Machine Learning Supply Contract Compliance

In an effort to reinforce an effective and compliant approach to responsible sourcing across the company, Ericsson’s leading artificial intelligence and data scientists recently began implementing a new solution based on the machine learning that is effective for:

  • reducing manual effort for audits by a substantial number of hours
  • ensure full coverage for all contracts
  • provide information and suggest proactive actions
  • reduce the risk of non-compliant contracts

Below, we’ll walk you through these benefits from a responsible sourcing perspective and dig deeper into our new machine learning-based solution.

The benefits of contract compliance automation

Let’s start with the background: what are the benefits of solutions like machine learning-based signature verification? To answer this, it makes it possible to establish the extent of the challenges posed by the traditional processes of the company, such as the control of the conformity of supply contracts in this case.

Today, Ericsson secures relevant software, services and hardware from vendor partners around the world covering the needs of all Ericsson units to meet business needs. The procurement team and specifically the procurement contract managers sign contracts with various suppliers of varying operational size, diverse geography and a wide range of legal terms and conditions. To ensure legal value, these contracts must be bilaterally agreed and signed by both the supplier and Ericsson’s agents to make them a legally compliant contract.

Ericsson’s sourcing team works with over 40,000 suppliers, resulting in 180,000 contracts worldwide in 170 countries.

While managing supply contracts of this gigantic scale and diversity, there can sometimes be unintended gaps which, in turn, pose a major risk to Ericsson. In case of missing signatures, the contracts become invalid and have no legal value. The impact of an undetected impropriety could lead to an ethics violation, high financial penalties and damage to the reputation of the Ericsson brand around the world. To manage the risk of non-compliant contracts, the sourcing team is very proactive in carrying out periodic audits. Since audits are both a manual and resource-intensive effort, automation could help significantly reduce the number of hours needed to perform such audits (see Figure 1).

It is also crucial to ensure complete coverage and to give the supply contract manager information on potential non-conformances so that he can take early and preventive action. This is another advantage offered by our machine learning-based solution, in that it not only helps to significantly reduce reliance on manual effort, but can also provide complete and uniform coverage and offer proactive information on potential non-conformities.

This clearly aligns with Ericsson’s focus area Responsible business and digital inclusionwhich includes a key strategic focus on Responsible Sourcing.

The responsible sourcing strategy states: “Managing social, ethical, environmental and human rights impacts in our supply chain is part of our value chain approach to integrating corporate responsibility into all of our activities. Strengthening the ability of our suppliers to meet high standards in all these areas is a fundamental part of our approach.” Supply contract compliance is one of the building blocks aligned to make the strategic vision a reality.

Figure 1: Manual steps behind Ericsson’s supply compliance process

Leverage machine learning models to ensure continuous compliance improvements

Identifying the most appropriate machine learning technique requires breaking down the business problem into the following components.

  1. Detection: Does the document contain a signature page?
  2. Verification: is the document signed?
  3. Identification: Is the document signed by Ericsson’s power of attorney?

This is also visualized in Figure 2 where the components of the problem are shown.

The different stages of the commercial challenge

Figure 2: The different stages of the business challenge

Figure 3 illustrates the end-to-end automated solution. It incorporates multiple models to get the results, and now we will dig deeper into each of the models and results.

Introducing the Sourcing Signature Detection Solution

Figure 3: Sourcing Signature Detection Solution Overview

Step 1: Detection – identification of signature pages

To cover all terms and conditions, contracts can range from three to 150 pages. During data analysis, it was realized that the signature page usually has common words that can be used to identify it as a page containing a signature.

Common Words in Signature Pages

Figure 4: Common words in signature pages

Using natural language processing (NLP), a text classification model based on machine learning techniques was used to identify a page as a signature page. Figure 4 shows common keywords that appear on signature pages.

Step 2: Verification – detection of signatures

The next step was to find the number of signatures and locate each signature on the page. To do this, we used YOLO, a family of pre-trained object detection models, which gives us the exact location of the signature(s).

Sample input image

Figure 5: Sample input image

Figure 5 shows the three signatures present in the contract page. The contract has three signatures, one manual and two digital. The result of the YOLO model is shown in the lower right part of Figure 5. It indicates the coordinates and the signature types (0 – manual signatures and 1 – digital signatures).

Step 3: Identification – validation of signatures offline

A person’s signature shows a high level of consistency and does not change much from time to time. For this reason, we used a machine learning model that detects irregularities and at the same time manages to detect falsified signatures that are very similar in case of qualified falsifications. A Siamese neural network was used to train it to approximate the similarity function which produces a score between 0 (similar) and 1 (different), see figure 6.

Example of similar (label=0) and different (label=1) signature pairs used for model training

Figure 6: Example of pairs of similar (label = 0) and different (label = 1) signatures used for model training.

Observations based on training machine learning models

We have shown that we can use machine learning to identify valid signatures in contracts and have created a scalable solution that can be used anywhere we have a sample of power of attorney signatures. As is the case with training machine learning models, the more high-quality data we have, the more accurate the solution. It is therefore not surprising that it was easier to find the signature pages in contracts written in English and contracts scanned in high quality.

The most difficult of the three tasks was verifying signatures, and the highest success rate was locating signatures. Signature verification passed for the most common signatures, but also gave satisfactory results for the less common ones. The full model shown in Figure 3, with all three steps, showed 85% accuracy. This is a result that would not have been possible a few years ago, but thanks to the development of neural networks, deep learning and its applications, we can achieve these results today.

The result

We believe that the application of machine learning models, coupled with a strong deployment strategy, can provide a much-needed foundation to enable a greater range of automation-based use cases with the improvement and standardization of process.

As we do for procurement, there is great potential to leverage the benefits of the solution in other business units as well, leading to minimal human intervention with increased oversight over several business-critical processes. ‘company.

To the future

In light of the growing volume of contract data consumed by the solution, we are further optimizing the models to provide a reliable and comprehensive solution for supply contract compliance. Given the sensitivity of contract data and subsequent decision making, the next step is to explore explainable AI solutions to respond to an understandable, transparent, interpretable and trustworthy system.

You want to know more ?

Other application areas of machine learning in Ericsson include predictive network planning, anomaly detection, ticket classification and management, BOM generation, node fault prediction, transportation management , freight forecasting, inventory optimization, supply planning, etc. Learn more about these areas and the other opportunities that await you on Ericsson’s Artificial Intelligence page.

Sherry J. Basler