection fires correctly, a poorly engineered audio flush path means the agent keeps talking for several hundred milliseconds after it should have stopped — long enough for the caller to finish their sentence first.
• The lost-words problem — recognizers that only activate after playback ends will silently drop the first words of every interruption, leaving the agent confused and the caller repeating themselves.
• The fix that costs more but works — running the speech recognizer continuously on the caller channel, even while the agent is speaking, so a partial transcript is already buffered when barge-in fires.
• Two diagnostic questions to ask any vendor or internal team before going live: the end-to-end latency from VAD signal to audio silence, and whether the recognizer is running continuously or only post-playback.

The episode frames barge-in as a measurable engineering promise rather than a toggle — and explains why teams that treat it as a feature flag tend to discover its failure modes only after real callers are on the line. If you're testing a voice agent before deployment, these are the exact scenarios worth scripting into your test cases. For a closer look at how per-minute costs shift when you add continuous recognition overhead, the Phony.ai blog's breakdown of AI phone call per-minute costs is a useful companion read. And because barge-in failure often ends with a caller demanding a human, it's worth having a well-designed human handoff path ready for when the agent falls short.

More from the show: if this episode got you thinking about what happens at the edges of a call, Transfer or Terminate: Designing the Handoff That Doesn't Drop the Caller covers the equally high-stakes moment when the agent has to exit the conversation gracefully.

Phony.ai

Here's what the episode covers:

• Transferring vs. completing: Why a raw SIP transfer is closer to a reset than a handoff — and why callers who have to repeat themselves lose trust immediately.
• Context delivery methods: The difference between a whisper message (audio played to the human agent before the call connects) and a screen pop (a structured data push to the agent's interface), and why neither works automatically just because both systems share an account.
• The integration gap: When an AI layer and a contact centre platform are separate services, context doesn't travel with audio unless someone explicitly builds the bridge — a design step that routinely gets skipped during prototyping. Phony.ai's human handoff tooling is built to close exactly that gap.
• Transfer timing: Why triggering a transfer the instant the condition is met can feel worse than a short delay — and how a simple three-beat spoken sequence (acknowledgement, hold, connection) makes a transfer feel intentional rather than like a system crash.
• Failure paths: What must be designed before go-live for every scenario where the destination doesn't answer — full queues, after-hours calls, and silent voicemail drops — including callback offers and structured voicemail with context already embedded. Thinking through call routing and fallback rules at the infrastructure level is what separates a reliable deployment from one that compounds the caller's problem.
• The takeaway framing: Design the transfer as the last impression the AI makes, because for most callers, it is. A smooth handoff with no repeated questions and a prepared human is invisible. A bad one is memorable for all the wrong reasons.

If this episode got you thinking about the latency and architecture questions that live just underneath the telephony layer, the earlier episode Why Sub-Second Voice AI Is an Architecture Problem, Not a Speed Problem is a natural companion. And if you're evaluating what it costs to run calls end-to-end once the handoff logic is built, the Phony.ai breakdown of AI phone call per-minute costs is worth reading before you finalize your deployment model.

Phony.ai

Here's what the episode covers:

• Why latency comparisons are almost meaningless. Vendors quote response times without defining the measurement window, making cross-vendor benchmarks impossible to interpret with confidence.
• The four-leg pipeline every caller pays for. Endpointing, transcription, model inference, and speech synthesis are sequential — and three of those four legs barely move regardless of how fast the model is.
• The honest floor for endpointing. Pushing the silence threshold below ~300 milliseconds causes the system to clip callers mid-sentence; the clipped fragment transcribes to nothing, and the caller experiences being ignored — a worse outcome than a longer wait.
• Transcription as the current bottleneck. On the web path, transcription accounts for roughly 830 milliseconds and is the single largest contributor — a constraint shaped by the fact that batch transcription can't begin until speech ends.
• The model is the one leg that's actually free to change. A small, fast voice-tuned model comes in around 690 ms to first token; a frontier model runs closer to 2.5 seconds — a factor of ~3.5× on the one variable that isn't constrained by physics or architecture.
• Two architectural moves that actually shift the floor. Speculative execution (generating against a predicted complete utterance and discarding if the caller continues) and speech-to-speech models that collapse the pipeline entirely are the levers that matter — not prompt caching or reasoning-mode toggles, both of which the team investigated and found innocent.
• Geography is a variable, not a footnote. Network distance between a host and a provider can swing results by 150 milliseconds. Any latency figure without a region attached is really a figure about one specific data centre.

The episode closes with a practical challenge: the next time a vendor quotes a sub-second response time, ask exactly what was measured between. Even with a hypothetical zero-latency model, endpointing, transcription, and synthesis still sit in the chain — and they don't disappear with a faster GPU. For more on how Phony.ai approaches product decisions around caller experience, check out the earlier episode Why There's No List Upload — And Why That's the Whole Point.

Phony.ai

The episode covers a lot of ground for a design decision that looks, on the surface, like a missing feature:

• Consent lives in the record, not the file. A CSV strips away the provenance of every number it contains — once it arrives, there's no reliable way to distinguish a genuine opt-in from a scraped or purchased contact.
• Three and only three paths to callability. A number becomes dialable through an inbound contact event, a form submission (captured with page-state context), or an active existing relationship — anything else is architecturally excluded, not policy-excepted.
• The honest counterargument. Plenty of lawful outbound calling does involve lists — legitimate newsletter audiences, properly disclosed partner arrangements, airtight conference sign-ups. The episode doesn't brush this aside; it acknowledges the real cost of refusing those too, and explains why the platform accepts it.
• Carrier trust is a shared resource. High-volume calls to unconsented numbers get phone numbers flagged, and flagged numbers degrade answer rates — and inbound reliability — for everyone else on the same infrastructure.
• What this design doesn't do. It doesn't make any caller compliant. Telemarketing obligations sit with the business placing the calls; the architecture just makes one specific failure — calling someone with no evidence of permission — hard to reach by accident rather than easy by default.
• The operational controls that do exist — consent gating, DNC scrubbing, account-wide opt-outs, per-tenant database isolation, audit logging, and retention ceilings — are listed in full, with no third-party certifications quietly implied.

