AI cold calling, done properly, means AI does everything around the conversation and a human has the conversation: the AI finds the signal, builds the list, dials several numbers at once, writes the brief, prompts the rep live and logs the call. It does not mean a synthetic voice calling your prospects, and in European B2B it should not. Here is what AI changes in each of the three layers of a cold call, what it measurably moves, and why the rep stays on the line.

What "AI cold calling" means, and the one version to avoid

The phrase covers four very different things: AI that picks who to call, AI that dials, AI that helps the rep during and after the call, and AI that talks to the prospect instead of the rep.

The first three are production tools with measurable results. The fourth is a demo. A cold call is 20 seconds of a stranger deciding whether to trust you. A voice bot loses that decision before the first question, and in Europe it has to announce itself: the EU AI Act requires that people be told when they are talking to an AI system. Add French B2B prospecting rules and the obligation to inform people when calls are recorded, and the bot is a compliance project with a bad meeting rate.

So the useful question is not "can AI make the call" but "what should AI do around it".

Layer 1: the list. AI decides who to call today

This is where most of the lift comes from, and where most teams start last.

Reps use only about 20% of the signals already sitting in their CRM. Half of Sonik's own pipeline came from signals that were already there: dormant leads, closed-lost deals reaching their anniversary, demo requests nobody called, past champions who changed jobs. AI's first job is to watch those sources every night and write a list of 50 contacts with a trigger each, timed by persona.

The timing matters as much as the names. In our benchmark of 1,000,000 B2B cold calls in France, 4pm connects at 15.8% and 8am at 9.8%; founders pick up around 10am, marketing around 4pm, HR around 1pm. The list should arrive already sorted into those blocks.

Layer 2: the dial. AI removes the waiting

On a manual dialer, most of a call block is spent listening to rings, voicemails and dead numbers. Parallel dialing has the AI dial several contacts at once, detect the voicemails and the bad numbers, and hand the rep the first live person. The result in our data is about 5x more conversations per hour from the same seat.

Two conditions for this to work in Europe. The voicemail detection must be accurate for French and German carriers, or the rep gets dropped into voicemails and the prospect gets silence. And the numbers you call from must be authenticated and rotated, or your pickup rate falls as carriers flag you. Ask any vendor for both figures by country.

Our north star is 20 conversations per SDR per day. Without layer 2 that number is not reachable; with it, it is a normal day.

Layer 3: the conversation. AI prepares, prompts and logs; the rep talks

Before the call: a four-line brief, the trigger, the persona's likely pain, a suggested opener. The rep reads it at connect, not ten minutes earlier for a contact who did not pick up.

During the call: a live prompt when the moment comes. The prospect objects, and the screen suggests "What do you mean?" before answering, which lifts conversion by 197% in our data. The rep is about to pitch, and the screen reminds them of the "how" question, which lifts it by 190%. 65% of reps open by explaining what they sell; a prompt that keeps the opener on the trigger moves the meeting rate from 2-3% to 7-10%.

After the call: transcript, summary, outcome, next step in the CRM, follow-up email drafted. The rep clicks once and dials the next person.

What the rep still does: hears the hesitation in the voice, decides to ask rather than answer, chooses the moment to ask for the meeting, and books it. That is the job. AI did not take it; AI made it possible 20 times a day instead of six.

What AI cold calling measurably changes

Three numbers, per rep, compared with the month before the change:

• Conversations per day. Layer 2. From a handful to 15-20 is the typical move.
• Share of calls opened on a trigger. Layers 1 and 3. From under half to above 80%.
• Meetings per conversation. Layer 3 plus the opener. This is the number that confirms the first two were worth it.

Dials per day is not on the list. A team that triples dials and keeps the same meeting count has bought a faster way to annoy its market.

In our experience European teams see conversations jump in the first week, the trigger share in the second, and the meeting rate by the end of the first month, once reps trust the brief enough to use it.

How Sonik approaches AI cold calling

Sonik is built on exactly these three layers. Signals from the CRM, the website, job changes and social intent build the list and assign each contact to the slot where its persona picks up. The parallel dialer runs the session and hands the rep live conversations only. AI call prep writes the brief, live coaching prompts the "how" question and the "What do you mean?", and the call is logged to the CRM automatically.

What Sonik does not do is talk to your prospects. The mission is 2x more meetings through 5x more conversations with the right contacts, held by your reps. The 1,000,000-call benchmark we built with Quentin Despas is the reason we are confident the human part is where the meetings come from.

Conclusion

AI cold calling is three layers of help around a human conversation, not a replacement for it. Let AI pick the list, run the dial, write the brief, prompt the question and log the call. Keep the rep on the line, especially in Europe. This week, measure your team's conversations per day and the share of calls that open on a trigger. Whichever is lower tells you which layer to fix first.