Executive summary

Anti-corruption laws expose careless companies and their executives to massive monetary and imprisonment risks.  

A world technology leader selected Neal Analytics to develop Machine Learning models that could proactively detect high-risk deals (HRD) for their volume licensing business.  

These models also had to be built and provided results that followed industry best practices regarding governance and ethical AI.  

The models allowed the customer to scale up its HDR compliance processes by both effectively tagging low-risk deals and proactively and automatically flagging high-risk ones. 

CORRUPTION AND THE CORPORATE RISK​

The effects of corruption on society are understood and well-documented. Corruption costs people—individuals and corporations—financially, but it also poses a risk to essential freedoms. Politically, corruption raises obstacles to democracy and the rule of law. Economically, it misbalances the way businesses operate and compete and risks undermining market economies due to inefficiencies and unfair allocations of resources. The enactment and enforcement of laws such as the U.S. Foreign Corrupt Practices Act (FCPA) illustrate this general point about corruption.   

Corruption can cost individual companies hefty liabilities (e.g., significant fines, imprisonment, government-imposed monitoring) as well as moral culpability (e.g., damage to reputation) if the company is negligent and fails to have adequate controls and policies to prevent corruption (e.g., bribery, collusion, embezzlement, misleading revenue, money laundering, fraud).   

Even though most companies offer employees and business partners some guidance on anti-corruption, the nature of today’s global business means that organizations face difficult challenges when trying to embed clear and effective controls for different regions of the globe where businesses may operate.   

Even well-respected global companies can find themselves on the wrong side of ethical behavior and the rule of law and are not immune to corruption.   

For instance, in 2008, Siemens agreed to pay more than 1.3 billion euros in fines to U.S. and European authorities to settle charges that it   

routinely used bribes and kickbacks as standard operating procedures to win public works contracts around the globe.   

In 2014, a Chinese court fined GlaxoSmithKline nearly half a billion dollars following a conviction for bribery and sentenced several GSK executives to multiple years in prison.   

What if machine learning could help reduce corporate corruption risk?

Neal Analytics leveraged Machine Learning to detect High-Risk Deals (HRD) proactively and automatically.  

Neal Analytics’ data science and risk governance experience and expertise, as well as its shared sense of ethical responsibility, were key selection criteria for the customer. 

Customer technology leadership to tackle corruption risk

Leading by example, the customer saw its global technological presence as an opportunity to set the standard for fighting corruption risk through clear ethical principles and procedural corporate governance.  

Using the customer’s vast amount of sales data (e.g., quotes, contracts), Neal Analytics helped design a Machine Learning (ML) decision-support solution that could early identify, monitor, and mitigate the customer’s corruption risk in the volume licensing space, i.e., High-Risk Deals (HRD).  

HRD is a modern finance solution that not only leverages data analytics to classify the riskiness of sales transactions for additional compliance oversight but also exemplifies how an AI solution can be built responsibly on clear ethical principles of accountability.  

The customer chose Neal Analytics based on the firm’s expertise in decision and data science in risk governance and its shared sense of responsibility in helping lead and educate clients on ethical values and principles.  

This ensured that the ML-based support systems were designed on intelligibility, fairness, reliability, and inclusiveness principles.  

Neal Analytics recognized that, in machine learning, automated recommendations and decisions might have an unintentional or undesirable impact on an AI system’s diverse stakeholders.  

It was, therefore, ready to support the customer in its responsible-AI corporate mission. 

THE HIGH-RISK DEALS ML-DRIVEN SOLUTION​

The High-Risk Deals solution replaces previous processes, which were essentially manual. An individual used to download the deals into Excel. Then, using mostly speculative approaches, compliance managers would interpret the risk intuitively. These previous control processes were inadequate and hard to manage, making deals and their reviews difficult to audit.   

The High-Risk Deals solution is now essentially automated. Risk is scored based on a predefined and clear set of criteria based on representations of anomalous deals. The analytical scoring framework (for which the ML recommendation and decision solution supplements) is based on risk attributes and criteria such as:   

1. 1. Whether a deal involves a state-owned entity customer or not, since historically, state-owned entities have been more susceptible to corruption than non-government entities.   
   2. Whether the deal’s business partner was awarded a sizeable deal discount above standard levels, e.g., deviation from similarly clustered deals and discount distributions.   
   3. Geography and corruption perceptions indexes where the Office of Legal Compliance keeps risk rankings for all countries where Microsoft does business.   
   4. Business partner historical behavior and instances of non-compliance as a proxy for doing business under risk with specific partners.   

High-Risk Deals ML (HDR-ML) has helped score over 25,000 enterprise-agreement deals, which account for over $25B in revenue, and has funneled over 1,500 deals for an additional compliance review in slightly more than a year. A small but significant number of deals have been found to have high risk.   

HRD-ML introduced several desirable qualities for helping fight corruption risk. For instance, it enhanced financial audits by providing quick risk assessments in uncomplicated deals. This permitted subject matter experts to focus on more complex high-stake deals, e.g., multi-million-dollar deals in strategic regions of the world, a.k.a. ‘countries in scope’ deals. Therefore, it alleviates the reviewer’s burnout and conserves financial resources.   

CONFIDENTIALITY, FAIRNESS, INCLUSIVENESS, EXPLAINABILITY, AND ETHICAL AI​

Including stakeholders to improve algorithms and build an ethical AI.  

Another important desirable feature of HRD-ML is its ability to offer subject matter experts posthoc counter-factual explanations, disclosing how HRD – through its ML models – came to its recommendation by providing the smallest change that can be made to achieve a desirable outcome, in this case, risk reduction.  

While offering intelligibility to ML outputs, it provided means for recourse to important actions (reverse a recommendation) while not requiring full disclosure of the ML code running behind HRD ML models.  

Space and client confidentiality precludes a detailed exposition of how Neal Analytics interpreted and applied other key ethical principles during the design and deployment of HRD-ML.  

However, it should be noted that fairness criteria and performance metrics such as classification parity, calibration, and control of error rates were fully employed.  

Also, Neal Analytics included inclusive, ethical criteria where representative stakeholders, such as finance, legal, and data science, had their respective outcome preference rankings elicited.  

These ranged from recall over precision and over parity and were aggregated according to prescriptive decision-theory principles to ensure no single stakeholder’s preference rankings dominate.  

The HRD program is growing. We now have a proactive process scoring enterprise deals ‘in the making,’ i.e., during quotes (and revisions) to customers.  

A separate compliance department now manages compliance reviews based on the outputs of HRD-ML data processes. HRD has also expanded to other licensing programs.   

By the end of its first year of operation, also HRD-ML flagged deals in multiple licensing programs the customer is offering.