Hiring AI: Why Implementing AI Systems is Not “Set It and Forget It”

When some companies initially consider AI tools for their quality system, they often think of them like any other software: install it, configure a few settings, and let it run. But this mindset misses a critical shift. Using AI effectively isn’t like buying Microsoft Word. It’s more like hiring a digital employee.

AI systems, both machine learning (ML) and large language model (LLM) based, are fundamentally different from traditional tools. They don’t just calculate or organize; they also reason, interpret, summarize, and make judgments - similar to a skilled employee. And just like employees, they need training, evaluation, feedback, and periodic re-training to stay effective.

1. Training and Onboarding: You Don’t Just “Plug and Play”

Just as you wouldn’t expect a new hire to excel on day one without training, AI tools need time and effort to learn your processes, terminology, data, and risk tolerances. While many LLM-based tools come pre-trained on general knowledge, they must be adapted to your specific quality system, often through fine-tuning the model, providing examples, or in-prompt training.  And there are very few off-the-shelf ML-based systems – each one must be trained from the ground up using your unique dataset, which solves a very specific problem for your organization.

Example: You might need to train your AI to distinguish between “correction” and “corrective action” in CAPA records—something an experienced quality engineer learns from training and experience, but an AI may not inherently or consistently understand.

 2. Performance Evaluation: AI Needs Feedback Loops

Human employees are reviewed and coached to improve. The same is true for AI.

AI tools should be evaluated on how consistently and accurately they perform their tasks.  If your AI is reviewing complaint files for completeness, but it sometimes misses unlinked investigation reports of missing product codes, it’s a signal that feedback is needed.  There are a variety of reasons why this might occur, but having a feedback system is the crucial first element to trigger an investigation and retraining.  Think of it as your post-market surveillance system (PMS), just like you would set up a PMS to monitor customer issues with your medical device.

In practice: Teams should implement regular performance checks (e.g., weekly spot audits or benchmark testing) to assess accuracy and consistency of how well the AI is performing—just like you’d review an employee’s work output.

Suggestion: Build a simple thumbs up/down feedback button into your AI interface, similar to ChatGPT, so users can rate the usefulness or accuracy of AI-generated responses in real time. This direct feedback can be reviewed by administrators and used to continuously improve the system.  You can even include text boxes to allow users to give comments beyond just good/bad.

3. Retraining and Maintenance: AI Needs to Keep Learning

Just as humans attend training to stay current on new regulations or product changes, AI needs updates to remain accurate.

Scenario: Your complaint audit AI has been working well for months, flagging missing investigation links and ensuring MDR decisions are consistently evaluated.

Then, a new product is released, and a novel design issue emerges: the connector pins in a home-use glucose monitor are prone to bending during insertion, leading to intermittent device failures reported by patients.

This failure mode was never present in prior data. At first, the AI misidentifies these complaints as user error, instead of correctly classifying them as a design issue.

To fix it, the AI needs to be updated - whether by adjusting its prompts, retraining it with new examples, or fine-tuning a model with labeled cases of the new failure mode. This is the AI equivalent of coaching an employee on how to spot a newly emerging issue.

This process isn’t optional. It’s essential for maintaining reliability and trust.

4. Accountability and Oversight: You Still Need Human Review

Even the best AI is not 100% autonomous. Like a junior auditor, it needs someone to check its work. In regulated environments, this oversight is not just best practice.

Think of AI as a force multiplier. It accelerates review cycles, highlights risks, and finds patterns you may have missed. But the responsibility still lies with you. AI helps make better decisions, but it doesn’t make them for you.

In Practice: For example, a complaint coding AI might suggest the top three most likely codes based on the complaint narrative and device history. But the complaint specialist still has to review those suggestions, apply their judgment, and select the correct code. This keeps decision-making with the human while letting the AI surface likely options and reduce cognitive load.

5. Adaptability to Process Change: AI Doesn’t Always Improvise Correctly

If you change your complaint categorization structure, or your MDR thresholds, a human can often adjust on the fly. An AI won’t unless it's explicitly told to.

So when processes or products evolve, which they always do, you must proactively re-align the AI’s understanding to match.

Final Takeaway: Think of AI as Hiring Digital Staff, Not Buying a Tool

Implementing AI in your quality system is not just buying software. You’re onboarding a digital employee, which means you need to dedicate resources to constantly monitor and maintain that digital employee. That means:

• Teaching it your process

• Checking its work

• Giving it feedback

• Retraining it as needed

• Holding it accountable

• And, most importantly, enabling it to help you scale quality

  
  

AI can supercharge your compliance efforts, but only if you treat it like the dynamic labor it is, not a static tool like Excel.

Want to see how myQMS.ai can work as a digital teammate in your quality system? Schedule a demo today