Supplier Mapping: How Your Company Can use AI to Find the Best Suppliers

by Erik Fubel18.10.2022

Some might have doubted Clive Humby’s statement “data is the new oil” back in 2006, but today nothing rings more true. With the release of Google Search and Amazon’s Alexa, or newer innovations such as the image generator DALL-E 2, it appears data has become the new universal currency. But how does the importance of data relate to the world of procurement? Read on to find out!

The importance of supplier data in procurement

In procurement, the central problem revolves around matching the best supplier to a given requisition under specific criteria. As a solution to this problem, Lhotse’s AI helps companies find the best suppliers faster and more accurately.

If a customer makes a request for a ‘Macbook Pro 13’, how can we actually identify the numerous electronic retailers who sell them? Simple: the answer is all in the data. At Lhotse, we have collected a vast quantity of data spanning from catalogs and keywords to certificates and product categories. All this information makes up a database of hundreds of thousands of suppliers! Additionally, we also enrich this data for each customer by combining it with data from all their past requisitions. Knowing that a customer has previously bought iPhones and Windows laptops from a specific supplier enables us to predict in the future that the supplier also sells Macbooks.

However, sometimes problems can arise if the supplier purchase history is not detailed enough. So, what happens then? Well, in many cases, several supplier branches exist that offer roughly the same products but are considered separate suppliers. Merging this information which is provided by the requisitions made by business users is crucial. This is because it helps the Lhotse App to make more confident decisions!

Figure 1: Comparison of available information with vs. without grouping of supplier branches


Grouping suppliers is harder than it looks

While the advantages of grouping suppliers are straightforward, the process is anything but. In simple cases, the difference between two suppliers can merely lie in the name of the city - in others it can vary greatly.

Suppliers may belong together despite their names differing in terms of:

  1. Acronyms (ABC GmbH vs. Anna Bauer Computers GmbH)
  2. Abbreviations (ABC Technology GmbH vs. ABC Tech GmbH)
  3. Business description (ABC Technology GmbH vs. ABC GmbH)

In other cases, though, two suppliers might sound very similar but should not be grouped together because their stock could be entirely different. They vary mostly in terms of:

  1. Description of the business unit (ABC GmbH Human Resources vs. ABC GmbH Logistics)

It’s clear then: basic computational heuristics simply do not suffice in the pursuit of accurately determining whether two suppliers belong together.

Instead, we could apply some comparisons, check online data for the two branches and then come to a confident conclusion. Seems quick and painless, right?

But there’s a catch:

Lhotse’s supplier base consists of hundreds of thousands of suppliers. Therefore, a supplier branch holds the risk of matching with every other supplier branch. Technically speaking, the number of potential matches grows quadratically with the number of suppliers. So, how quickly would a result be produced, if a comparison were only to take a second? You would be waiting a while because the answer is more than 10,000 years! It goes without saying that this is simply not a feasible option.

The implications are, therefore, two-fold: firstly, we must ensure comparisons are fast but simultaneously accurate, and secondly, we must reduce the number of comparisons that occur as much as possible.

Figure 2: The increase of computational complexity with a larger supplier base

Lhotse’s solution

‘Blocking’ and relevance scores

The first step is to eliminate as many irrelevant potential matches as quickly as possible! This process is named ‘blocking’ and in order to implement it, we construct an extremely fast yet accurate algorithm which can eliminate all pairs that are clearly not a match.

Take these three suppliers as an example:

  1. ‘Jupiter Berlin GmbH’,
  2. ‘Jupiter GmbH Branch Cologne’,
  3. ‘Hürth Car Repair KG’,

Here, only ‘Jupiter Berlin GmbH’ and ‘Jupiter GmbH Branch Cologne’ would pass the blocking process as they are the most likely to belong to the same supplier. Simultaneously, the algorithm can decipher that ‘Hürth Car Repair KG’ is too different from the other two suppliers to potentially match them, so these supplier pairs are not further considered.

In this way, we can reduce 99.93% of the billions of potential pairs within mere seconds.

Figure 3: Blocking irrelevant combinations

What’s next?

Now that only the relevant potential pairs remain, we must make a final decision as to whether the two suppliers belong together. Having, fortunately, eliminated many potential pairs during the blocking process, we can now divert our attention to more elaborate methods. These may take longer to compute but are more successful in taking into account all the tiny, yet, extremely relevant details of each supplier.

For this, we trained a Machine Learning model on thousands of samples. By providing it with different scores on the similarity of two suppliers, it can learn what makes two suppliers more or less likely to match. As a result, it can predict the likelihood of any previously unseen pair of suppliers belonging together.

Figure 4: Estimating the probability of two suppliers matching

For this reason, we know which pairs of suppliers are likely to match. But in order to make the most of our data, we need to also create groups of suppliers, rather than just pairs.

To do so, we connect suppliers through a graph where two suppliers are connected if their probability of matching is above a certain threshold. All suppliers that are connected within one component of the graph are grouped together.

Figure 5: Using graphs to group suppliers

Using this set of techniques, the assignment of a supplier branch to a particular group is correct in 98% of all cases!


For Lhotse to correctly match suppliers to a given requisition, it is essential to collate and combine the past data of several supplier branches. Simple heuristics coupled with state-of-the-art AI solutions enable Lhotse to automate the entire process of grouping supplier branches. This also all occurs with very high accuracy and little to no computational effort. It’s all in the data!

Fully leveraging supplier data is paramount to the success of any procurement team. This is because making the most of data aids in the understanding and general performance of supplier capabilities allowing companies to make better decisions regarding their suppliers. In short: it reduces costs and improves transparency!

If you wish to find out more about how Lhotse can help you to improve your procurement processes, book a demo with us today!


Copyright © 2023 Lhotse. All rights reserved