AI SDR: Where It Holds Up and Where It Breaks, From People Who Build AI Voice

Almost everything written about the AI SDR comes from one of two places: people selling one, or people who have never had to make one work. We are in a third position, and it is worth saying up front because it is the whole basis for this article. SalesArmor runs live voice conversations between salespeople and an AI buyer every day. We do not sell an AI SDR. But we build the same machinery one runs on — speech in, a model deciding what to say, speech out, in real time — and we have spent a long time finding out exactly where it fails.

So this is not a verdict on whether SDR AI "works". It is a map of which parts of the job it does well, which parts it does badly, and the mechanism behind each, so you can judge a specific product against your specific use rather than against a demo.

The short version: the jobs where an AI SDR holds up are asynchronous, and the places it breaks are synchronous. Almost everything else follows from that.

What an AI SDR actually is

"AI SDR" is sold as one thing and is at least three:

• An AI email and messaging agent that researches accounts, writes and sends outbound sequences, and replies to responses.
• An AI appointment setter that answers inbound enquiries — by chat, text or phone — qualifies them and books a meeting on a human's calendar.
• An AI voice agent that places or takes live phone calls and holds the conversation itself.

They share a label and very little else. The first never has to speak. The second mostly works in text, with a phone option. The third is doing the hardest thing in the category: holding an unscripted spoken conversation with a stranger who did not ask to talk to a machine. When a vendor says "AI SDR", the first useful question is which of the three they mean, because the evidence for one says almost nothing about the others.

The four jobs where it holds up

1. Speed to lead

An inbound enquiry is at its warmest in the minutes after it is submitted. The person is still at their desk, still thinking about the problem, and often filling in the same form for two of your competitors in the same sitting. The first response to reach them frames the conversation, and the others arrive to someone who has already had it.

Humans are bad at this for structural reasons, not effort: they are in meetings, asleep in another time zone, or working a queue in order. An AI that replies in ninety seconds with a reasonable message beats a human who replies the next morning with a better one — not because the AI is better at selling, but because it is there. This is the most defensible job in the category and it is almost entirely asynchronous.

2. Qualification against explicit criteria

If you can write down what makes a lead worth a human's time — company size, role, timeline, budget authority, a specific use case — an AI can ask for those things consistently and route on the answers. It does not skip questions because it is tired, and it does not talk itself into a weak lead because the call was friendly.

The limit is in the word explicit. It is exactly as good as the criteria you wrote. If your best reps qualify on something they could not articulate — a tone, a hesitation, a detail that reminds them of a deal that went bad — the AI will not find it.

3. Follow-up persistence

Most sequences die of neglect rather than rejection. A human sends two follow-ups and moves on to newer leads; the fourth and fifth touches, which are often where a reply finally comes, never go out. An AI does not get bored or embarrassed and does not deprioritise last month's list. For a job that is mostly remembering to do the next thing, persistence is the product.

4. Scheduling and the administrative tail

Proposing times, handling time zones, rescheduling, sending the confirmation, logging the activity in the CRM. None of this is selling and all of it consumes SDR hours. Automating it is uncontroversial, and it is where an AI appointment setter earns its name.

Notice what the four have in common. None of them requires holding a live conversation with someone who can interrupt.

The four places live voice breaks

This is the part we can speak to from our own engineering, because each of these is a problem we have had to solve, work around, or accept.

1. Interruption and overlap

Human conversation is not turn-based. People talk over each other constantly — starting before the other person has finished, murmuring "right" and "mm-hm" while listening, cutting in to correct something. Real calls are full of overlap, and it is how people signal attention.

A voice AI decides who has the floor by detecting voice activity. When it hears speech while it is talking, it has to choose: stop, or carry on. In our system, when the model is interrupted we drop the audio it still had queued and keep the transcript of what it had said up to that point. That part is solvable. What is not cleanly solvable is that the detector cannot reliably tell a backchannel from a takeover. Tune it sensitive and it stops mid-sentence every time the human says "mm-hm", which sounds nervous and deferential. Tune it less sensitive and it talks over people who are genuinely trying to cut in, which sounds rude. There is no setting that behaves like a person, because a person is not deciding by volume.

2. Latency reads as tone

For a voice system to reply, it first has to decide you have finished. It can only know that by waiting for a stretch of silence — so every reply carries a built-in pause before the model has even started. Shorten the window and it cuts people off when they pause to think. Lengthen it and every reply arrives after a noticeable gap.