The episode closes by connecting the list-upload decision to a broader design principle: an agent that won't place a call it can't evidence consent for is the same agent built to know when to stop mid-conversation rather than push past the point of a necessary human handoff. The refusal isn't a gap in the product — it is the product. For more on how Phony.ai's billing model reflects similar transparency principles, listen to The Itemised Receipt: Why Your AI Phone Bill Is Lying to You.

AI phone software

Here's what the episode covers:

• The model-cost assumption is usually wrong. Teams instinctively treat the language model as the big-ticket item, but on a typical short call, it's the clock-billed layers — telephony and speech-to-text transcription — that dominate the invoice.
• Clock billing vs. token billing changes everything. Telephony and transcription run for every second a call is live; the LLM is billed by tokens per turn. Understanding which layers bill by the clock completely reframes where cost reduction actually lives.
• Call length is a conversation design problem, not a procurement problem. If wall-clock layers are the real cost driver, swapping in a cheaper model won't move the needle — but tightening the dialogue, cutting dead air, and triggering handoffs on time will.
• Latency has a price tag. Silence while an agent waits for a slow response is still being billed by the carrier and the transcriber. Sluggish orchestration isn't just a quality issue — it shows up directly on the receipt.
• Five-line receipts didn't actually help buyers predict costs. Phony.ai launched with fully itemised billing and a bring-your-own-keys model, then pulled back: splitting the AI layers into five separate numbers helped explain a bill after the fact but made upfront comparisons harder — and price arithmetic reads as a more expensive price whether or not the maths works out that way.
• Telephony stayed itemised for a reason. International rates vary too much to bundle honestly, and customers with their own Twilio accounts or SIP trunks can connect them directly — meaning carrier costs can bypass the platform invoice entirely, with no markup added.

The episode closes on a point that applies well beyond pricing design: a blended rate that buries the clock-billed layers will keep directing engineering and procurement effort toward a cheaper model, when the real lever is a shorter, tighter, better-designed call. The receipt isn't just an accounting document — it's a strategy document. More from the show: if this episode made you think about the fundamentals of how AI voice platforms handle calls, Buying a Number and Answering a Call Are Two Different Things is a useful companion listen.

Phony.ai

Here's what the episode covers:

• Provisioning vs. voice are entirely different integrations. Buying a number involves a handful of API calls. Handling an inbound call means implementing a carrier-specific state machine in real time — they are not the same column on a feature matrix, even when a platform treats them that way.
• Carrier support is a claim that deserves scrutiny. A carrier can appear on a "supported" list for provisioning purposes while having zero functioning inbound voice path — and nothing in the UI will tell you that until a real caller experiences the silence.
• Silent failures are the norm, not the edge case. When a number is assigned through a carrier the platform hasn't fully implemented for voice, every screen reports success. The failure is invisible by design — until it isn't.
• Four questions worth asking in any demo, right then, out loud: Which carriers can actually answer a live call (not just sell a number)? What happens when the model errors mid-call? Can the caller interrupt the agent mid-sentence? And what's the platform's data retention policy — with specifics?
• Barge-in is a structural test, not a prompt problem. A system that won't stop talking to listen is revealing something about its architecture that better instructions cannot fix.
• The honest carrier matrix looks narrower. A platform that counts provisioning and voice separately may appear to support fewer carriers than a competitor — but the accurate comparison is the one that matters when the phone actually rings.

If this episode made you rethink what "supported carriers" actually means, you might also appreciate The Hard Part Is Knowing When to Stop Talking, which examines another structural challenge that can't be patched with a better prompt. More from Phony.ai on the questions worth asking before any AI voice deployment.

Phony.ai

Here's what the episode covers:

• Why keyword-based escalation fails the callers who need it most — the polite, frustrated caller who repeats themselves without ever hitting a trigger word, and the cost of catching frustration too late.
• The silent failure mode — how an agent can sound confident and fluent while answering the wrong question entirely, and why that's invisible from the dashboard but consequential for the caller.
• Repetition as a threshold — if a caller states the same intent twice, the first answer didn't land; why two repetitions is the moment to act, not three.
• Empty retrieval as a stop signal — agents that always return a best guess will eventually return a confident wrong answer; the fix is building a retrieval layer that can surface uncertainty and hand off before that answer gets spoken.
• Irreversible decisions as a hard boundary — money, medical, legal, and cancellation topics warrant automatic escalation not because AI accuracy is necessarily poor, but because the downside of being wrong is asymmetric.
• What a proper handoff actually carries — transcript, understood intent, what couldn't be answered, and crucially the reason the agent stopped; why that last field is the one teams leave out and the one that makes the difference.
• The out-of-hours trap — an agent should never promise a transfer it can't complete; when no one is available, acknowledgment and a committed callback beats a phone ringing in an empty office.

The throughline is straightforward: an agent that only stops when told to stop is an agent that won't stop when it most needs to. Designing for the trigger word is designing for the easy case. The real work is in the conditions nobody says out loud.

Phony.ai