For leaders, the main lesson is not about adopting more AI; it is about deciding where AI is allowed to act autonomously and where it must remain tightly supervised. In a payroll and HR environment, โmostly rightโ is not a workable standard. The management implication is that AI value has to be filtered through control requirements first, then productivity gains. That changes how use cases are approved, who signs off on them, and what evidence is needed before scaling.
This also points to a different operating model for AI delivery. ADPโs lab behaves like a gated innovation pipeline, with clear success criteria and an explicit path from experiment to product. For CIOs and transformation leaders, the trade-off is speed versus trust: broader access to AI can accelerate engineering and service work, but only if guardrails are strong enough to prevent inconsistent outputs in regulated processes. The next question is whether your organisation has comparable entry and exit criteria for AI use cases, or whether projects are being evaluated inconsistently by each business unit.
There is also a portfolio issue. Not every AI opportunity deserves the same level of model sophistication. The article suggests a practical split between high-risk, deterministic use cases and lower-risk productivity tools for staff. Leaders should ask which tasks can be safely automated, which need human-in-the-loop review, and which should remain non-AI until governance matures. That framing helps avoid two common mistakes: over-engineering low-value work, and under-controlling critical workflows.
For boards and executives, the key question is what proof of control will be required before expanding AI into customer-facing or decisioning roles. The answer will shape vendor selection, audit readiness, staffing, and the pace of deployment more than the technology itself.
Regulatory compliance looms large over how ADP deploys AI. Yet chief AI officer Roberto Masiero finds ways to give his engineers an edge with the technology.
Building an innovation pipeline
Some of the new ideas and products ADP explores evolve from external connections and sources. Masiero said the lab works with other parts of the company, such as the ADP Research Institute, which he said has a strong connection to Stanford and other universities across the country. The lab also works with ADP Ventures, the company’s corporate venture arm, to build relationships with startups and ventures the company might back. This can include investments or commercial partnerships. “We learn a lot from those relationships,” he said. This is the latest installment in Outside In, InformationWeek’s series that explores how large enterprises scout, cultivate and scale innovation from external technology partners and startups. The innovation lab is held to certain criteria for success, such as revenue and the number of clients a new product or deployment could reach, Masiero explained. Projects that meet those goals advance to further development and adoption. For example, ADP Mobile was born in the lab and now has more than 20 million users, he said. The ADP Marketplace, a platform for sharing data across HR systems and connecting with ADP APIs, likewise got its start in the lab. Masiero said ADP has introduced conversational AI to its payroll management platform for small businesses, which did require constraints on generative AI for the sake of compliance.A long-term investment in AI at ADP
The innovation lab started about 13 years ago, and the dive into AI began some six years ago, though the technology was not called AI then, Masiero said. “It was more machine learning, it was more using models that we train to do specific tasks like understanding the intent of the user or using it for entity recognition, obviously for translation and things like that,” Masiero said. As large language models emerged, however, they introduced a new challenge. “[Generative AI ] has a tendency to not follow instructions to the letter,” Masiero said, requiring careful oversight. “A big part of what we do in the lab is to try to put in the guardrails, to ground the model. We need to be 100% right.” The lessons learned in the lab now shape how ADP uses AI internally. Masiero said that one way ADP uses AI is to help its service staff learn from collective experiences and content in order to make better decisions and drive productivity. ADP’s IT teams are also benefiting from AI. “On the development side, we use AI to accelerate our development cycles and code,” he said, noting that all engineers at the company have access to AI. “Some of them don’t even write a line of code. They basically describe what they want and let the AI do all the programming or all the syntax for whatever feature they’re creating,” he said.Guardrails bring focus to next steps
ADP is also considering offering more AI on the client side to create job descriptions and job posts when clients seek new hires, Masiero said. AI might also be used to organize work schedules by analyzing patterns to distribute personnel and skills where needed within companies. “Instead of spending two hours trying to put people in the right boxes and in the right shifts in the right store, let AI do it,” he said. As ADP uses AI internally to automate certain tasks, AI image and video generation has been explored for training purposes. The company is also exploring small language models for certain applications but has not yet deployed them, mindful that it operates in a compliance-heavy industry, Masiero said. That focus on compliance includes letting AI handle specific jobs rather than giving full control to an AI agent to oversee a crucial operation without oversight. “I don’t think we’re there yet,” he said. “But we need to obviously think about it and understand how to create the guardrails necessary to move the low-value, repetitive tasks we have agents that can complete them.”Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.