Management – Bakery and Confectionery Review, December 2025
The architecture of the recommendation system is deliberately simple, yet highly effective operationally. Its key components are a financial filter aligned with the declared limit of the monthly instalment and a popularity ranking of the proposed vehicle.
The aim of this paper is not only to describe the method, but also to argue why a “simple bicycle” is, in a real organisation, often more effective than a “racing car” whose construction drags on for months. The system is intended to be ready for deployment in a short time, measurable and easy to scale; its advantages lie in predictability, decision discipline among consultants and robustness to “noise” in narrow customer segments.
A ready-to-use application, described in this publication and written in Python, is available in a GitHub repository at:
https://github.com/ff-wm/reco-A1
Business context and motivation
Let us imagine the call centre of a company brokering the sale or lease of vehicles. In a classic scenario, a consultant answers a phone call from a bakery owner and hears the question: “Which vehicle will be suitable for me?” If the answer is off target – for example, a very expensive car without financing is proposed in a situation where the client expects leasing with a specified instalment – the conversation in practice ends immediately, even if politeness keeps it going for a short while longer. The first hit is crucial, because it reduces the client’s cognitive effort and signals fit: “These people understand my needs.” The opposite situation triggers an avoidance mechanism: the client does not return, because they have experienced dissonance and feel their time has been wasted.
Organisations often postpone building their own recommendation system, waiting for the “ideal” moment or telling themselves: “We are efficient – why change what works well?” Ultimately, daydreams arise about some future large-scale system that will revolutionise sales. This is a costly illusion. Business needs results here and now, not demonstrations of new technologies or advanced artificial intelligence. The idea of “low-hanging fruit” consists in implementing solutions which – despite their simplicity – deliver immediate effects, improve key indicators such as conversion, average revenue per contact, or shorten the time needed to handle a transaction.
A small recommendation system as a low-hanging fruit leading to a measurable competitive advantage
In B2B sales interactions, the impression formed in the first seconds of the conversation has a strong impact on subsequent decisions. A call centre that, right at the outset, proposes an appropriate configuration of a delivery vehicle shortens the client’s decision path and reduces their uncertainty. In this text I present a conceptual, inexpensive and simple recommendation system for a call centre selling delivery vehicles up to 3.5 tonnes to bakeries.
W-MOSZCZYNSKI ppic 12-25
Problem framing and data assumptions
We consider the sale of commercial vehicles to bakeries. We focus on the segment up to 3.5 tonnes, because for bakeries this range is particularly attractive. It combines the availability of drivers with a category B driving licence with functionality sufficient for daily delivery of baked goods. We assume that the call centre has access to a database of clients who call regarding new vehicles. The database contains information on the maximum monthly budget, that is, the limit of the leasing instalment, as well as basic client data, such as scale of production and region of operation. Clients with a purchase history have stored in the database information about vehicle segments, models, equipment variants and forms of financing. The call centre also has a vehicle catalogue containing information such as segment, model, equipment version (trim), form of financing, and the estimated price together with the resulting estimated loan or leasing instalment.
It is important to separate two roles of the data:
• first – they serve as a decision constraint in the form of the client’s budget, which is then translated into a hard filter for clients;
• second – they provide a signal of preferences in the form of purchase frequencies across financial classes.
This makes it possible to build recommendations that are both feasible (fit within the instalment limit) and consistent with the purchasing habits of similar clients.
Recommendation mechanism
The method consists of two sequential steps, but both parts form one compact recommendation procedure. First, for a given client we determine their income position relative to the population. In practice, the most convenient approach is to divide the distribution of the declared instalment limit into six equally numerous intervals (sextiles). In other words, we divide clients into six economic classes. One can choose any number of affluence classes; the number itself is not of major importance here. Such a division balances class sizes and facilitates the estimation of stable purchase frequencies.
Next, we restrict the candidates for recommendation to those vehicle configurations whose forecast monthly instalment does not exceed the client’s declared limit. This hard financial filter is a critical element: it removes random and sometimes embarrassing proposals from the call centre that “blow up” the first impression and lead the client to withdraw from the transaction.
