The AI revolution is fueling far-reaching business transformation. This short series of ‘AI & REGTECH’ posts has been created by the data scientists and benefits specialists at BENEFITSCAPE to help employers understand better the emerging benefits of AI for ben admin and compliance.
In Part 2 BENEFITSCAPE digs deeper into the different types of AI engine and their uses — as well as misuses. But Before we do that, it is worth repeating a key point made at the outset of Part 1.
In Part 1 we strongly advised all AI use to adhere to the National Institute of Standards and Technology’s AI Risk Management Framework [AI RMF]. AI can help employers with compliance ben admin tasks but also raises its own compliance issues, such as data privacy regulations and legal liability. If you have any uncertainty about your company’s adherence to AI RMP guidelines, please contact BENEFITSCAPE. As leading specialists in the application of advanced technologies to ben admin and compliance, we can help you review your practices.
In Part 1 we also stressed the importance to understand the diversity of different types of AI engines, their different uses, training, and configurations. Below we lay out a framework for understanding these differences in the specific context of ben admin governance, risk & compliance.
_______________________
Weak vs Strong AI
These labels for two different types of AI can be misleading.
‘Weak’ means AI designed for specific tasks — and not designed to demonstrate a more wide-ranging and adaptive human-like intelligence. AI designed just to play chess is an example of ‘weak’ AI but does this one task brilliantly well. The famous Alan Turing AI test can be something of a red herring for useful applied AI. Turing focused on AI’s ability to fool us into thinking it was human and not on the ability to outdo us at specific thought processes or tasks.
For many compliance purposes, a specialist focus makes ‘weak’ AI more relevant, reliable, and valuable. AI used for ben admin compliance needs to analyze complex but specific, finite data sets for errors, anomalies, or other unusual and therefore high-risk patterns — and to do so far more powerfully, rapidly, and accurately than a human can.
‘Strong’ AI, such as ‘large language models,’ is far more wide-ranging but also therefore far more subject to errors, since it learns from analyzing open-ended data sets, including erroneous data, outdated data, or data that is itself AI generated. Hallucinations are a symptom of strong AI.
Sometimes the rules are just the rules, however complex; the overdetermined use of strong AI models can be detrimental when total accuracy is required for compliance to pre-determined rules.
_______________________
Generative AI [ vs Discriminative AI ]
Generative or Gen AI is fun and easy to play with. We’ve all been dazzled by the likes of Chat GPT, and Gen AI steals a lot of the headlines. It’ll write you a passable Valentine’s poem or fake a new background for your selfie. It’ll even structure an argument for or against a hot topic. Once the engine is trained on the available data, it can generate new content based on the learned patterns.
You can think of Gen AI as very clever plagiarism, analysing and mashing up many sources to create something ‘new.’ Hence the intellectual property and other legal issues that have been opened up as AI generates its own new compliance questions.
_______________________
Gen AI Natural Language Processing & Output
Key to the popular appeal and usability of Gen AI engines’ is their ability to process natural language data — in terms not only of the questions we ask, the tasks we set, and their output but also the engines’ ability to learn and self-train by processing unstructured natural language content.
For what might be called ‘softer’ applications of AI such as ChatBots handling routine customer queries, Gen AI engines can be a huge boost to cost-saving, and when deployed appropriately even improve customer satisfaction.
For example, you can ask any of the more popular AI engines to tell you about different aspects of ACA reporting and compliance. What’s offer affordability? When are the IRS filing deadlines? And the answer will be largely right and certainly helpful to a degree — though it will, for example, on occasions get tax years and calendar years muddled or miss out the granularity needed for your specific operations and unique employee population. BENEFITSCAPE is supporting clients and plan providers in exploring the successful deployment of AI-assisted ChatBots to provide support to employees about their benefits.
