Written by Dr. Richard Soltero, President, InstantGMP, Inc.

Quality has a way of finding the hole in your test plan.

When I was at NOBEX we uncovered a problem with an oral insulin tablet I developed. The tablets didn’t hold up in packaging even though they met their friability spec. It turned out that friability got worse over time, but we only checked right after production.

I have been thinking about that batch a lot lately. Every quality conference I attend now has a session on using artificial intelligence to predict quality issues. The implication, sometimes stated and sometimes left hanging, is that a model trained on enough manufacturing data could have flagged that drift before it ever showed up on a packaging line. Maybe. But probably not for the reasons most people assume.

The Question I Keep Getting

Quality managers in pharmaceutical, dietary supplement, food and beverage, automotive parts, cosmetics, medical device, and contract manufacturing operations all want to know the same thing: can AI actually predict the next deviation, the next out-of-spec result, the next CAPA before any of them happen?

AI is becoming an essential tool in large scale quality management systems. That part is true. The harder question is whether it can be predictive at the scale most manufacturers actually run at. The honest answer is “sometimes, and probably not yours.”

What AI Is Genuinely Good At

Where AI excels is in researching and cataloging information from large data sets. Hand a model years of batch records, environmental monitoring trends, deviation summaries, supplier certificates of analysis, and complaint history, and it will surface relationships a human team would never have time to chase down. The link between a temperature excursion in March and a dissolution failure four months later. The supplier whose particle-size readings drift just before a process upset. The deviations that cluster around a particular line, shift, or starting material.

In my opinion, this is where AI excels. We are adapting AI as a tool to improve our electronic software solutions and to improve our quality management systems, and most of the work we are doing inside iQMS.ai sits squarely in this category: helping people search faster, summarize cleaner, and write better SOPs. Useful work. Not crystal-ball work.

Why Scale Is the Whole Argument

Here is the part most vendors leave out of the deck. Using AI to predict quality issues is like using a backhoe to dig a hole for a tree. It’s overkill for planting a flower. AI needs a lot of data to make an extrapolation of what might happen.

The operations where prediction actually works tend to be the same ones that already had pattern-detection programs before AI got fashionable. Pharmaceutical manufacturers running multi-site operations under 21 CFR Part 211. Automotive parts plants under IATF 16949 producing millions of components a year. Food and beverage plants under FSMA Preventive Controls running continuous lines. Those facilities generate enough repeatable history for a model to find a real signal.

Where AI is not effective in making predictions is in smaller manufacturing operations. There just is not enough data to have a history of repeatable problems. A 503B compounding pharmacy producing forty batches a year of a sterile injectable does not have the dataset machine learning needs. A small dietary supplement contract manufacturer running short campaigns under 21 CFR Part 111 has the same problem. A medical device startup under 21 CFR Part 820 making clinical-trial product has even less. The math is simply not there.

People Find Quality Gaps Before Models Do

Most quality problems that occurred in my 30+ year career started with a single issue that exposed a gap in the quality system. The NOBEX friability story is the clearest example I can give. The gap was not hiding in our data. It was hiding in our test schedule. We never thought to retest friability after the tablets sat in the package, because nothing in our protocol told us to. A predictive model trained on the testing we actually did would have seen nothing wrong, because nothing was wrong by the test we ran.

I have built several quality systems over the years and even with the best consultants’ help there have been gaps that show up only after the system is put into use. That has not changed. What has changed is that companies now want to skip ahead to AI before they have closed those gaps. Sequencing matters, and getting it backwards is expensive.

Dashboards First, Models Later

InstantGMP is taking another approach to using AI in quality management systems. We will expect customers to build their own dashboards that track information that they think is predictive based on data in their manufacturing execution systems.

The reason is straightforward. The people closest to the process know which indicators matter. A QA director at a cosmetics manufacturer under MoCRA already knows that ingredient-lot variability and adverse event signals are her early warnings. A quality manager at a chemical plant under ISO 9001 and ISO 14001 already knows that solvent recovery efficiency tells him when a batch is drifting. A 503B director already knows that cleanroom particle counts in the hour before fill are the canary. Those are not insights that come out of a generic AI model. They come from operators who have lived the process.

Our job, as the platform underneath, is to make those indicators visible, traceable, and tied to the live data already flowing through the manufacturing execution system. Build the dashboard that watches the metrics you already know are predictive. Let the system alert you when those metrics drift. That is a far more honest use of digital infrastructure than asking a model to guess at things it has never seen enough of.

In my experience, the companies that I saw succeeding with quality data were not the ones with the fanciest software. They were the ones whose batch records, electronic batch records, inventory transactions, and deviation history all lived in the same system and could be queried together. Companies that still rely on spreadsheets will not be well served by AI since the data is disconnected and often incomplete.

A Practical Path Forward

If you are running a regulated operation and trying to figure out where AI belongs, the answer is sequencing. Get your batch records electronic and timestamped. Get your deviations and CAPAs into a system that links them back to the affected materials, equipment, and SOPs. Get your environmental and process data into the same place where it can be reviewed in context. Once that foundation is real, your operators can build the dashboards that catch trouble early. AI gets layered on top of that foundation as a research tool. Not a forecast.

The friability problem at NOBEX would have been caught by a stability-aware test plan, not by a model. It still would. The same is true of most of the issues showing up on your floor today.

If you remember one thing from this post, make it the sequence: clean data first, dashboards next, AI last. Skip steps and you are buying a backhoe to plant a flower.

---

About the author: Richard Soltero, Ph.D. is the founder and president of InstantGMP, Inc., a global manufacturing platform serving pharmaceutical, dietary supplement, food and beverage, automotive, cosmetics, medical device, chemical, and contract manufacturing operations. He has spent more than 30 years in pharmaceutical development, GMP manufacturing, and laboratory information systems. Learn more about iQMS.ai and the InstantGMP platform.