Agentic AI meets regulated quality: Inside Honeywell’s record processing agent

Replacing a manual, human process with AI requires thoughtful consideration and loads of teamwork, especially when the process involves a highly regulated industry. Honeywell Life Sciences’ Martin Dowdall (pictured), chief product officer, and Eric Hartye, principal architect for AI, pulled back the curtain on a recent co-innovation project with a global medical device manufacturer at this year’s Honeywell Users Group meeting in Phoenix.

The scenario will be familiar to many in the industry: rising complaint volumes, a mandate to do more with less, and the unrelenting demands of regulatory scrutiny — all against a fixed deadline.

“I liken it to the year 2000,” Dowdall noted. “On December 31, 1999, you were ready or you weren’t. Nobody was going to push the date back. That was a key driver to getting things done very quickly.”

The regulatory reality is uncompromising. Authorities do not accept a lack of personnel as an excuse for unmanaged or delayed quality records. Everything must be traceable, compliant and thoroughly documented. The project required a scalable solution capable of maintaining a complete audit trail while processing data at a speed humans alone could not match.

Enter the record processing agent

The cornerstone of the solution is Honeywell’s record processing agent. Built on the Salesforce platform, this agentic capability uses AI to automate the completion of quality records. Crucially, the system is architected to operate independently of Honeywell’s TrackWise Digital quality management system (QMS), yet it remains fully capable of orchestrating TrackWise Digital properties.

Hartye explained how the agentic architecture was broken down into composable units called skills:

• The LEGO analogy: Think of skills as individual LEGO blocks. One skill might break an incoming complaint into component pieces; another determines reportability; a third applies specific regulatory classification codes.
• The transformation skill: The initial prototype focused on transformation — taking unstructured input context, such as text or excerpts from a standard operating procedure (SOP), and using it to populate a standardized quality record.
• Architected for reuse: Because these skills are modular and not tied to a single workflow, a capability built for complaint processing can be repurposed across other quality records in the future.

Training AI vs. training humans

A key insight from development was how AI agents interact with SOPs compared with humans. Hartye explained that optimizing an AI agent is like onboarding a new employee. “You wouldn’t take your 800-page SOP, thump it down in front of them at their computer, and say, ‘Let’s get to work,’” Hartye observed. Instead, teams must curate high-value, contextual excerpts from the SOP to provide the agent with precise, actionable instructions. Treating SOPs as data — structured, curated and continuously refined — proved to be one of the most transferable lessons of the project.

Trust by design

In a regulated environment, “black box” AI is unacceptable. “Every decision is transparent, traced and improvable,” emphasized Hartye. “Requirements were shaped by the co-innovation from day one.”

Honeywell integrated a robust observability layer anchored by three primary pillars:

• Human in the loop: Configurable at the individual skill level, ensuring human operators remain in the driver’s seat to approve or reject specific agent contributions within their existing daily workflows.
• Rationale: The agent leverages natural language generation to explain not just what it decided, but why, eliminating opacity.
• Audit intelligence: Generates digestible artifacts and actionable insights directly within the system to ensure seamless traceability during regulatory audits.

Co-innovation in new territory

“Both organizations were in genuinely new territory,” said Dowdall. “There was no playbook for agentic AI in life sciences. We brought our platform depth, AI architecture and composable skills approach. Our partner brought deep QMS domain expertise — real-world workflows, regulatory requirements and the operational context that made the solution production-grade. The learning ran both ways. That mutual investment is what future customers build on.”

Today, the complaint pipeline — from parsing multilingual, unstructured narratives to executing downstream logic — is orchestrated autonomously by AI skills, with humans approving where it matters.

For organizations looking to take a similar path, the lessons are clear: assemble your brightest technical and quality experts, establish firm target dates, treat SOPs as data, and find a partner willing to share the operational risks of the AI frontier.