Human resources gets AI upgrade with ADP Assist

Human resources (HR) has been the butt of many a joke in popular culture (see Toby from “The Office” on NBC), and has in my experience in real life been viewed alternatingly by employees as an annoying scold they get sent to for making impolite jokes; as the kindler, gentler smoke screen of the company’s more outwardly amoral (or immoral) top brass; or, in the best case scenario, the last refuge for employees dealing with workplace toxicity from managers or colleagues.

But one thing is for sure, the demands on HR professionals are not slowing down, especially at a time when the workplace has become more fractionalized and widely distributed than ever before. HR and employee benefits software systems provider ADP (formerly known as Automatic Data Processing) thinks generative AI can help HR through this time of upheaval and increased demands and is introducing a new tool, ADP Assist, to just that.

A new kind of HR chatbot

ADP Assist is a new chatbot assistant rolling out to select ADP customers — eventually all of them are planned — that HR managers can turn to for help answering common questions and flagging brewing problems.

It is also designed to help offload some of the inbound, tedious and repetitious questions they get from employees.

The chatbot aims to facilitate better decision-making by providing data-driven insights and simplifying HR tasks with intuitive, conversational interfaces.

HR managers can ask it questions about company policies and get a response grounded in their data, and employees can do similar if they want to know where to find information about their benefits, time off, payroll, or other common questions.

A compelling example screenshot of ADP Assist shown off on the company’s website displays a conversation between a user and the chatbot asking it to file taxes on the user’s behalf.

One key feature of ADP Assist is a “payroll inspector” that ADP says allows its clients to “increase payroll accuracy” by analyzing and flagging errors in payroll to managers.

Another feature, “smart analytics,” claims to offer the ability to show HR managers and executives where to focus their time when it comes to employees, such as detecting if turnover rates or overtime are increasing in specific departments.

“It’s not just another technical solution; it’s a comprehensive experience that combines ADP’s unmatched HCM [ human capital management] dataset and expertise to empower HR professionals, leaders and employees,” said Maria Black, president and chief executive officer of ADP, in a press release from the company issued last week.

ADP also posted the following video on YouTube featuring its chief data officer Amin Venjara describing some of its capabilities and time-saving processes.

What’s under the hood?

Asked by VentureBeat for more information about what kind of foundation model was powering ADP Assist — and whether it was an LLM such as OpenAI’s GPT-3.5/4/4 Turbo, Anthropic’s Claude, Cohere’s Command, or open source alternatives such as Meta’s Llama 2 or Mistral’s Mixtral — ADP chief product owner Naomi Lariviere provided the following response to VentureBeat via an email from a spokesperson:

“At ADP, we’ve been working with AI and machine learning in our products for many years, leveraging the power of our unmatched HCM dataset to help businesses succeed.

We understand the great potential that new technologies like generative AI can provide as we design and develop innovative solutions that address the critical needs of our clients and associates.

As part of our broader generative AI strategy, we have the infrastructure to incorporate these new innovations in the right way for the right reasons and scale them across our business.

ADP’s extensive and unique dataset is used to train ADP’s AI to anticipate what users want and need. As the power of AI grows, we’re remaining laser focused on the human experience.“

As you can see, the response doesn’t really provide much clarity one way or another one what underlying models ADP might be using for this new chatbot.

Yet the product owner was able to share with VentureBeat that the new tool was trained on its own proprietary dataset “the industry’s largest and deepest HCM dataset with over one million clients spanning 140 countries and 41 million wage earners globally.”

“ADP’s extensive and unique dataset is used to train ADP’s AI to anticipate what users want and need. ADP Assist only uses the data required to provide the insight, next step, or information the user needs to make the next best decision.”

To me, that reads that the private companies trusting ADP with their employee data now have themselves to thank for providing to ADP the information it used to build this new feature, something we’ve seen similar outlines of before with Adobe using its Adobe Stock user-generated content to train its Firefly model, as well as Meta training its Imagine image generator model on Facebook and Instagram user submitted photos.

Safeguards, security, and availability

ADP stated in its press release that Assist is “in early release with some ADP clients today with a planned rollout for all clients.”

Asked which clients are using it now, Lariviere told VentureBeat via email: “We have clients ranging from small businesses to large enterprises using ADP Assist which operate across several industry sectors, including manufacturing, logistics, and others.“

She further noted that ADP Assist would be made available only to current customers of ADP’s numerous other software offerings — ADP has at least 10 different product lines listed on its website, including Workforce Now and TotalSource — not as a stand alone offering. But she noted it would be made available to current customers free of additional charge.

As for ADP’s approach to data security and reducing hallucinations in AI responses, Lariviere said this:

“We’ve adopted a rigorous set of principles and processes to govern the use of newer technologies, including operational monitoring of recommendations made by AI/ML technologies. Our models and algorithm are routinely tested to ensure they perform as expected and we are addressing the potential for AI/ML to introduce, perpetuate, or amplify bias. In some cases, when appropriate, ADP’s AI & Data Ethics Board will review proposals to make sure data is being used fairly and in compliance with both legal requirements and our own ethical standards.“

As for who controls what data the models have access to, and whether customers can limit access of ADP Assist to specific data, Lariverie said:

“We provide our clients with information about how our AI and ML models operate, their proper use, and their limitations, so that clients can implement those models in accordance with their design and purpose, operate them effectively, and use their outputs as intended.“