Humans do not have this problem because they do not wait for silence. They predict the end of your sentence from its grammar and intonation and start their reply before you finish. The machine is always a beat behind, and to a human ear a beat behind does not sound like processing. It sounds like hesitation, or evasion. The same half-second that is invisible in a chat window becomes, on a call, the moment a prospect decides they are talking to something that is not sure of its answer.

There is a difference between fast enough to be real and fast enough to be natural. A system can clear the first comfortably and still fail the second on every single turn.

3. The names it mishears

Speech-to-text is very good at common words and noticeably worse at proper nouns — people's names, company names, product names. On a sales call those are precisely the words that matter. A rep who calls the prospect by the wrong name, or mangles their company, has lost the call in a way no follow-up recovers.

We see this from the other side. The component that picks the coaching note after each of our users' calls is told, in so many words, that speech-to-text "mishears proper nouns constantly", and to treat any wrong-name mistake as a transcription artefact — because the name in the transcript is often not the name the rep said, and coaching someone on the microphone's error is worse than useless. In practice software that is a small correction. In an AI SDR it is the system itself believing it heard a name it did not, and then saying it back to the person.

4. The unexpected turn

An AI SDR is built around a path: open, qualify, handle a known set of objections, book. On the path, it can be good. Off the path — a joke, a complaint about an invoice from a different department, a question about something adjacent, a brush-off that means the opposite of what it says — it degrades quickly, because there is nothing to match the moment against.

The brush-off is the clearest case. "Send me some information" is almost never a request for information; it is a polite way to end the call, and good reps know how to answer it rather than comply with it (we wrote about handling "send me more information" separately). A system that takes language at face value will send the PDF and log a warm lead.

Keeping a voice model on instructions is harder than it looks even when the instructions are simple. Our AI buyer is sometimes told to apply one specific kind of pressure at one specific point in the call. Getting it to do that reliably took three independent layers: the instruction itself, written as a contract; the model reporting whether it had done it, with a correction sent mid-call if the deadline passed; and an evaluator confirming afterwards. Any one of those alone leaked. That was for a single behaviour, in a conversation where a slip costs nothing.

The accountability question nobody asks before signing

When a human SDR over-promises — a discount they cannot authorise, a feature on a timeline that does not exist — there is a person, a manager and a correction. When an AI SDR does it, the company said it. There is no individual to have misspoken.

This is not theoretical. In 2024 a Canadian tribunal held Air Canada liable for a bereavement-fare policy its website chatbot had described incorrectly, rejecting the argument that the chatbot was responsible for its own statements. The principle travels: what your agent tells a prospect is what your company told them.

Regulation is moving in the same direction. In the US, the FCC ruled in February 2024 that AI-generated voices count as "artificial" voices under the Telephone Consumer Protection Act, so outbound calls using them generally need the recipient's prior express consent. In the EU, the AI Act's transparency rules — people must be told when they are interacting with an AI system — apply from August 2026. Neither makes AI voice outreach impossible. Both make it something to clear with counsel before it is switched on, not after.

Practice versus substitution

Here is the argument from our own position, and we have an obvious interest in it, so weigh it accordingly.

Every one of the four failures above changes character depending on who is on the other end and whether they chose to be there.

In substitution — the AI standing in for your rep with a real prospect — the prospect did not ask to talk to a machine, every artefact of being synthetic costs trust, and every mistake is a statement by your company. The technology is being asked to pass as something it is not, and it gets caught on the interruption, the pause and the mispronounced name.

In practice — the AI standing in for the buyer so a rep can rehearse — the same properties become neutral or useful. The rep knows it is synthetic and does not care. A beat of latency reads as a busy buyer thinking, which buyers do. An off-path turn is exactly the thing the rep needs to learn to handle. A mistake costs nothing and the call can be run fifty times, against a buyer who can be made harder on purpose. Being synthetic is the feature: you cannot book a real CFO for fifty rehearsals.

That is why we build the second and not the first — and why, if you are evaluating both, it is worth asking which one your problem actually is. If leads wait too long for a reply, automate the reply. If reps lose the calls they do get, practise the call. We have written about how these practice systems work in what is a coaching bot and sales avatars for AI training, and about AI inside the rep's own workflow in AI guided selling. If you are comparing practice tools, that is a separate shortlist, and we keep one at the best cold call practice tools.

