In short: In sales, AI wins before and after the conversation – research, personalization, notes, follow-up, training. In the conversation itself it loses, because closing runs on trust, and trust isn't a template.
The limit first: what AI breaks in the conversation
A deal happens when someone feels understood. That feeling comes from pace, pauses, follow-up questions and the courage to bring up something uncomfortable. A model that says the perfect sentence produces exactly the opposite: polished distance.
That's why there's no clone on the phone in my sales and no automatic replies in chats with prospects. What runs automatically is everything around the conversation – and that's the bigger part of the work. I've written more about this stance in why emotion beats systems.
The seven places that carry weight
1. Building and pre-qualifying lists
Companies, contacts, industry, size, the state of their website – all from public sources. The machine sorts, the human decides who gets a call. In my website campaign aimed at restaurants, this pipeline produced 616 finished demo websites – every business already had a visible result before first contact.
2. Personalization no human could manage
One real, verified reference per recipient – not “Hi {first name}”. Across 522 letters sent, each one had its own opening sentence based on what was actually on the recipient's website. Important: every sentence gets checked against a list before it's allowed out. A wrong personalized sentence is worse than none.
3. Preparing for the call
A half-page brief: what does the company do, what stands out, which two questions open the conversation, which objection is likely to come up. Five minutes of preparation decide the first thirty seconds.
4. Call → CRM, without typing
A recording or dictation gets transcribed, summarized and written into the right fields: status, next step, objection, date. This is the spot with the biggest invisible win, because salespeople otherwise don't do it – and without an entry, a lead is worthless after two weeks.
5. Follow-up that doesn't get forgotten
Sequences that send reminders on their own, with prepared drafts and clear stop conditions. A human hits send. What matters here isn't the wording, it's that the follow-up happens at all.
6. Objection training against your own recordings
The underrated tool. Transcribe your own calls, find the moments where they tipped, and practice exactly those moments. Here a model is allowed to play the customer – in training, artificial smoothness isn't a problem, it's repeatability.
7. Numbers that are simply there in the evening
Calls, appointments, deals, open threads – from chat, CRM and spreadsheet into one daily report. In my company a system writes it by itself every evening. What gets measured gets managed; what gets measured by hand doesn't get measured.
Cold outreach and AI: the rules in Germany
A machine doesn't change the law – it just makes a violation scalable. I work in Germany, so these are the German rules I clarify in every project. Other countries regulate outreach differently; check what applies where you sell.
- Marketing emails generally require prior consent under Section 7 of Germany's Unfair Competition Act (UWG). That applies to B2B too, and it applies when a model wrote the text.
- Phone calls to businesses can be allowed on the basis of presumed consent; to private individuals they can't. What “presumed” means depends on the individual case.
- Letters are the channel with the least friction – one of the reasons my campaign ran on paper.
- Data from public sources isn't automatically free to use; the GDPR still applies, including access and deletion requests.
This is my working basis from practice, not legal advice. For a binding assessment you need a lawyer – in a campaign with five hundred addresses, those two hours of advice are the cheapest line item in the plan.
The stack I work with
| Task | What I run | Why |
|---|---|---|
| CRM | Close | phone, email, smart views and workflows in one place |
| Transcription | a local model on my own machine | call content never leaves the building |
| Copy and drafts | Claude | longer context, usable structure |
| Chains and triggers | my own services plus a no-code builder | traceable and without subscription sprawl |
| Numbers | my own dashboard built on the daily reports | one source of truth, not three spreadsheets |
What matters isn't the list of tools but that the pieces connect. Five connected steps beat fifteen separate subscriptions.
Frequently asked questions
Can I send cold emails written by AI?
AI doesn't change the legal situation. In Germany, marketing emails generally require consent under Section 7 UWG, including B2B. Writing without consent risks cease-and-desist letters – with a template just as much as with a model. Phone calls to businesses may be possible in some circumstances; letters are the least critical channel. Rules differ by country. Not legal advice; if in doubt, ask a lawyer.
Will AI replace closers and setters?
Not the conversation. What gets replaced is the groundwork: lists, research, notes, follow-up reminders, reporting. In my sales team, the result is that the same people have more conversations instead of typing.
How do I personalize hundreds of messages without landing in spam?
One real, verified reference per recipient from a source that actually exists – and a check of every sentence against a list before it goes out. On top of that, cap the daily volume, set up your domain properly and switch channels if in doubt. Wrong personalization does more harm than none.
Which AI tools do I really need in sales?
A CRM where everything lands. Transcription for calls. A language model for drafts and summaries. A chain that sends reminders. That's all you need to start – and the four have to be connected.
How do I get calls into the CRM automatically?
Transcribe the recording or dictation, generate a short summary with fixed fields from it – status, next step, objection, date – and write those fields into the CRM. The human checks, the machine types.
Can I have AI running live during the call?
Technically yes; I don't do it. If you're looking at suggestions during the call, you're not listening to the person in front of you – and listening is exactly why people buy. After the call, the analysis is gold.
Mario De Luca I build sales systems and AI-powered software for businesses and founded the Clone Looping System (CLS). Every figure in these guides comes from my own projects, not from studies. Anything I haven't measured myself is labeled as an estimate. More about me · Journal
Who's writing this
I'm Mario De Luca, an AI systems builder from the DACH region and founder of NextCore and the Clone Looping System. My story, my projects and the principles behind these guides are on my about page. In the journal I write about what I've built, tested or messed up myself.
About Mario De Luca →