In the second step, we order the list of six economic classes, i.e. the clients’ budget capacity. To do this, we evaluate each vehicle variant in the context of the client’s financial class by its conditional popularity in each class, specifically the purchase frequency in a given interval, as well as by its global popularity, i.e. its frequency in the entire client population. The signal indicating the most frequently purchased configuration within a given financial interval can be unstable when the number of observations is small. A single “exotic” purchase may shift the ranking.
Why do we use additive smoothing rather than, for example, a weighted mean or a median?
Because we aim to maintain stability despite data sparsity in some classes. Fixed weights ignore the fact that classes differ in size. The M-estimate automatically scales the influence of the global component to the sample size, ensuring that classes with few transactions do not “burn through” the ranking by accident. Let us assume that the middle class contains only five transactions, and within these transactions one client has bought seven vehicles from the Indian manufacturer Tata Motors. A single rare transaction can disrupt the stability of a recommendation system in a class containing few transactions; therefore, it is necessary to stabilise it using global means.
The goal of the recommendation system is to provide a ranking of the most popular vehicles in specific classes. A kind of top-5 list of vehicles with the highest probability of sale in a given class is created. In other words, a recommendation list is produced, ordered by increasing probability of a “hit” and, secondarily, by decreasing monthly cost (instalment). As a result, the consultant receives a set of proposals that are both statistically credible and financially acceptable. As mentioned earlier, in cases of extremely rare vehicles on the market that appear in a specific, very small class, we mitigate extremes by incorporating the global ranking, which prevents domination by exotic brands that are a kind of cultural anomaly.
Intuition and stability
Clients react not only to price and specification, but also to the impression of fit and time savings. The first sensible proposal reduces uncertainty and shortens the negotiation distance. From a statistical perspective, the method exploits the fact that preference distributions within similar financial constraints resemble one another. The admixture of global popularity acts as a regulator that “calms” the ranking when the local sample is too small.
From a business point of view, the solution is transparent: we can explain to the consultant why a given vehicle ranked high. It was purchased by “people like you”, so it is popular and meets the budget constraints, which fosters trust in the call centre and facilitates the process. Everyone knows that a popular vehicle model means easy access to spare parts and service, as well as a large amount of user information available online.
Operational implementation without code
It is not necessary for the consultant to understand the details of the algorithm. From their perspective, the interface is a field in which the client identifier is entered; a list of five vehicle proposals with a forecast instalment and a brief justification then appears. From the viewpoint of the psychology of sales, providing specific recommendations is critical. A recommendation opens the conversation but does not close it – the consultant asks two or three follow-up questions, for example about body type preferences or the availability of a driver with a category B licence, then confirms the choice and proposes an alternative variant within the same financial class. Such a scheme limits embarrassing proposals, “offers that miss the mark”, and builds an impression of professionalism and self-confidence.
Measuring the effect in production conditions
The organisation should assess the system from two perspectives.
First – through the prism of operational indicators: conversion of contact into sale, revenue per contact, and time to decision per contact.
Second – from the point of view of ranking quality: whether the purchased vehicle appears in the first five proposals, i.e. on the recommendation list, and in which position on that list.
The most pragmatic experiment is a simple division of call-centre consultants into a test group and a control group. Data “before” and “after” implementation should be collected for the same employees testing the new solution, or a parallel effectiveness test should be conducted in the same time windows.
With normal distributions and “before/after” comparisons, we use Student’s t-test for dependent samples; where normality is violated, we use the non-parametric Mann–Whitney or permutation test. We adopt a conservative decision criterion (e.g. significance level 0.05) and, in addition to p-values, we also report effect sizes (Cohen’s d), which facilitates communication of business value. Put more simply: implementing a simple recommendation system also requires testing on a small sample of call-centre employees.
Risks, ethics and change management
Simplicity does not relieve one of responsibility. Particularly sensitive is the way in which we categorise clients – we should limit ourselves to the financial correlate of the transaction (instalment limit), and not to indirect socio-demographic characteristics that could introduce unjustified bias. The system must not “penalise” the client for being different beyond what is economically justified. In other words, if a citizen of India wants to buy Tata vehicles, there is little point in persuading them that “the computer says no”.
Equally important is preventing excessive exploration of niche configurations. This is safeguarded by the mechanism of smoothing using global popularity. Finally, implementation should be accompanied by a short training path for consultants – with emphasis on understanding the principles of operation and consciously formulating justifications for the client.