How to evaluate an AI SDR if you are buying one

Do not evaluate from the demo. Demos are the path. Evaluate off it.

1. Call it yourself and interrupt it. Cut in halfway through its opening. Then say "mm-hm" while it is talking and see whether it stops. You will learn more in two minutes than from the deck.

2. Give it a name it will mishear. Your own, if it is unusual, or a company name with an odd spelling. See what it says back to you.

3. Leave the path. Ask about something adjacent to the product, complain about something unrelated, make a joke. Then say "just send me some information" and see whether it recognises a brush-off.

4. Try to get a commitment. Ask for a discount, a feature by a date, a contract term. What it agrees to under pressure is what it will agree to with your prospects.

5. Ask whether you are talking to a machine. Hear what it says. Then check that answer against the rules where your prospects are.

6. Read the unselected transcripts. Ask for twenty consecutive real conversations from an existing customer's deployment, not a curated set. The highlight reel shows what the system can do. The consecutive run shows what it usually does.

7. Start asynchronous. If the case for it is speed to lead and follow-up, you may be able to get most of the value from the text channels, where the failures above mostly do not apply, before deciding whether live voice is worth the risk.

Common questions about AI SDRs

What is an AI SDR? Software that performs some or all of a sales development rep's job — researching accounts, writing and sending outreach, replying to responses, qualifying leads and booking meetings. The label covers three quite different products: AI email and messaging agents, AI appointment setters, and AI voice agents that hold live phone calls.

Can SDR AI replace a human SDR? For parts of the job, yes: fast responses to inbound leads, consistent qualification against written criteria, follow-up persistence and scheduling. For live, unscripted phone conversations with strangers, it remains weakest exactly where the job is hardest — interruptions, natural timing, proper nouns and anything off the expected path.

What is an AI appointment setter? An AI system that responds to inbound enquiries, asks qualifying questions and books meetings on a human's calendar, usually by chat or text and sometimes by phone. It is the most mature form of the AI SDR, because most of its work is asynchronous.

Why do AI voice calls sound unnatural? Mostly timing. A voice system has to wait for a stretch of silence to know the speaker has finished, so every reply starts a beat late, and it decides who has the floor by detecting sound, so it either stops for every "mm-hm" or talks over genuine interruptions. Humans predict turn endings instead of waiting for them.

Are AI SDR cold calls legal? It depends on where the recipient is and how the call is made. In the US, the FCC ruled in 2024 that AI-generated voices are "artificial" voices under the TCPA, which generally means prior express consent is needed for such calls. In the EU, the AI Act requires people to be told when they are interacting with an AI system. Get legal advice for your own markets before deploying.

Who is responsible if an AI SDR promises something it should not? The company deploying it, in practice. A tribunal has already rejected the argument that an AI agent is responsible for its own statements. That is why what the agent may commit to, and how that limit is enforced, belongs in the evaluation.

What is the best use of AI in sales development? The asynchronous jobs — replying to inbound leads immediately, persistent follow-up, qualification against explicit criteria and scheduling — where the known weaknesses of live AI voice do not apply. For the live conversations themselves, AI is currently more useful for practising them than for having them.

How should I test an AI SDR before buying? Call it and interrupt it, give it a name it is likely to mishear, take it off its script, try to extract a commitment it should not make, ask whether it is an AI, and read twenty consecutive real transcripts from an existing deployment rather than a curated selection.

A note on sources

This article contains no figures for meetings booked, reply rates, cost per meeting or pipeline generated by AI SDRs. Nearly all of the ones in circulation come from vendors describing their own best deployments, and the ones that do not rarely define what they counted as a meeting. The widely quoted speed-to-lead multipliers have the same problem at one remove — they come from older studies of web-form response times, run before any of these products existed — so this article describes the mechanism rather than repeating the number.

The engineering observations are from our own system: how interruptions are handled, the silence-detection trade-off behind response timing, the rule that excludes name errors from our scoring, and the layered checks needed to keep a voice model on a specific instruction. They describe a practice product, not an AI SDR, and we have tried to say only what transfers.

The Air Canada decision is Moffatt v. Air Canada (British Columbia Civil Resolution Tribunal, 2024). The FCC's ruling on AI-generated voices under the TCPA was issued in February 2024. The transparency obligations are in Article 50 of the EU AI Act. These are summarised for buyers, not offered as legal advice. The asynchronous-versus-live framing and the evaluation tests are ours.