AI in sales: what works today — and where it costs you customers

The short answer

AI works in sales today wherever it takes writing and searching time off the salesperson: preparing for the meeting, summarising call notes, first drafts of the quote and follow-up email, role-plays for practice. It works badly wherever it is supposed to replace the conversation: voicebots in cold calling, automatically “personalised” mass emails, forecasts from a CRM nobody maintains. I have seen teams gain hours a week with a simple chat assistant, and teams that burned their reputation in a region with expensive automation. The rule behind it: AI makes the salesperson faster before and after the conversation. The conversation itself stays theirs.

Where AI saves real time today

The following overview is an assessment from practice, not a study:

Task	Is AI suited to it?	What to watch
Meeting preparation (company, sector, questions)	yes	check facts against website and register
Summarising call notes	yes	next step with a date must be in it
Draft quote, follow-up email	yes, as a draft	never adopt prices and commitments unchecked
Objection training in role-play	yes	supply objections from real conversations
Cold calling via voicebot	no	law and reputation, see below
Revenue forecast	only with clean data	rubbish in, rubbish out

Preparation. Before a first meeting I have AI give me three things about a business: what it does, what has changed recently, which three questions I should ask. That does not replace my own research, but it shortens it. How the questions then fit into the conversation is covered in the article on first-meeting structure.

Follow-up. This is the biggest lever. After the meeting the salesperson dictates two minutes of bullet points into their phone, and AI turns them into the note in the format “Status / customer’s open question / next step on …”. That is exactly the entry that is missing from the CRM salespeople actually use when it has to be written from memory in the evening.

Practice. A language model can play a sceptical buyer who says “too expensive”, five times in a row and slightly differently each time. For new salespeople that is a good sparring partner before their first real meeting — provided you supply the objections that actually come up in your own market.

A calculation on follow-up

A model calculation with freely chosen figures, to show the order of magnitude:

• 12 customer meetings per week
• Follow-up so far 10 minutes per meeting = 120 minutes
• With dictation and AI summary 3 minutes per meeting = 36 minutes
• Gain: 84 minutes per week per salesperson

With ten salespeople that is 14 hours a week. Redo the calculation with your own figures. The real gain, however, is not the time but the quality: a note written five minutes after the conversation contains the next step. One written in the evening often does not.

Where AI costs you customers

Voicebots in cold calling. The customer notices within seconds, and the conversation is over before it has started. Then there is the law: § 7(2) no. 2 UWG requires the addressee’s prior express consent for advertising using an automated calling machine — unlike a call made by a human, this also applies to businesses. Whether an AI voicebot legally counts as a calling machine has, to my knowledge, not yet been firmly decided; I would treat it as one. What applies to human calls is covered in the article on cold calling by phone.

Mass “personalised” emails. The company name in the first sentence is not personalisation. Anyone who sends a thousand such emails trains their market to ignore them.

Forecasts from dirty data. An AI revenue forecast is only as good as the entries in the CRM. If half the opportunities have no next step, the model is calculating with memories. Data first, then the forecast.

Unchecked commitments. A draft that invents a discount, a delivery time or a contract clause is quickly sent. Responsibility lies with the sender, not the tool.

Recording, transcribing, evaluating: the legal position

Recording only with consent. Anyone who, without authorisation, records the privately spoken word of another person on a sound carrier commits a criminal offence under § 201(1) no. 1 StGB. A customer conversation on the phone or in the office is private. So ask beforehand:

“I’d like to have our conversation recorded so that I can send you a clean summary. Is that all right with you? If not, I’ll take notes as usual.”

The answer belongs in the note. Anyone who does without a recording and only has their own bullet points summarised does not have the problem.

Transparency for bots. Since 2 August 2026, Article 50(1) of the AI Regulation has applied: AI systems that interact directly with people must be designed so that the person is informed they are dealing with an AI, unless that is obvious. Formally the obligation falls on the provider of the system. Anyone using a chatbot on their website should nevertheless check that the notice is switched on.

AI literacy. Article 4 of the AI Regulation obliges companies that use AI to contribute to the AI literacy of their staff. According to KPMG’s summary, the Digital Omnibus amendment of 2026 weakened this obligation from “ensure” to “promote”. A short training session on what may and may not go into the tool is still the minimum.

Works council. § 90(1) no. 3 BetrVG expressly names the use of artificial intelligence: the works council must be informed in good time and with documents. To assess it, it may call in an expert under § 80(3) BetrVG; to that extent this counts as necessary. If a tool evaluates the calls or activities of individual salespeople, co-determination under § 87(1) no. 6 BetrVG is added.

Action list

1. Pick one task, not ten: follow-up after meetings almost always brings the most.
2. Define a format for the call note that AI fills in, with a next step and a date.
3. Write down the rule: a human checks prices, discounts, delivery times and contract content before anything goes out.
4. Recordings only with consent; add the question as a standard sentence to the call script.
5. No voicebots for outbound advertising calls without the express consent of those called.
6. Define which customer data may go into which tool, and train the team on it.
7. Inform the works council before introduction and regulate evaluations by works agreement.
8. Measure after four weeks: minutes of follow-up per meeting and share of CRM opportunities with a next step.