Almost every sales tool has added 'AI' to its name in the last few years. Some of it genuinely changes how work gets done. Some is a decade-old feature in new packaging.
What works today
The applications where language models are genuinely strong share one property: they process text that already exists rather than predicting the future.
- Summarising conversations. Turning fifty minutes into a structure of agreements and tasks is a task where models are solid and predictable.
- Extracting tasks and dates. Recognising 'I'll send pricing by Thursday' and turning it into a dated task works very well.
- A first draft of the follow-up. Not the final version — the first. A human reviews and sends, but does not start from a blank page.
- Mapping content to CRM fields. Assigning fragments of a conversation to the right fields is a classification task, and models do it better than a tired human at 6pm.
Promising but immature
The second category works in demos and disappoints in production — usually because it requires data the company does not have.
Close probability prediction. A model can find correlations in deal history, but if that history was entered from memory on Friday afternoons, it mostly learns the team's documentation habits.
Real-time call guidance. Live prompts sound attractive but in practice split the rep's attention precisely when it should be on the customer.
Fully automated sequences. Sending without a human in the loop scales beautifully — including the mistakes. One inappropriate email to an important account costs more than a hundred automated ones save.
What is marketing
The third category is features that existed before and acquired a new label. Rule-based lead scoring did not become artificial intelligence because the rules were rewritten as weights. An email template with a merge field is still a template.
A simple test: ask the vendor what specifically the model does that an if-statement could not. If the answer stays general, the answer is probably 'nothing'.
The useful question for an AI vendor is not 'how does it work' but 'what happens when it is wrong'.
Questions to settle before choosing
Beyond functionality, three matters determine whether a tool can be deployed in a European company at all.
Where data is processed, and whether transcripts reach models outside the EU. Whether customer data trains models — an answer that belongs in the data processing agreement, not the privacy policy. And how long audio is retained; the cleanest answer is that it is not retained at all.
In Luminote, analysis runs on Mistral AI in Paris, transcription on Voxtral, hosting in Frankfurt or Warsaw. No OpenAI and no data leaving the Union. That is not a functional advantage — it is the precondition for legal to join the conversation at all.
Key takeaway
AI in sales is genuinely good at processing what has already been said and weak at predicting what will happen. Buy the first, treat the second with scepticism.