How Can Artificial Intelligence Support a Salesperson?

January 2026

A rapid recommendation system for call-centre service

W-MOSZCZYNSKI-PPIC-1-2026

A salesperson who does not yet know the customer

Imagine an office selling production machinery—for example ovens, packaging lines and small bakery machines. The office operates like a conventional call centre: the telephone rings, a consultant picks up, and there is someone new at the other end—a new customer. No purchase history, no relationship. This is so-called cold service: no information and no support telling the salesperson how to talk to the customer.

After a few seconds, the typical dance begins: the customer says something, asks something, hesitates a little and complains a little. One customer is impatient and wants a decision “right now”; another examines every technical detail. The salesperson intuitively tries to adapt, but does it by feel because they know neither the customer’s history nor character. Moreover, the salesperson is burdened by their own personality and may be unable to adapt in a fraction of a second, detect the type of customer, or may be unmotivated or tired. Support would be useful.

What kind of support?

The idea is simple: we record a transcript of the conversation, and after several sentences a language model extracts the customer’s psychological decision profile—not their full personality, only what matters to the purchasing process. On that basis, the system advises the salesperson how to conduct the conversation: what to discuss extensively, what to discuss briefly, and what is better left untouched.

With artificial-intelligence support, the customer experiences “chemistry”—the conversation proceeds at a pace and in language that suits them, while the salesperson does not have to guess how to speak because clear guidelines appear on the screen. This is a new type of salesperson: a salesperson supported by artificial-intelligence systems who receives customized guidance in real time.

For such a system to work, two things are required:

1. a sensible set of dimensions describing the customer’s decision psychology;
2. very simple, even graphical KPIs for the salesperson—the employee will not read instructions during a conversation and must receive quick graphical information on how to speak.

Five dimensions that genuinely help in sales

The literature contains many theories of personality and decision styles. A call centre selling machines needs something less academic and more suitable for immediate use. I propose five dimensions.

The first dimension: promotion focus versus prevention focus

• A customer in promotion mode says things such as: “I would like something more powerful; a richer package would be good”; “I am interested in a newer model, something modern.”
• A customer in prevention mode says: “The most important thing is that it does not break down”; “I want to be certain that the instalment will not surprise me.”

For the salesperson, this is a simple map:

• promotion → talk about development, capacity and new possibilities;
• prevention → talk about stability, guarantees and low risk.

Of course, a person is not sufficiently quick and motivated to grasp the customer’s focus instantly. A machine does it for them. From the transcript, a language model can extract a simple indicator: the customer is closer to promotion, closer to prevention, or somewhere in between.

The second dimension: solution-search style—maximizer versus satisficer

• Maximizer: asks many questions, compares and returns to previous options: “What other versions are available?”; “How much does that amount to over five years?”; “How does it compare with that model?” The customer is afraid of deciding too quickly and regretting it later.
• Satisficer: looks for a solution that is “good enough” and quickly accepts a reasonable proposal: “If you say this is a good configuration, let us take it. We do not need to keep looking.”

In a call centre:

• maximizers benefit from brief but sound comparisons, figures and TCO;
• satisficers respond better to one or two well-described options without an extensive list.

The third dimension: price sensitivity and risk aversion

• Price sensitivity: “What further discount can we obtain?”; “That instalment is too high; we must bring it down.”
• Risk aversion: “I am afraid of these new drives/variable rates/used machines”; “A fixed instalment and a long guarantee period are best.”

The artificial-intelligence model can immediately indicate that the customer has high price sensitivity and will negotiate firmly, or high risk aversion and will not accept complicated financing with an uncertain final cost. The resulting guidance is direct:

• high price sensitivity → clear discounts, cost comparisons, and the argument “it pays”;
• high risk aversion → fixed instalments, long guarantees and no surprises.

The fourth dimension: classic consumer decision-making styles

In machine sales, we do not need the entire theory; several labels are enough:

• Quality-conscious/perfectionist—it must be solid, not the cheapest.
• Price-value conscious—it must pay; the best price-to-quality ratio.
• Brand loyal—I buy this brand because I have always done so.
• Impulsive/habitual—acts quickly, sometimes according to a pattern, “as always”.

For the salesperson:

• perfectionist → go deeper into parameters, workmanship quality and certificates;
• price-value → show a specific equation: “this is what it costs; this is what it saves”;
• brand loyal → do not overwhelm them with a competitor’s offer; show the new model of their favourite brand.

It cannot always be inferred from one conversation, but in many cases the customer provides the signals themselves—for example, “I have only used X for years.”

The fifth dimension: how the customer speaks, not what they choose

Communication styles: directness and detail orientation.

Direct versus indirect: does the customer say, “I want a machine for PLN 3,000 per month; give me specifics”, or do they talk at length about the context, employees and company history?

Detail-oriented versus big-picture: do they ask about detailed parameters—motor power, sheet thickness and bearing type—or rather say, “it should be convenient; I do not want to worry”?

This allows the length and style of the consultant’s statements to be adapted: short, specific sentences for direct customers, and a more descriptive approach for customers who need a “story”.

What can be extracted from a brief conversation?

