March 2026
This article describes a method of analysing telephone conversations in a company that supplies the food industry—including bakeries and confectioneries—with machine spare parts, components, replacement elements, consumables and selected semi-finished products. In such a business, some customers are regular and place orders automatically, for example through system integrations, while a large group order by telephone, ask about parameters, negotiate prices, and agree delivery dates and conditions.
W-MOSZCZYNSKI-PPIC-3-2026Conversations often concern expensive items such as machine modules, drives, control systems and precision elements, but also inexpensive consumable elements that generate large turnover on a monthly scale.
The most important objective of the analysis described here is practical and business-oriented, while also being very specific. The customer is to be recognized quickly from the conversation and then routed to the salesperson who best matches the customer’s conversational style and manner of making decisions. A well-matched customer–salesperson pair increases the probability of a sale, while a poorly matched one can interrupt a transaction even when the customer genuinely needs the part and has the budget. Because telephone sales are handled in a call centre, conversations are recorded, transcribed and then analysed in a repeatable manner that is as stable as possible over time.
Stage one—recording and transcription of the conversation
Telephone conversations are recorded by the call-centre system, and the scale is large because there may be thousands of calls per month. At the outset, the customer is identified mainly by telephone number, which makes it possible to combine conversations into a contact history and link them with sales data and data from the company’s systems.
After the conversation ends, transcription is performed—that is, speech is recorded as text—using OpenAI’s Whisper-1 model. The statements must then be divided between two roles by determining which fragments belong to the customer and which to the salesperson. In practice, many models do not do this reliably because voices overlap, people interrupt one another, and sometimes the conversation is fast and chaotic. The pipeline therefore includes an additional language model—in this approach, OpenAI’s o1—which reads the content and assigns statements to roles on the basis of conversational context. This step is critical because further analysis must be certain who said a given thing; otherwise, the conclusions will be mixed up and useless.
The result of the first stage is simple in form: the conversation text divided into roles and stored in a database as a table, most often a DataFrame and frequently a CSV file, so that characteristics can easily be calculated and the results combined with other data.
Stage two—description of the customer’s and salesperson’s characteristics
In the second stage, the role-separated transcript is passed for analysis in which a language model calculates characteristics separately for the customer and the salesperson. Two major personality methodologies are used: OCEAN, known as the Big Five, and HEXACO, which extends the description with an additional dimension. Each characteristic is assessed on a scale of 0–100, where 0 denotes no intensity of the characteristic and 100 denotes maximum intensity, although in practice most results fall in the middle ranges rather than at the extremes.
It is very important for the characteristics to be stored separately in different columns because characteristic names from OCEAN and HEXACO may sound similar but are not entirely identical across the methodologies, so later models must see the full data structure. This is supplemented with analysis of Cialdini’s influence methods, where each method is encoded in binary form—that is, as information indicating whether the method appeared in the conversation. Such analysis is performed for both customer and salesperson because influence in telephone sales operates in both directions and the communication style is part of the result.
Heuristics are calculated in parallel: simple content and behavioural characteristics that can be read from the conversation without entering personality theory. These include signals indicating whether the customer has already selected a part or specific machine, whether they are decided, negotiate the price strongly, ask about technical matters and compatibility, ask about formalities, guarantee, law and taxes, press for a deadline, prefer calm arrangements, or display a cautious or risk-taking orientation. The same observations can be made on the salesperson’s side because one salesperson conducts the conversation procedurally, another relationally, a third technically, and a fourth attempts to close the sale quickly.
The result of the second stage is a set of characteristics describing the conversation and both parties, recorded numerically in the data so that it can be used for further division, comparisons and tests.
These are simply the levels of characteristics detected by the next—the third—artificial-intelligence model in this project.
Stage three—classification into groups and the sales matrix
In the third stage, a language model may be used to assign customers and salespeople, on the basis of the collected characteristics, to several descriptive groups called archetypes in practice. For salespeople, this work adopts four types: a salesperson who imposes their own will; a salesperson who guards procedures; a relational and friendly salesperson; and a spontaneous salesperson who acts more visionarily and flexibly. An analogous division can be created on the customer side, providing a simple structure for analysis.
A sales-probability matrix is then built, matching customer type with salesperson type, with the percentage of successful transactions in the cells. Information about sales success comes from company systems such as CRM and accounting, which show whether an order was placed and completed. Such a matrix has practical value because it easily shows that certain combinations of people—customer–salesperson pairs—work better while others work worse, directly supporting decisions about routing calls to the appropriate salespeople.
Limitation of the archetype approach and the stability problem
In practice, this stage has a serious limitation because archetypes created by a language model are not full statistical clustering, while some genuine human characteristics are hidden, mixed and situation-dependent. Research also reveals a problem with stability over time because sales probabilities for the same pairs can change from week to week. A particular pair may look very good for several weeks, after which the effect weakens or disappears. This means that it is difficult to build a commercial routing system intended to operate continuously and predictably on such a division.
An important conclusion follows: a more stable method is required—one that also uses a greater number of characteristics and does not rely solely on four rigid groups.
Stage four—statistical clustering with dimensionality reduction
In the fourth stage, classic clustering is used, but the large number of characteristics must first be handled because the data may contain dozens of variables at once. Principal component analysis, or PCA, is therefore applied. It compresses the information into several dimensions while preserving as much content as possible and organizing the data in space. After this reduction, the k-means method can be used to find natural groups in the data.
The number of clusters is selected using the elbow method—analysis of a plot showing the point at which further increases in the number of clusters cease to provide a large quality gain. This provides a reasonable starting point for building customer and salesperson clusters.
Stage five—a new sales matrix and experimental variants
Once customers and salespeople have been assigned to clusters, a sales-probability matrix is created again, but this time it is based on clusters rather than descriptive archetypes. The assumption is that the arrangement should be more stable because it results from the data rather than imposed labels.
Designing experimental variants is also very important because different sets of characteristics may create different groupings and clusters. In practice, it is worth testing different combinations—for example, heuristics alone with HEXACO; OCEAN with Cialdini’s methods; or the complete combined set. One should also consider quantitative characteristics from the conversation, such as its length, number of words and the proportion of speaking by customer and salesperson. Such variables can be surprisingly informative, especially in negotiations where the manner of conducting the conversation can be as important as its subject.
Stage six—stability and reliability tests
The final stage consists of tests intended to answer whether clustering genuinely stabilizes sales results. In practice, one checks whether sales probabilities for cluster–customer/cluster–salesperson pairs are repeatable in successive weeks and whether the dispersion of those probabilities is small. If the results fluctuate strongly and the standard deviation is large, the customer-routing system will be risky and unreliable. If dispersion is small and the relationships are repeatable, then there is a basis for implementation.
Summary and conclusion
In the bakery and confectionery industry—and more broadly in supplies to the food industry—telephone conversations often concern expensive equipment and costly elements, as well as large orders for consumables, so the quality of the conversation has a direct effect on the result. Analysis of call-centre conversations based on recordings, transcripts and calculated characteristics is therefore a way to recognize the customer quickly and route them to the best-matched salesperson. The key condition for success is a stable classification method because routing—assigning the appropriate salespeople to the appropriate customers—must work stably without unnecessary unpredictability.
The archetype approach is a good step for demonstrating the direction and discussing it with the business, but research indicates that it may be unstable over time. The natural move is therefore towards statistical clustering with dimensionality reduction and stability testing. In practice, stability measured over weeks and across many conversations is what determines whether the system can be implemented as a tool that genuinely increases sales probability through better customer–salesperson matching.
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.
Dodaj komentarz