What next: modelling purchase sequences
Once a simple recommendation system starts delivering results, its complexity can be increased gradually. Natural directions include incorporating seasonality and regional context or using route logistics. One can move towards modelling sequences of vehicle purchases, i.e. dependence between consecutive vehicles over time.
Modelling purchase sequences means adding to the recommendations information on what clients usually choose as the next vehicle after their previous choice and when they usually do so. In other words: not only what is popular in a given budget class now, but also what, with a certain probability, usually follows a given earlier transaction. For example, after 36 months of leasing an Atom van from Ford, clients most often switch to the refrigerated version AtomFrost+. The “what after what” mapping is based on a simple principle of transition frequencies, e.g. from segment A to B within the same budget class with light smoothing, and their influence grows the more recent and longer the client’s history log is. This method is quite popular among vehicle manufacturers; it is said that cars “grow”¹.
Further development: Learning-to-Rank
A simple, functioning recommendation model that earns additional revenue can be expanded in an evolutionary way. The key is to maintain measurement rigour: no extension should weaken the simplicity of interpretation during the conversation with the client.
We treat recommendation as a ranking problem: the model learns which vehicles should be higher for a given client. Instead of predicting “whether they will buy”, we predict the relation “Ford before Renault”. In practice, we feed the model with features of the client (budget), vehicle features from history, financing, instalments and labels from logs, i.e. information on the course of the purchasing process².
Example: for a client with a monthly instalment budget of 1,800 PLN, the model learns that “SUV-lease_48” usually outperforms “wagon-loan_60”, so in the top-5 it moves SUVs higher, because historically they more frequently end in a scheduled contact or a purchase.
Further development: the multi-armed bandit mechanism
This is an evolutionary development approach that boils down to two balancing directions:
• exploitation – we tell the model: show what has worked so far;
• exploration – we tell the model: sometimes show something less familiar to check whether it performs better.
Each direction is a variant of recommendation, a specific configuration or rule for ordering the list. Simple algorithms include ε-greedy (e.g. in 5
For example, suppose we have three list-ordering strategies:
(S1) pure popularity (i.e. what we have described in this article),
(S2) popularity + “what-after-what” sequences,
(S3) popularity + greater emphasis on leasing_36, which happens to be exceptionally profitable for the call centre.
The bandit assigns consultants’ conversations to S1/S2/S3; after each conversation we update the “reward”. Over time, the algorithm increasingly often selects the strategy that yields the highest profit, while still occasionally testing the others to capture any trend shift.
In summary, the multi-armed bandit approach is a simple online learning method that distributes conversations among several strategies (e.g. S1/S2/S3), updates their “reward” (click/scheduled meeting/purchase) after each conversation, and over time increasingly chooses the strategy that delivers the highest profit – while still occasionally testing the others so as not to miss a better option.
Conclusions
Even the simplest recommendation system described here – budget matching combined with a popularity ranking and stabilising smoothing – can significantly improve call-centre sales outcomes by increasing the accuracy of the first proposal, and thereby sales conversion.
The advantage of this recommendation system stems from three features:
• it is quick to implement,
• it is transparent to explain,
• it is robust to data sparsity in narrow segments.
Contrary to intuition, “toy-like” does not mean frivolous: in many contexts it is precisely this type of system that constitutes the first step enabling an organisation to begin the journey towards more advanced forms of personalisation and revenue optimisation.
Wojciech Moszczyński
Wojciech Moszczyński – graduate of the Chair of Econometrics and Statistics at Nicolaus Copernicus University in Toruń, specialist in econometrics, finance, data science and management accounting. He specialises in the optimisation of production and logistics processes. He conducts research in the field of development and application of artificial intelligence. For many years he has been involved in popularising machine learning and data science in business environments.
¹ Golf VII (since 2012) moved to the modular MQB platform – the same family on which the Passat B8 (since 2014) was built. In practice this meant a “floorpan” and key chassis elements that were shared or similar, so the Golf clearly “grew” towards the former size of the Passat.
² Example models: LambdaMART, XGBoost ranker, pairwise Logistic Regression. Metrics: NDCG@k, MAP, Hit@k.