Why AI Model Updates Need Change Control: AAPS PharmSci 360 2026

At AAPS PharmSci 360 2026, Manish Baviskar, senior validation engineer, Hubot CDMO, speaks with PharmTech about how validation and quality professionals can prepare to work with AI-enabled systems. The interview previews his Careers 360 session, "Building Artificial Intelligence Skills in Pharmaceutical Manufacturing and Validation," which covers GAMP 5 risk-based validation, ALCOA++ data integrity, and digital quality systems.

Baviskar says becoming AI-ready starts with the fundamentals. "My recommendation will be to start with the basic things about AI," he says. He suggests spending the first 3 months learning how AI models work and where they fit in pharmaceutical processes, such as using language models to draft standard operating procedures. After that, professionals need to keep up with the FDA's changing expectations. Baviskar notes that the agency has issued 6-7 AI-related draft guidance documents over the past three years. Those documents now sit alongside the 10 guiding principles for good AI practice that the FDA and the EMA released in January 2026. AI output still needs independent verification by quality and subject matter experts, he says. The FDA reinforced that point when it cited a manufacturer for failing to review AI-generated GMP documents.

The main difference Baviskar sees is between static software and systems that keep learning. Conventional automation software runs a fixed version that changes only when a developer updates it. AI keeps learning. "AI will learn from the data you are feeding to that model, and it automatically updates," he explains. Data from hundreds of batches builds up, a model can show which way attributes such as pH are drifting. That ability brings a control obligation. "As a pharmaceutical professional, we need to make sure that AI is not updating the version by itself," Baviskar says. He adds that model updates should go through change control so they stay traceable.

Before deploying AI under GMP, he says, teams should define its purpose, the business need, and the regulatory requirements. They also need to understand model drift and bias, an area other experts call the thinnest part of AI governance.