January 2026
In the sector manufacturing advanced machinery for the food industry—from packaging lines to precision dosing systems—the sales process is rarely quick. Transactions involve high amounts and the decision cycle is complex. In this environment, the call centre ceases to be a simple service office and becomes a strategic hub where engineering meets negotiation.
Moszczynski-PZM-1-2026The purpose of this study is to present the evolution of an analytical system intended to optimize Customer–Salesperson matching, or Smart Routing. A poorly matched consultant can squander the chance of a contract worth hundreds of thousands of zlotys, while an optimal match between psychological and competence profiles substantially increases conversion.
Phase 1: digitization and structuring of voice
The first challenge is to transform thousands of hours of telephone calls into processable data. The scale of the operation requires complete automation.
- Transcription—the Whisper-1 engine is used to convert speech into text.
- Speaker identification (diarization)—this is a critical stage. Because recording systems often provide a single audio track, it is necessary to distinguish who is speaking. An advanced reasoning model was implemented here—the best for this at present is o1—which, on the basis of context such as the use of quotation questions versus technical-specification questions, precisely tags statements as “Customer” or “Technical Adviser”.
Phase 2: extraction of psychometric and behavioural characteristics
Instead of relying on intuition, the system performs an in-depth analysis of every interaction, building a multidimensional feature vector for both sides of the conversation.
- Personality profiles—language is analysed according to the Big Five model, OCEAN, and HEXACO. Every characteristic—for example extraversion, conscientiousness and honesty—is quantified on a scale of 0–100.
- Persuasion techniques—the algorithm uses Cialdini’s seven principles of influence, with binary 0/1 encoding, checking whether the salesperson uses, for example, scarcity or authority and how the customer reacts. It also checks whether customers use these methods.
- Industry heuristics—variables specific to machine sales, such as “customer’s level of technical knowledge”, “fixation on price”, “urgency of implementation” and “propensity for technological risk”.
Phase 3: limitations of the archetype approach
Initial categorization attempts were based on assigning speakers to four archetypes defined by an LLM, such as “Engineer–Formalist”, “Visionary” and “Negotiator”. Although this model was easy for management to understand, it proved empirically unstable. Success matrices based on these labels showed large variation over time—a lack of repeatability from week to week—which made it impossible to build a reliable routing system.
Phase 4: the current standard—dimensionality reduction and clustering (PCA + K-means)
In response to the instability of archetypes, an approach based on hard statistics—the baseline—was implemented. It is an unsupervised method that searches for natural groupings in the data without knowing the sales result.
- Principal component analysis (PCA)—because of the large number of variables, approximately 50 psychological and behavioural characteristics, PCA is used to compress the data. This eliminates noise and correlated variables, reducing the space to several principal dimensions.
- K-means algorithm—the k-means algorithm is run in the reduced space. The number of clusters, k, is selected using the elbow method, permitting the optimal number of customer and salesperson groups to be determined mathematically.
Disadvantages: this method groups people who are similar, but not necessarily those who generate sales. A cluster may be psychologically coherent but heterogeneous in terms of conversion, which still generates a certain amount of “noise” in the decision matrix.
Phase 5: the future—alternative methods (learning to match)
To eliminate the disadvantages of the unsupervised approach and link profiling directly to sales success—the target variable 0–1—we propose implementing five advanced analytical methods.
Alternative 1: topological success mapping (supervised UMAP + HDBSCAN)
Instead of linear data projection with PCA, we use nonlinear UMAP dimensionality reduction in supervised mode.
Mechanism: the algorithm receives information about whether the conversation ended in a sale. It “stretches” the feature space to separate cases of success from failure and only then groups the data.
Advantage: the resulting clusters are, by definition, strongly correlated with sales effectiveness. In addition, using the density-based HDBSCAN algorithm makes it possible to reject “noise”—atypical conversations—which increases data purity and translates into stable results.
Alternative 2: tree embeddings (random forest proximity)
Ensemble models, Random Forest or XGBoost, are used to create a new distance metric.
Mechanism: we train a “Sale Yes/No” classification model. We do not use it for prediction, but to build a similarity matrix. Two customers are “close” if, across thousands of decision trees, they arrive at the same terminal nodes.
Advantage: the method handles nonlinear relationships perfectly—for example, “A technical customer buys only from an Expert Salesperson unless the salesperson is extremely self-confident”—and is highly stable over time.
Alternative 3: Siamese neural networks (deep metric learning)
This is a deep-learning approach inspired by facial-recognition systems.
Mechanism: we build a Siamese neural network that learns vector representations, or embeddings, of the salesperson and customer. The network is trained to minimize the distance between pairs that completed a transaction and maximize it for unsuccessful pairs.
Advantage: we create a dedicated “business matching metric”. The system learns by itself which combinations of characteristics—for example, salesperson extraversion plus customer conscientiousness—lead to success.
Alternative 4: biclustering (spectral co-clustering)
Instead of grouping customers and salespeople independently, patterns in the matrix are sought simultaneously.
Mechanism: the algorithm looks for “blocks” in the data—subgroups of customers who react positively only to specific subgroups of salespeople and negatively to others.
Advantage: this detects niche relationships, or local correlations, that disappear in the global averaging of K-means. It is ideal for identifying specific segments of the machine market, such as “Innovator Customers” versus “Conservative Production Plants”.
Alternative 5: latent probability (latent class analysis/GMM)
Mixture models—Gaussian mixture models—are used instead of hard assignment to groups.
Mechanism: the model assumes that every salesperson and customer is a “mixture” of different styles with specified probabilities—for example, 80% adviser and 20% salesperson.
Advantage: this is crucial for stabilizing results over time. If the salesperson’s style evolves slightly from week to week, the probabilistic model corrects it smoothly. In hard K-means, a small change could move the salesperson into a different group, drastically changing the matrix statistics. GMM prevents such jumps.
Summary and operational conclusions
Moving from the experimental stage—archetypes—to the engineering stage—PCA + K-means, and ultimately supervised methods—is essential if a sales-support system in the machinery industry is to become predictable.
It is recommended that A/B tests begin with supervised UMAP or tree embeddings because they offer the best balance between matching precision and resistance to changes in the data over time. The overriding objective remains to build a stable probability matrix that permits automatic call routing with minimal risk of error.
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