The hard part of AI cold calling is not the voice

The short version

• An AI cold calling bot is four bought parts plus the orchestration between them: transcription, turn detection, a language model, and speech synthesis over a phone line.
• Turn detection decides whether a call feels human, not voice quality. The silence you wait before replying is a direct trade between interrupting people and sounding slow.
• Compliance controls belong in the dialler and the database, not in the prompt. A model asked to honour a do-not-call request will usually honour it, and usually is the wrong standard.
• Cold lists are the hardest case for a voice agent. The same stack performs better on inbound enquiries, old leads and appointment confirmations.
• If the offer does not convert when a competent human delivers it, an outbound AI calling agent will only deliver a weak offer to more people.

What does an AI cold calling bot actually do on a call?

It runs one loop, over and over, on a phone line. Audio comes in and is transcribed. Something decides whether the person has finished their sentence or is just drawing breath. The transcript and the call history go to a language model. The reply is turned into audio and played back down the line. That loop is the whole product.

So AI cold calling software is an assembly: speech recognition, turn detection, a model, speech synthesis, and a telephony provider to carry the call. None of those parts is scarce. You can wire a working outbound AI calling agent together in an afternoon, which is exactly why every demo looks good. The components are commodity, and they improve every few months without you doing anything.

What is not commodity is the orchestration. What happens when the model takes two seconds. What happens when the prospect interrupts halfway through the second sentence. What happens when the line is noisy, when two people are in the room, when the call connects to an automated menu, when somebody says a thing the script never anticipated. That is where the gap between a convincing agent and an embarrassing one sits, and it is mostly engineering rather than prompting.

Why does latency matter more than the script?

Because people measure silence, and on a phone call they measure it closely. A gap of a second reads as a dropped line. The person fills it, the agent starts talking at the same moment, and now there is crosstalk that neither side handles well. No script survives that. A mediocre opener delivered with human timing beats a brilliant one delivered a beat late.

The delay is cumulative and mostly serial. The transcriber will not commit to a final result until it believes the speaker has stopped. The model cannot start until it has the transcript. Synthesis cannot start until it has text, though it can start on the first clause instead of the whole reply. The audio still has to travel. Each stage is quick on its own. Stacked, they are what the caller hears as hesitation.

The mitigations are all about starting earlier: stream partial transcripts, synthesise from the first sentence, keep the opening line pre-rendered, use a smaller model for turns that do not need a large one. What you cannot engineer away is the turn detection threshold. You do not know somebody has finished until some amount of silence has passed, and the length of that silence is a trade: shorter means you talk over people, longer means you sound slow. Pick it deliberately, then test it against recordings of real callers rather than your own voice in a quiet room.

What breaks first when the demo meets a real list?

The list. Before any conversation design matters, a large share of numbers will not reach the person you intended: disconnected lines, wrong numbers, switchboards, voicemail, mobiles answered in a car with the window down. The agent that was eloquent in testing now spends most of its attempts talking to an answering machine or a receptionist. On cold data those are not edge cases, they are the common path, and they need first-class behaviours rather than fallbacks.

The second thing to break is the set of questions a real stranger asks. Who is this. Is this a recording. How did you get my number. Why are you calling me at work. Take me off your list. Somebody who did not ask to hear from you opens with suspicion, and a model left to improvise will improvise, which is the failure you least want on a recorded line.

The third is the write-back. Calls happen, conversations go well, and nothing lands in the CRM in a shape anybody can act on: no disposition, no transcript against the right record, no task for the rep, duplicate contacts created because the dialler could not match a number. This is dull integration work. It is also the part that decides whether you end up with pipeline or with anecdotes.

Which behaviours should be defined before the first real dial?

• •Is this a robot? Answer honestly, in one short sentence, and carry on.
• •Take me off your list. Write to a suppression table immediately, confirm, end the call. The model does not get to treat this as an objection.
• •Voicemail detected. Leave one defined message or hang up. Never let the agent hold a conversation with a beep.
• •A switchboard or keypad menu. Decide in advance whether you navigate it, ask for a named person, or drop the call.
• •How did you get my number? A true, boring answer about the source, available to the agent as data rather than invented on the spot.
• •Call me later. Capture the time, confirm it, schedule it. A promise the system cannot keep is worse than a refusal.
• •Price, contract terms or product claims. Narrow the agent to what it may state, and hand off rather than guess.

