Almost every AI SDR on the market is an email product. Some add LinkedIn. Practically none touch the phone, which is interesting, because the phone is where the highest-value outbound conversations still happen.
The gap is not an oversight. Voice is a much harder problem than text, and the ways it fails are considerably more embarrassing.
Here is where the technology actually is, and what the useful version looks like.
What AI voice can do now
Genuinely, not from a demo reel.
Handle a scripted qualifying call. Confirm a role, check headcount, ask two or three qualifying questions and route the result. This works. Latency is low enough that the pauses are not obviously wrong.
Confirm and reschedule meetings. Short, bounded, predictable. One of the clearest wins.
Answer inbound and route it. A caller reaching a number they were given, with a known reason for calling. The conversation has context and a narrow range of outcomes.
Leave a voicemail. Trivially, and better than a human at consistency.
The common factor in all four is that the conversation has a small number of possible paths. Where voice AI works, it works because the script is real.
Where it fails
Cold opens to a stranger. The first eight seconds of a cold call are an adversarial negotiation. The prospect is deciding whether to hang up, and the signals they use are tone, hesitation and how you handle the interruption. Voice agents handle interruptions poorly, and being interrupted is the defining feature of a cold call.
Anything off-script. A prospect asks something unexpected, or answers a question with a question, or is mildly hostile. Text gives a model time and a bounded input. Voice gives it neither.
Disclosure. In several jurisdictions, and increasingly as a matter of basic practice, you should disclose that a caller is automated. Doing so changes the call, usually for the worse. Not doing so is a risk that grows every year as regulation catches up.
Reputation. The volume of automated calling has made people hostile to anything that sounds synthetic. An AI voice that is nearly convincing may perform worse than one that is clearly a machine, because the near-miss reads as deceptive.
The recovery problem. When a text sequence goes wrong you look slightly robotic. When a voice call goes wrong you have wasted somebody’s time in real time and they remember it. The downside is asymmetric in a way it is not for email.
Where the actual value is
Not the AI on the call. The AI around it.
This is the part that gets overlooked because it is less impressive in a demo, and it is where almost all the measurable benefit sits.
Deciding who to call and when. The largest lever in calling economics is connect rate, and connect rate is mostly a function of list quality and timing. A system that prioritises a queue by signal — someone just changed jobs, their company just posted three roles, they engaged with something relevant this week — raises connect and conversion rates more than any improvement in the conversation itself.
Preparing the call. Pulling the last touch, the account’s recent activity and a suggested talk track into the screen before the rep dials. Reps lose a large share of their non-talking time to reconstructing context.
Handling the aftermath. Recording, transcription, outcome logging, updating the pipeline stage, scheduling the follow-up. This is pure administrative load and it is the difference between 30 dials an hour and 70.
Routing what to do next. No answer becomes a voicemail drop and a WhatsApp voice note. A callback request becomes a task at the right local time. A gatekeeper becomes a LinkedIn touch instead.
Those four together change calling economics substantially, and none of them require the AI to speak to anybody.
A reasonable position
If you are considering AI voice, the boundary that holds up is whether the person on the other end expects the call.
Inbound and warm: reasonable. Someone who requested contact, or a confirmation call for a meeting they booked. Disclose, keep it short, hand off on anything unexpected.
Cold outbound: not yet, for most B2B. The failure modes are public, the recovery is poor and the regulatory direction is unfavourable. The economics also matter less than people assume — as the cost model shows, rep time is around 80% of cost per meeting, so automating the conversation attacks the right line, but only if connect quality holds. It currently does not.
Internal and administrative: entirely reasonable and under-used. Transcription, summarisation, coaching feedback from recordings, pipeline updates.
What to ask a voice AI vendor
- What does it do when interrupted mid-sentence?
- What happens when the prospect asks something outside the script?
- Does it disclose that it is automated, and can that be turned off? (If the answer is that it can be turned off and they recommend doing so, that tells you something.)
- What is the actual connect-to-conversation rate, not the connect rate?
- Can I listen to ten unedited calls rather than a reel?
That last one settles most evaluations quickly.
The short version
Voice AI handles bounded, expected conversations well and unbounded, unexpected ones badly. Cold calling is the second kind, which is why the category has not moved there despite obvious commercial incentive.
The version that pays off today is AI around the call rather than on it: signal-driven queue prioritisation, context on screen before the dial, and automatic logging afterwards. Those change dials per hour and connect rate, which are the two numbers that actually move calling economics.
Ask for ten unedited recordings before believing anything else.
Next: what cold calling costs per meeting, or AI SDR versus hiring an SDR.
Frequently asked questions
Can AI make cold calls?
Technically yes. In practice it performs poorly on cold opens to strangers, because cold calls are defined by interruption and off-script responses, which is where voice models are weakest. It works well on bounded conversations like meeting confirmations and inbound routing.
Are AI sales calls legal?
Rules vary by jurisdiction and are tightening. Several require disclosure that a caller is automated, and consent rules for recording differ by state and country. Treat disclosure as the default and check the rules for the markets you dial.
What can AI do for cold calling?
The valuable parts are around the call: prioritising the queue by buying signal, putting context and a talk track on screen before the dial, and handling recording, transcription, outcome logging and follow-up scheduling afterwards.
Do AI SDRs make phone calls?
Almost none do. The category is overwhelmingly email with some LinkedIn, because voice is a harder problem with more visible failure modes.