AI Cold Calling Just Hit $0.05 a Minute. The Constraint Moves to Your List.

The cost floor under the phone channel just collapsed

On September 10, OpenAI shipped GPT-Live-1 into the API at $0.05 per minute for the front-end voice layer, with native telephony support. That single line in a pricing section is the most consequential thing to happen to AI cold calling this year, and most sales leaders read past it because it was buried in a week that also produced the Agents API, seven named Agentforce agents, and an Anthropic essay about agent swarms taking over the internet.

Here is why it matters. Until this month, an AI voice agent that could hold a real B2B conversation was a chained system: speech-to-text, then a reasoning model, then text-to-speech. Every handoff added latency, and every pause in the prospect's speech risked the bot talking over them. That stop-start rhythm was the tell. Prospects hung up not because the words were wrong but because the timing was inhuman.

GPT-Live-1 is full-duplex — it listens and speaks at the same time inside one model. OpenAI reports a 30 percentage point improvement on Full Duplex Bench over GPT-Realtime-2.1, with the largest gains in turn-taking latency and interactive behaviour. Language-learning platform Speak found in early evaluations that the model cut interruptions during learners' thinking pauses by almost 80% compared with turn-based systems. Yelp's CTO reported "meaningful improvements in call handling rates" on reservations and food orders, and noted that callers started speaking in fuller, more natural sentences — which is the more interesting observation, because it means the person on the other end stopped adjusting their speech for a machine.

For outbound teams the takeaway is not "AI can call people now." It could already. The takeaway is that the cost, the latency, and the uncanny-valley problem all moved at once, and the binding constraint on the phone channel is no longer how many dials you can afford. It is who you dial and why. That is a targeting problem, and it is the one most teams have not solved.

What AI cold calling actually costs per conversation now

Let us do the arithmetic honestly, because the vendor math floating around this week is fantasy.

The $0.05 per minute is the voice layer only. A working outbound voice agent also needs a backend reasoning model for anything beyond scripted turns, carrier minutes, a telephony provider, CRM writes, recording storage, and an orchestration layer. Call it $0.10 to $0.20 per connected minute all-in for a competent implementation, and more if you route every turn to a frontier reasoning model.

Cognism's State of Cold Calling 2026 report, built with WHAM on an analysis of over 200,000 calls, puts average call duration at 82 seconds and the average dials needed to reach a prospect at 1.55. So a connected conversation is roughly 1.4 minutes of voice, and about a third of your dial attempts never connect at all — those cost carrier pennies, not model minutes.

Against that, an entry-level SDR in the US averages $55,018 a year in base pay according to ZipRecruiter's August 2026 data, most falling between $42,000 and $61,000. Fully loaded with tax, tooling, management and ramp, $80,000 is a conservative annual number. A rep making 70 quality dials a day across 20 working days produces about 1,400 dials a month.

The honest read on that table: the cost per dial fell by roughly a factor of twenty, and the cost per good conversation did not fall nearly as much, because the thing that makes a conversation good was never the dialling.

Why cheaper dials will make the channel worse before it gets better

Every channel that has had its marginal cost crushed in the last three years followed the same arc. Cold email hit near-zero marginal cost, volume exploded, mailbox providers responded with authentication requirements and complaint-rate ceilings, and reply rates fell to a point where the average sits around 3% while disciplined senders still clear 10%. LinkedIn connection requests followed the same curve, which is why acceptance rates are now a vanity metric.

The phone is next, and it has a harsher immune system: carrier-level spam labelling, STIR/SHAKEN attestation, and a buyer who can hang up in two seconds at zero cost. Cognism's data shows the channel recovered to a 2.7% success rate in 2026, up from 2.3%, largely because lists got better and the reps who stayed were dialling reachable people. A flood of cheap AI dials into that same numbering plan is the fastest available way to undo that recovery.

So if you are modelling AI cold calling as "same conversion, twenty times the volume, twenty times the pipeline," you are modelling the one outcome that will not happen. Volume is the thing the channel is about to start punishing.