Long questionnaires are unnecessary. Strong signals are already visible in the first minutes:

• Promotion versus prevention—practically from the first sentence: “I want greater capacity” versus “the most important thing is that it does not break down”.
• Maximizer versus satisficer—after several questions: either the customer digs deeply or says “that is enough”.
• Price sensitivity/risk aversion—in every question about a discount and every remark about risk.

A language model can collect these signals from the transcript and generate a structured description: the customer is more promotion-focused, has high risk aversion and moderate price sensitivity, is probably a maximizer, and is quite direct.

More difficult are:

• deep decision styles, such as perfectionist versus brand-loyal, which may not appear at all in the first conversation;
• abstract values, such as an approach to work and life, which we do not need here.

It is therefore worth retaining the strong dimensions that can be read quickly: promotion/prevention, maximizer/satisficer, price sensitivity, risk aversion, directness and detail orientation. Generally speaking, everyone has known such an approach for at least 150 years, but no one used it because people are not machines. They cannot manage several dimensions simultaneously while continuing to analyse the message and converse at the same time. This is a subject for artificial intelligence.

Why pragmatic cold instructions are useless to a salesperson

Suppose the model calculated that the customer is:

• Maximizer—70%;
• Satisficer—30%.

It sounds clever, but what should the salesperson do with it in the middle of a conversation? Speak 70% of the time as though to a maximizer and 30% as though to a satisficer? This information is operationally illegible. It is good for publication in an article and terrible for use on a headset.

A salesperson needs simple signals, preferably ones that can be seen graphically on the screen: red, yellow and green; an icon meaning “say a lot”, and another meaning “do not touch”. What matters is:

• whether to give the customer two offers or five;
• whether to talk extensively about financing or merely outline it;
• whether to make comparisons with competitors at all, or whether that discourages the customer.

Numerical and percentage information is therefore too “dense”. It must be reduced to simple KPIs that agree with common sense.

Common-sense KPIs: “SAY_A_LOT”, “SAY_LITTLE”, “DO_NOT_SAY”

Instead of showing the salesperson four radar charts and a table of percentages, it is better to ask several very simple questions:

1. How should technical details be discussed?
2. How should financing be discussed?
3. How should failures/reliability be discussed?
4. How should competitive advantage be discussed?

For each question, the answer can be reduced to:

• SAY_A_LOT—a key subject; develop it and return to it;
• SAY_LITTLE—mention it, but do not elaborate;
• DO_NOT_SAY—it is better to avoid this thread; the customer does not want to hear it.

This can even be drawn: four icons with colours—for example, solid green, a thin yellow bar, and crossed-out red. In the background, the system can use all the psychology, but it outputs only a simple instruction to the salesperson.

How can the prompter model be launched in five minutes?

Here is the address of a file containing the most important things required to launch the artificial-intelligence model.

The first step is to enter one of the Playgrounds. These are so-called play areas where selected artificial-intelligence models can be tested.

Example addresses:

• https://platform.openai.com/chat/edit?models=gpt-4.1
• https://app.fireworks.ai/playground

At the top, there is usually a list of models to choose from. Somewhere near the bottom, a ready prompt—the protocol instructing the model what to do—must be pasted. A prompt created specifically for our call centre is in this file: https://github.com/ff-wm/PPIC/blob/main/A2.ipynb.

We can now proceed according to the instructions in that file. The file is used to test the responses of different LLMs to selected customer dialogues. It contains three customer dialogues: “Dramatic dialogue”, “Neutral dialogue” and “Sensitive-person dialogue”. The listed websites allow an API protocol to be launched—that is, the call-centre workstation can be connected to an artificial-intelligence model and direct guidance on how to talk to the customer can be used.

Conclusion: a new type of salesperson

This system is not an oracle. A brief conversation means:

• less certainty about the classification;
• a risk that the model will invent things the customer did not say.

It is worth adding an internal confidence indicator—high, medium or low—and showing it simply in the interface, for example by colour intensity. It is better to ask the model about specific behaviour in the conversation, such as how the customer reacted to the subject of price, than about their general “character”. The result should be combined with hard data: company size, type of activity and previous purchases.

Even with these limitations, the simple “SAY_A_LOT/SAY_LITTLE/DO_NOT_SAY” KPI division is a huge step forward compared with a salesperson entering a conversation completely blind. A language model will not replace the consultant, but it can put them on the correct track in the first minutes. Instead of abstract percentages such as “Maximizer 70%”, the salesperson receives four simple messages: do not touch the technicalities; say a lot about financing; strongly emphasize reliability; do not give a lecture about competitive advantage.

This is precisely the new type of AI-supported salesperson: a person conducts the conversation, while the system continuously suggests which threads to develop and which to avoid. The customer feels that someone “understands” them, the conversation flows, and the call-centre office ceases to be a lottery in which everything depends on the consultant’s mood and pure intuition.

Wojciech Moszczyński

Wojciech Moszczyński—a graduate of the Department of Econometrics and Statistics at Nicolaus Copernicus University in Toruń; a specialist in econometrics, finance, data science and management accounting. He specializes in optimizing production and logistics processes. He conducts research into the development and application of artificial intelligence. For years, he has been involved in popularizing machine learning and data science in business environments.