_______________________
Decision-Making Limits of Gen AI
Gen AI can seem like ‘strong’ AI given its range and natural language interface. But despite its seemingly magical output capabilities, Gen AI engines excel in relatively weak applications — in other words, at well-defined narrow tasks. They lack the general and adaptive understanding needed to address broader, open-ended challenges such as strategic decision-making or ethical dilemmas or to generate valuable new ideas. See ‘Understanding The Limitations Of Generative AI,’ Forbes, May 9, 2024.
The cautionary and amusing website Spurious Correlations also illustrates well the explanatory limitations of Gen AI with its Gen AI generated explanations of remarkable but spurious data correlations. See, for example, Spurious Correlations, ‘Popularity of the First Name Cameron correlates with UFO Sightings in Florida.’
_______________________
Discriminative AI
The purpose of Gen AI is to generate new data [text, images, music, videos]; the purpose of Discriminative AI, in contrast, is to categorize existing data. Is this image an apple? Is this email spam? Do this medical data indicate a diagnosis? Do these patterns of transaction suggest fraud?
In many ways, Discriminative AI is the more traditional AI and predates the popular explosion in more casual Gen AI use. Discriminative AI has long been established in scientific research, looking for patterns in vast quantities of experimental data; and since its foundation, BENEFITSCAPE has used intelligent tools to analyse ben admin data and help improve the performance of its regtech and compliance services.
Unlike Gen AI, discriminative models use supervised learning as well as unsupervised and reinforcement learning. Their training includes known input data and labels to learn how to discriminate between diverse classes. They focus on minimizing classification errors and maximizing the probability of accurate class predictions. This training makes them valuable tools at identifying irregularities, anomalies, or outlier patterns in complex data sets. This can make suitably trained Discriminative AI a valuable tool for ben admin and compliance.
The emphasis here is on suitably trained. The garbage-in, garbage-out data adage does not go away with AI. And as said before, some rules are just rules, and these can be set in regtech diagnostics without recourse to AI.
_______________________
Discriminative Diagnostics
BENEFITSCAPE regtech services are supported by Discriminative AI diagnostics, trained on known ‘clean’ data to look for anomalies and irregularities in other data sets, such as an employer’s ACA reporting and compliance data.
AI-assisted in this way, BENEFITSCAPE’s specialist ACA Flag & Fix diagnostic services can carry out pro-active risk assessments for employers, auditing real-time employee data from complex HCM databases and other relevant sources well before any of the processed data is subjected to the IRS’s own AI-assisted technologies.
It is worth noting again, as discussed in Part 1, AI and regtech have greatly increased the IRS’s own capacity to audit even the most complex corporate tax filings and this, in turn, has raise the bar on compliance. See ‘Regtech Transforming IRS Powers.’
_______________________
Automation vs Intervention
Some BENEFITSCAPE Flag & Fix diagnostics can be automated. These are the more ‘predictable’ and self-evdiently non-compliant errors of data categorization or input commonly encountered by regtech audits of ben admin compliance data. Other anomalies or irregularities may just be atypical data outliers. These flagged errors or warnings, which could easily be missed without regtech diagnostics, require intervention by BENEFITSCAPE or the employer’s own subject-matter experts to interpret and if necessary correct.
In the realm of compliance and risk management, mission-critical processes should always err on the side of validating human oversight. AI and regtech can hugely transform the accuracy and efficiency of ben admin and compliance — when deployed correctly. Poorly understood and deployed, AI can be as source of its own processing errors with potentially costly implications.
When outsourcing ben admin reporting & compliance services and support, employers should make sure their service provider has both proven benefits and technology expertise. This is truer than even when AI is deployed to support these services. AI should be viewed as a huge, often game-changing support and not a substitute for subject-matter expertise.
_______________________
In Part 3 of ‘AI & REGTECH’ we ill conclude this series of posts from the data-scientists and benefits specialists at BENEFITSCAPE with a concise checklists of insights and considerations for the employer when thinking about how to take best advantage of the breakthrough advances in AI to support benefits administration & compliance.
In the meantime if you have any questions regarding any of the topics raise in these posts — or indeed any other aspect of ben admin on you HCM — please do not hesitate to contact BENEFITSCAPE.
