What compliance innovators learned bringing AI into regulatory research

Public LLMs for regulatory compliance can speed up research, but out of the box they can’t guarantee that an answer is current, cited, or drawn from a credible source. That gap matters when a product’s market access depends on the answer. In 3E’s Ask the Innovators panel, leaders from Milliken & Company and ERM joined 3E to discuss what they learned bringing AI into regulatory research, from the risks of generic tools to what success could look like a year from now. Here are the takeaways.

Why AI adoption pressure is rising for compliance teams

Compliance teams have always faced growing regulatory complexity. What’s changed is the push from leadership. Across more than 9,000 compliance conversations 3E held over the past year, AI adoption pressure came up in 43% of calls in September 2026, up from 20% a year earlier. Enterprise Copilot rollouts, leadership efficiency goals, retiring experts, and headcount constraints are all driving it.

Steve Smith, technical partner at ERM, sees four drivers across clients: efficiency, accuracy, outside expectations, and automation. “Compliance teams today are being asked to monitor more and more regulations, and organizations are not adding staff to the same extent,” he said. Leadership and investors also expect companies to show how they use AI, not just that they’ve rolled it out.

For Liv Gamble, advanced data analyst at Milliken & Company, the urgency comes from volume and pace. As AI models respond faster, the expectation that compliance teams will deliver insights at the same speed keeps rising.

Why public LLMs fall short for regulatory decisions

Both panelists see a place for general-purpose AI tools. The risk comes from using them as-is for regulatory research.

Data privacy comes first. Milliken uses commercial LLM tools inside an enterprise environment, with do-not-retain agreements in place. Accuracy is the bigger issue for regulatory work. “We’re dealing with legal documents and things that have true legal impacts on our business,” Gamble said. That’s why AI outputs need to trace back to trusted, verified sources.

ERM learned this firsthand. About two years ago, the firm prohibited its consultants from using responses straight out of public LLMs. “They make it sound very believable,” Smith said, and separating accurate answers from false ones took significant time. Even with data security managed, someone still has to reread the source text to confirm each answer, which gives back much of the efficiency gain.

What trusted AI for regulatory research requires

3E’s conversations with compliance teams point to four business requirements for AI in product compliance workflows:

• Cited sources: links to the actual regulation behind every answer
• Explainable logic: reasoning and rules written for regulatory experts
• Data protection: no training on customer data, with certified AI governance
• Controlled access: IT sets what agents can read and write

When ERM’s consultants began testing the 3E Agent, they challenged answers they initially thought were wrong. In most cases, the 3E Agent’s answer held up. “The whole thing is only as good as the data that it’s grounded and curated on,” Smith said. He added that 3E Agent surfaced relevant requirements in later chapters of a regulation, the kind of detail a time-pressed reviewer can miss.

Keep humans accountable for the final call

Grounded data doesn’t remove the need for expert judgment. “Even with a private model, with trusted sources, it’s best practice to have the human validating that output and ultimately giving the final judgment,” Gamble said. Smith agreed: “There will always be that verification step.”

The goal is to change where experts spend their time. With a trusted starting point instead of a blank slate, specialists spend less time gathering information and more time interpreting what it means for the business.

Three ways to bring regulatory intelligence into your AI ecosystem

Bobby Shackelton, VP of product management at 3E, outlined three ways teams can work with the 3E Agent for regulatory research:

• In the AI tool your team already uses: connect the 3E Agent to Microsoft Copilot, ChatGPT, or Claude, and get answers grounded in 3E data and cited to full-text sources.
• In the 3E AI Platform: while IT completes setup, teams can explore and trial the 3E Agent in a ready environment and see additional workflows the 3E Agent can support.
• Agent to agent: a company’s own agents can call the 3E Agent as part of a scheduled job or routine process, such as horizon scans against a live portfolio.

In a live demo, Shackelton asked Copilot for one CAS number’s status on U.S., EU REACH, and Canadian inventories and got a cited, jurisdiction-by-jurisdiction answer in a single response.

From asking AI for data to asking AI for results

Smith’s view of success a year from now: “We’ve got to stop asking the AI for data, and we’ve got to start asking the AI for results.” Instead of asking what changed in a regulation, a team would ask what the implications are for a product line next year.

Gamble described a similar goal: AI tools that scan for changes continuously, put them in context with the organization’s own data, document them, and route them to the right subject matter expert. Through a launch partnership with 3E, her team is piloting workflows and plans to track success metrics as it scales.

Advice for compliance and EHS leaders getting started

Gamble offered three pillars:

• Start with the problem: bring the right people together, find where your team spends the most time gathering information, and define the business problem first.
• Invest in training: AI changes quickly, and using it well is a skill teams have to keep building.
• Set governance early: decide which models you’ll use, public or private, and make sure you trust the sources AI responses cite.

Smith’s advice was more pragmatic: teach people to prompt. When users tell him an agent isn’t giving them what they want, about 80% of the time the fix is a better prompt. You have to “start thinking of your AI tool as an assistant,” he said.

Hear the complete discussion: watch the Ask the Innovators panel on demand.