The three jobs AI voice agents are genuinely good at today

There is a version of this technology that creates real leverage this quarter. It is narrower than the demos and it is mostly not cold.

1. Speed-to-lead on inbound and hand-raisers

This is the unambiguous win. Someone books a demo, fills a form, or replies to a sequence, and a voice agent calls within sixty seconds to qualify and schedule. The call is expected, consent is cleanest, the conversation is short and structured, and the alternative is a rep who gets to it four hours later. Full-duplex matters here because hand-raisers interrupt constantly — they already know what they want and they talk over the script.

2. Warm callbacks on a genuine signal

A prospect viewed your profile, engaged with your post, commented on a competitor's thread, or posted about the exact pain you solve. A voice agent calling with a specific, verifiable reason — "you commented on X yesterday, is that a live problem for your team?" — is a different artefact from a cold dial. The opener does the work. This is where signal-based outbound and voice agents actually compound: the agent's weakness is improvised persuasion, and a strong signal means less persuasion is required.

3. The unglamorous middle of the funnel

Data verification, no-show recovery, meeting confirmations, reactivating closed-lost from last year, checking whether a champion still works there. None of this gets a conference keynote. All of it is repetitive, scripted, low-stakes, and currently eating hours of SDR time that could go to the fifteen accounts that matter. Run the voice agent here first and you will learn your failure modes on calls where a bad outcome costs you nothing.

Where it still fails

• Multi-threaded discovery. Anything requiring the agent to hold four stakeholders' competing priorities in mind and decide what not to say.
• Gatekeepers and switchboards. Unstructured, adversarial, and full of the exact edge cases prompts do not cover.
• Novel objections. The agent will generate something fluent and confident. Fluent and confident and wrong is worse than a rep saying "let me find out."
• Anything where being caught matters. And you will be caught — assume the prospect knows within fifteen seconds, and design the call so that is fine.
• Regulated buyers. Financial services, healthcare and public sector procurement teams are actively writing policies about agent-initiated contact right now.

The compliance wall nobody put in the pricing page

This is the part that should slow you down, and it is not optional. None of what follows is legal advice — take it to counsel before you dial.

The FCC's February 2024 Declaratory Ruling established that calls made with AI-generated voices are "artificial" under the Telephone Consumer Protection Act. That classification is the whole ballgame. It means AI voice calls inherit the TCPA's artificial-and-prerecorded-voice rules rather than sitting in some new unregulated category, and it applies even to non-marketing calls.

Practically, per current compliance guidance for voice AI builders:

• Prior express written consent is required for marketing calls using an AI-generated or prerecorded voice to mobile numbers. Signed, affirmative, specific to automated calling, and not conditioned on purchase.
• Identification at the top of the call — the calling entity by name plus a contact number or address — is already required for artificial-voice calls.
• A pending FCC NPRM would add an explicit AI-disclosure requirement in plain language at the opening of the call. As of mid-2026 it is not finalised, but treating it as twelve months out is the prudent planning assumption.
• State law is already ahead of the FCC. Texas SB 140 requires AI voice technology to be disclosed within the first 30 seconds of a call and bans cloning identifiable voices without consent. Other states are following.
• Statutory damages remain $500 to $1,500 per call. At AI volumes, that number compounds into something that ends a company, not a quarter.

The B2B escape hatch people reach for — "TCPA is about consumers" — is weaker than it looks, because the mobile number on your prospect's LinkedIn profile is a mobile number regardless of who pays the bill, and because multiple 2026 compliance playbooks now treat B2B mobile dialling as inside the risk envelope rather than outside it.

A practical checklist before your first AI-dialled call

• Disclose AI in the first sentence, not the first thirty seconds. Earlier than the law requires, because trust is the asset you are protecting.
• Maintain a consent record per number, with source and timestamp, that you could hand to a regulator without redaction.
• Scrub against the national DNC list and your own suppression list on every run, not on import.
• Check two-party recording-consent states before you record anything, and default to disclosed recording everywhere.
• Set a hard cap on dial attempts per contact and per day, and monitor your numbers for carrier spam labelling weekly.
• Give the agent an unconditional escalation path to a human, triggered by the prospect asking once.
• Log every call with full transcript, model version and prompt version. When something goes wrong you will need to prove what the agent was told.