Do you have to tell people they are talking to an AI?

Assume yes, disclose it in the opening line, and build as though the strictest rule applies to you. The honest position on the law is that it varies by country, varies between US states, and has been moving quickly. Disclosure of artificial voices, consent to record, registered do-not-call lists, permitted calling hours and what counts as marketing all differ. We build the controls. Your own counsel decides what your markets require. Anyone telling you the rules are simple is selling something.

The useful part is that the controls look much the same whichever rule set applies, and they belong outside the model. Suppression is enforced before the dialler places the call, not by a prompt asking the agent to be careful. An opt-out triggers a database write and a confirmation, not a sentence. Calling windows are enforced by the scheduler against the prospect's own time zone. Disclosure is a fixed string in the opener that the model cannot rephrase away. Recordings have a retention period and an access rule.

The reason to put these outside the prompt is not tidiness. A model asked to honour a do-not-call request will usually honour it. Usually is the wrong standard for the one instruction that carries a penalty. Anything that must be true on every call should be true because the code made it true, not because the model agreed.

Where does outbound AI calling actually earn its keep?

At the warm end, not the cold end. Calling an inbound enquiry within a minute of the form being submitted is the most defensible use of an outbound AI call agent: the person asked to be contacted, the call is expected, and speed is something software is genuinely better at than a rep with a full calendar. Reactivating old enquiries comes next, because the list is yours, the context exists, and nobody was going to call those rows anyway.

Pure cold calling is the hardest case, and it is the one most of the marketing is about. It can work for a concrete offer that explains itself in two sentences and qualifies with three questions. It works badly for a consultative sale where the value of the call is the judgement of the person making it. If your reps' conversations are the product, automating the opening of them is a narrow win at best.

There is also a category where this is not worth building. If you have forty named accounts, call them yourself. If the conversation requires regulated advice, do not put a model in it. If one clumsy call permanently damages a relationship you cannot replace, the expected value is poor however good the agent is. And if the offer does not convert when a competent human delivers it, an AI voice sales agent will only deliver a weak offer to more people, faster.

Which outbound jobs suit an AI agent for outbound calls?

How do you know it works before you point it at the whole list?

Grade recorded calls against written criteria, and keep the set. Choose the behaviours that must hold: did it disclose, did it honour the opt-out on the first request, did it state only what it is permitted to state, did it book the slot it said it booked, did it end the call cleanly. Score real transcripts against those, including the ugly ones, and replay the whole set after every prompt, model or voice change. Without that, every change is a guess and regressions arrive silently.

Then measure the funnel at the list level rather than the call level. Connected, conversation held, outcome agreed, appointment actually attended, and what happened after. Completed calls is a vanity number: an agent can complete a thousand of them and produce nothing. Meetings held and qualified is the figure worth putting next to the cost.

On cost, the structure matters more than the per-minute rate. You pay for transcription, the model, the voice and the telephony leg, and you pay for attempts that reach nobody as well as the ones that convert. Your effective cost per booked meeting is therefore driven mostly by connect rate and list quality, which are properties of your data rather than of your vendor. Work it out on your own numbers before believing anyone's pricing page.

Should you buy AI cold calling software or build your own?

Buy the call layer. Build the edges. The speech stack is commodity and improving faster than any individual team can track, and the platforms selling outbound voice have put real engineering into turn detection and telephony reliability. That work is dull, specific, and not a differentiator for you.

What deserves your own engineering is everything a vendor cannot know: which records get called and in what order, what the agent may say about your product, suppression and consent state, write-back into your systems, handover to a human at the right moment, and the evaluation harness that tells you whether a change helped. That is integration and policy work. It also survives a change of voice vendor, which you will make.

The decision to take early, before any tooling, is what a successful call means and who picks up the result. Teams that cannot answer that end up with a working AI cold calling bot and no process waiting at the other end of it. That looks like a technology failure, and it is not.