The Sidekick Becomes a Coworker: What a Quality Leader Learned About Trusting AI With Compliance

A year ago, most quality leaders would have called AI in compliance mostly hype. Christian, a quality and compliance leader who has spent his career in radiopharmaceutical manufacturing, said as much on a recent Accelerate conversation with Qualio. Then he watched the same teams move from asking AI to write an SOP, to feeding it context about their company and their process, to something else entirely: treating it like a coworker.

That progression is worth sitting with, because it maps almost exactly onto the gap between where most quality management systems are today and where they need to go.

The backlog is not a discipline problem

Christian has seen the same failure pattern across small and mid-sized life sciences companies for years. Regulations tighten. Processes get more detailed to close every gap. And the people running production, QC, and quality assurance, often just a handful of people total, are left executing an SOP that has grown more thorough and more burdensome every year it gets revised.

"The challenge that I see is actually that we do not have any support in the process," Christian said. "We still have to manually go through the process, and this is sometimes, and I see this very often, is really the bottleneck." Thirty or forty open deviations. Tens if not hundreds of change requests. A backlog that never fully clears because the team never gets a day without new inputs arriving faster than they can close the old ones.

This is the scaling pain every quality leader recognizes: headcount and complexity grow together, and the team runs out of hours before it runs out of backlog. Adding another auditor-facing checklist does not fix that. Removing the manual bottleneck does.

From sidekick to coworker

The distinction Christian draws is the useful one. A sidekick answers questions when asked. A coworker looks at the work as it happens, flags what needs attention, and can be handed a defined piece of the process to run, with a human checking the result.

"AI becomes a coworker of mine because this is the opportunity when I can give AI certain tasks like I would do to a coworker that will help me to reduce my workload on a daily basis," he said. Not writing the SOP. Reading the CAPAs, the deviations, the change requests as they are generated, and catching the pattern a person under deadline pressure will miss.

His favorite example is root cause. Ask most teams why a deviation happened and the first answer is often "human error," because it is the easiest one to reach for under time pressure. Christian described asking a coworker, human or AI, a few more questions before accepting that answer: has this come up before, did the last corrective action actually hold, is there a system cause underneath the individual one. That is not a faster way to fill out a form. It is a better root cause, arrived at because someone, or something, had the bandwidth to ask the next question.

This is the same principle behind bounded autonomy: an agent that can run a gap analysis, triage a CAPA, or draft a corrective action, with the human still the one who approves it before it becomes the record of truth. The team gets the second set of eyes it never had the bandwidth for. The audit trail still shows a person made the call.

Reactive by design, until now

Christian made a point about the entire category that is easy to miss because it sounds obvious once said out loud: quality management systems, as built, are backward looking. Something happens, you document it, you analyze it, and you use that analysis to do better next time. That is not a flaw in any one company's process. It is how the systems were designed to work, decades before there was a way to look at the data while it was still being generated.

"For the first time now that we have more power to what we can do," Christian said, "our view angle shifts from looking backward, what is happening right now." He pointed to something as simple as the rate of entry corrections on handwritten forms, a number that quietly tells you when a team is under strain, well before that strain shows up as a deviation. Catch the drift in that number three weeks before the finding, and the conversation with the team is "what happened, how can we help," not a corrective action after the fact.

That shift, from a record of what already happened to a signal on what is happening now, is the entire argument for a compliance system built to act continuously rather than one built to store artifacts after the fact. A system of record can tell you what went wrong. It was never built to tell you it is about to.

No need to reinvent the rulebook

One thing Christian pushed back on, gently but firmly: the instinct to write an entirely new regulatory framework for AI because it feels unfamiliar. His suggestion is more useful and considerably less work. Treat AI the way the industry already treats an analytical method or a new coworker. Qualify it. Give it an onboarding, a training plan, a way to check it understands what it is doing in its own words. Validate it against defined statistical expectations, the same way a lab validates an assay, and revalidate it on a schedule.

"No need to reinvent the wheel just because it is a new concept," he said. "Just apply the same concepts over it for the right level of control." That is the case for validated, regulatory grade AI: not AI that is impressive in a demo, but AI whose outputs hold up to the same qualification rigor the rest of the quality system already runs on.

What the team gets back

The payoff Christian described is not efficiency for its own sake. It is time. Time to stop surviving the backlog and start asking what the team should do next. "This needs this state of mind where you are relaxed, where you know, I am in control," he said. "Let's have a breakout session, a little brainstorming, grab a coffee because we have the time today."

He was careful to draw a line here that matters: AI does not replace the human in that picture, it makes room for the part of the job only a human can do. Judgment, creativity, the next idea nobody has had yet. The coworker handles the deviation review. The person gets the five minutes back to think.

That is the case for agentic compliance in one interview: not a better filing cabinet, but a system that works the backlog down in real time so the people running quality can do the part of the job that was never supposed to be reactive in the first place.

Want to see where your own team sits on that spectrum, from reactive backlog to continuous, proactive compliance? Get your quality score →

Prefer to hear it straight from Meg and Christian? Watch the full Accelerate session on demand →