How payroll giant ADP practices AI governance and human oversight

For the human capital management and payroll company, embedding AI in its products and maintaining strong AI governance practices are not mutually exclusive.

While ADP uses AI internally within its IT team for coding and to automate tasks such as AI image and video generation, it has also included generative AI features in its ADP Assist tool. ADP Assist is a suite of AI agents that work across HR and payroll workstreams. The agents help reduce pressure on HR and payroll teams.

As an organization with a pervasive presence in enterprise HR and payroll, it has made responsible AI a key ingredient in the tools it releases to its clients.

In this Q&A, Helena Almeida, vice president, managing counsel and AI legal officer at ADP, discusses the company’s responsible AI practices and principles, at a time when much public attention is focused on AI risk and safety.

How has ADP’s responsible AI shifted as generative AI has become more mainstream?

Helena Almeida: ADP's been using different versions of AI and machine learning for years. We have had a responsible AI program for a long time. We have had a set of ADP AI and data ethics since 2018 that govern how we develop and deploy AI.

Around the same time, we formed our AI and data ethics council, comprising internal subject matter experts and external experts, spanning legal, privacy, ethics, HR, and product expertise from both internal and external providers. We have a team that is always looking at uses of technology and AI to make sure we have the right perspective.

We have always been thinking about responsible AI. As AI technology has shifted to include GenAI uses, we keep looking at our processes, we keep looking at our cross-functional teams within ADP to look at AI governance, and we keep updating them. The essential premises are making sure it is fair, transparent, secure and that a human is involved. The key decisions that affect people have stayed the same.

How does ADP handle varying governance layers across states and countries, especially given that U.S. states and countries have different laws and regulations?

Almeida: The common principles among all these laws are the same. The principles are recognizable transparency for users, explainability, risk management, privacy and security, and maintaining monitoring programs. All these things together appear in different regulatory schemes.

We feel these are principles that we have had in place for a long time, and we've weathered a lot of change.

We have guiding principles that help us design products. We are staying attuned to changes and regulations. But we have noticed that theme, and we feel comfortable that our program is considering things so that these regulations are covered.

Does the possible arrival of regulation determine what AI products you use as an organization or how you embed AI within the tools that your customers use? For example, if the U.S. were to create a federal AI regulation framework or agency, would that make a difference?

Almeida: There are parts of the world that have prohibited certain uses of AI that would be unsafe for end users. Those are not things that ADP would be involved in, whether there were laws or not. What impacts us when we look at the laws is how we can ensure that this product has the levels of transparency and explainability required for a particular use case.

We are always very careful with what we do because, at least in the US, we pay one out of six workers. These workers have personal experience with ADP and receive their paychecks from ADP.

One of your principles is to ensure that a human is involved. How do you ensure that humans are not just automatically approving what the AI is doing?

Almeida: We cannot have it be a ceremonial checkbox; the human takes a particular step just to say we have a human loop. We really need a meaningful opportunity for a human to participate. We need to make sure that the person understands the rules of AI and has enough context to figure out when there are unusual or questionable results.

Human review should not be a reflexive requirement for every AI-supported task, because not all AI-supported tasks carry the same risk. So, a design that affects somebody's pay, their employment or their legal rights deserves a design that creates moments for human interaction. Low-risk administrative tasks can be safely automated and do not require human intervention. That is part of what I do every day: think about where we need a human to be involved and then work with the product people to design.

Some are obvious, some are harder to anticipate upfront.

For example, there is something called automation bias. If the AI tells you that there is a right answer, people tend to believe it and move on. We want to make sure that is not happening. We always pilot things we use internally. We look at our clients’ use of them, and sometimes we see that people are spending way more time on something a little more administrative and safer than something we thought they would spend more time on. Then we adjust the design to ensure we are getting the right incentives to spend time on what matters, because it is important to give our clients tools that make them more productive.

Editor’s note: This Q&A has been edited for clarity and conciseness.

Esther Shittu is an Informa TechTarget news writer and podcast host and producer covering AI software and systems.