What this does to the rest of your outbound motion

Three second-order effects are worth planning for now.

Answer rates will get defended. Expect carriers, handset makers and buyer-side AI assistants to screen aggressively. Your prospect's phone is about to have the same relationship with unknown numbers that their inbox has with unknown senders. The Salesforce Agentforce launch on September 11 put a long-horizon outbound agent into pilot the same week; buyer-side screening agents are shipping just as fast. The end state is agents negotiating access to human attention, and the currency in that negotiation is relevance.

Trust becomes the differentiator, not throughput. Buyer research keeps landing in the same place: 69% of B2B buyers say they want to validate AI-generated insights with a human rep. That is not anti-AI sentiment, it is a request for accountability. The teams that win will use agents to earn the human conversation, not to replace it — which means measuring your voice agent on qualified human conversations created, not calls completed.

The list becomes the entire competitive advantage. When dialling is free, everyone dials. What separates a 2.7% team from an 11% team in the Cognism data was never the dialer — it was ICP discipline, signal-based timing, and calling when something actually changed at the account. Cheap AI dials do not fix a bad list. They industrialise it.

This is the part that matters for how you spend the next quarter. If your plan is to point a voice agent at a purchased list, you are buying a faster way to burn your phone numbers and your brand. If your plan is to point it at people who have just shown intent — a profile view, a comment on a relevant post, a hiring signal, a public complaint about the tool you replace — you are buying leverage on a scarce resource. That is the motion Updately is built around: capture the warm signal, score it against your ICP, research the person properly, and then reach out with a reason that survives contact with a sceptical human. The channel you use for that reach-out matters less than whether the reason is real.

A 90-day plan for teams that want to move

Days 1–30: pick the boring use case. Deploy a voice agent on meeting confirmations, no-show recovery, or data verification. Instrument everything: connect rate, completion rate, escalation rate, and the share of calls where the prospect asked whether it was a bot. Get your compliance documentation built while the stakes are low.

Days 31–60: move to warm callbacks. Route only signal-triggered contacts — people who took an observable action in the last 72 hours — to the agent, with a human rep as the escalation target. Compare booked-meeting rate against your reps on the same signal type. If the agent is more than 40% below your reps, the problem is almost always the opener, not the model.

Days 61–90: decide what stays human. By now you will know which conversations degrade when the agent handles them. Give those back to reps permanently and stop trying to fix them with prompts. Redeploy the time you have freed into account research and multi-threading, which is where human sellers still win outright.

Run each phase with a named owner and a kill switch. An outbound voice agent is an autonomous system making contact with your market under your brand name — it deserves the same review gate as a production deploy, which most GTM teams still do not apply to anything.

Takeaways

• The price of an AI-dialled conversation fell roughly 20x. GPT-Live-1 at $0.05 per minute with telephony makes voice agents cheap enough that volume is no longer the constraint.
• Full-duplex is the real unlock, not the voice quality. Natural interruption handling is what stops prospects from hanging up in the first five seconds.
• Cheap dials will degrade the channel. Expect carrier screening, buyer-side filtering and falling answer rates as volume floods in. Plan for a worse channel, not a bigger one.
• Start with warm and boring. Speed-to-lead, signal-triggered callbacks, and funnel hygiene are where voice agents produce value today. Cold spray is where they produce complaints.
• The TCPA exposure is real and quantified. AI voices are "artificial" under the FCC's 2024 ruling, statutory damages run $500–$1,500 per call, and state disclosure laws are already in force. Get counsel before you dial.
• Your list is the moat. When everyone can dial infinitely, the only durable advantage is knowing which 40 people to call this week and why.

The teams that get value out of AI cold calling in the next twelve months will not be the ones who dialled the most. They will be the ones who had a reason to call.