Large Language Models in Sales: What Managing Directors Need to Know

Almost every new sales tool now advertises "generative AI". Behind it are large language models, or LLMs for short. This article explains without jargon how large language models work in sales, which tasks they handle reliably, where they make mistakes and which questions you should ask a provider before you sign.

What a large language model is

A large language model is a program trained on very large amounts of text. In the process it has learned how language is structured: which words go together, how questions and answers relate, what a polite letter sounds like. When it receives an input, it calculates step by step the most likely fitting continuation. The result reads like text written by a person.

One point is key to your understanding: an LLM does not know what is true in the human sense. It produces plausible language. If it is to state your opening hours, your prices or your free slot on Thursday, it has to get this information from a connected source, such as your knowledge base, calendar or CRM. Without that connection, it guesses. This is exactly what separates a useful sales tool from a toy.

What large language models do well in sales

In day-to-day sales, four areas have proven themselves where an LLM's strength, understanding and producing language, brings direct benefit:

• Holding conversations: understanding and answering enquiries in chat, via WhatsApp, email or on the phone, even when customers express themselves awkwardly or change the subject mid-sentence.
• Summarising: turning a ten-minute phone call into a short note for the CRM, with the request, budget range and next step.
• Classifying: sorting incoming messages by request, for example new enquiry, appointment change, complaint or billing question.
• Writing drafts: preparing quote texts, follow-up emails or answers to standard questions that a member of staff checks and sends.

Example from an estate agency

An estate agent receives an enquiry about a three-room flat on a Sunday evening. The prospect writes: "Is it still available, and are dogs OK? We could move in from March." A rule-based chatbot would probably have failed at this mix of three questions. An LLM-based assistant recognises the requests, checks the listing to see whether pets are allowed, gives the move-in date, asks about household size and proof of income and offers a viewing. On Monday morning the agent finds a complete note in the CRM. The difference between the two approaches is described in more detail in the article Rule-based chatbots vs. AI agents.

How the model gets your company knowledge

The language model itself does not know your business. For it to answer sensibly, it is supplied with the right information for each enquiry. Technically, this usually works as follows: the customer's question is matched against your knowledge base, the relevant sections are picked out and passed to the model together with the question. The model then formulates the answer. Specialists call this retrieval; what counts for you is the principle: the quality of the answer depends directly on the quality of your documents.

To get started you do not need data scientists, but you do need order. A short collection of the following has proven useful:

• services and products with binding prices or price ranges
• opening hours, directions, responsibilities and cover arrangements
• the twenty most frequent customer questions with your best answer
• clear no-go topics the assistant must not comment on

If you set up these documents properly once, you benefit twice: new employees also find their way around faster.

Limits you need to know

The best-known weakness is called hallucination: the model formulates a convincing-sounding answer that is simply wrong. In sales this can be expensive, for example if an assistant promises a discount that does not exist or quotes a delivery time nobody can meet. Reputable solutions limit the risk by allowing the model to answer only from approved sources and to hand over to a person when unsure.

Other limits concern judgement and responsibility. An LLM does not understand company policy, knows nothing of a long-standing customer relationship and bears no liability. Negotiations, pricing decisions, goodwill gestures and sensitive conversations therefore belong in human hands. A sales assistant should not answer legal, tax or health questions either.

What managing directors really have to decide

You do not need to know how a neural network calculates. But you do need to know which decisions are yours. The following questions help in conversations with a provider:

1. Where does the knowledge come from? From which sources does the assistant answer, and who maintains them?
2. What happens when it is unsure? Is there a clear rule for when it hands over to a person?
3. Where is data processed? In the EU or outside it, and is your customer data used for training?
4. How is it checked? Can you view conversation histories and correct mistakes?
5. What may the assistant promise? Appointments yes, prices only as per the list, discounts never: set such limits in writing.

Whether you run your own model or use a ready-made platform is a separate trade-off. It is covered in the article Open source vs. commercial AI in sales. For most small and medium-sized businesses, a configurable solution is the more practical way in. After a few months of operation, it is easier to judge whether more customisation is needed.

Data protection and the EU AI Act

As soon as an LLM works with customer data, the GDPR applies. You need a data processing agreement, a legal basis for the processing and clear information in your privacy policy. Under the EU AI Act, customers should be told that they are talking to an AI. An open notice at the start of the conversation is simple and avoids trouble. What SMEs should know about regulation is summarised in the article The EU AI Act for SMEs.

Neurobots' digital employees use language models for phone, chat, WhatsApp and email, process the data on servers in Frankfurt and pass results to CRM systems such as HubSpot or Pipedrive. Setup is handled by a certified partner, typically in about seven days.

When an LLM assistant is not worth it

If your customers almost always ask the same three questions, a good FAQ page is often enough. And if nobody has time to keep the knowledge base up to date, even the best language model will soon give outdated answers.

Frequently asked questions

Is ChatGPT the same as an AI sales assistant?

No. ChatGPT is a general application based on an LLM. A sales assistant uses such a model, but it is connected to your data, calendars and rules and configured for your business.

Does the model learn from my customer conversations?

That depends on the provider and the contract. Ask explicitly and get written confirmation of whether your data is used for training.

How can I tell whether a provider has hallucinations under control?

Test with questions that have no answer in your documents. A good assistant will then say that it does not know and offer a call-back instead of making something up.

Does my sales team need training for this?

A short briefing is usually enough. Your team should know what handovers look like, where to find conversation histories and how to report wrong answers so the knowledge base can be corrected.

Conclusion

Large language models in sales are strong at understanding language, summarising and holding conversations. They only become reliable with well-maintained sources, clear limits and human oversight. What this looks like for property enquiries is shown on the page Digital assistant for estate agents.