Content Benefit

Content Strategy | UX Writing | Content Design | Globalization | Communications

The factory floor won’t retire

The factory floor won’t retire

Generative AI in pharmaceutical manufacturing at Angelini Pharma

To me, the arrival of generative AI on the factory floor feels like a second industrial revolution: manual processes get automated and lead times shrink. Yet assembly lines never sent skilled workers into retirement, and this time too the people who know the machines are staying put, with an even more central role. A good Italian example was covered by Agenda Digitale and presented at the AI & Automation Talk hosted by the Digital Innovation Observatories of Politecnico di Milano: Angelini Pharma, working with Class Reply, has turned decades of notes written by line operators into real-time guidance for getting a stalled plant running again.

Thirty years of notes on machine stops

Angelini Pharma is a family-owned group founded in 1919, with more than 1,500 employees in Italy and a presence across Europe, Asia and the Americas. Its lines produce, among other things, Tachipirina, the acetaminophen brand that sits in nearly every Italian medicine cabinet, along with several epilepsy treatments.

For twenty or thirty years, after every plant stoppage, staff logged the cause of the fault and the fix they applied. The result was a huge body of knowledge, but written as free text, with no categories and no structure at all. Raffaele Cardi, the company’s Lean Project Manager, framed the starting point as a question: how could the operators’ established know-how be put to work to raise production efficiency? The answer was to hand generative AI the cognitive troubleshooting that every technician runs through in their head while the line is down.

How the Cognitive Layer works on Microsoft Azure

The project began as a Minimum Viable Product on a single production line. Class Reply, the Reply Group company specializing in Microsoft and Databricks technologies, then took it to enterprise level by adding a Cognitive Layer to the existing platform: an application layer that classifies anomalies and suggests the remedy. Marco Trampaglia, AI Specialist at Class Reply, walked through the architecture:

  • Sources: real-time alarms read directly from the machines, alarm history exported to Excel, and fault reports entered by hand;
  • Storage: Azure Cosmos DB for alarms and faults, SharePoint for the knowledge base;
  • Processing: Azure Databricks works with Azure OpenAI inside Microsoft Foundry, using GPT models and text embedding models;
  • Interfaces: an app for operators on touch screens and iPads, Power BI reports, and an admin panel for whoever oversees the content.

Angelini set one explicit requirement: its own people must be able to manage and validate the knowledge base without relying on the vendor. On a new line, the base is seeded with the alarm history; on lines already running, it grows every day with the operators’ confirmations. Either way, before anything is saved for good, an internal reviewer checks what the GenAI has produced. The final word stays with a person.

Two taps instead of minutes at the keyboard

For the line staff, the biggest shift has been a reversal of routine. Before, they fixed the fault and only once the stop was over typed up a comment after the fact. Now, when an alarm goes off, the dashboard shows the machine, product, format and length of the stop, while the app suggests grouped categories, each with its own confidence percentage. The operator confirms the diagnosis and the fix with two taps on the screen, and can still write a free-text report when the case is not covered.

The hardware choice came from the shop floor too. The project team asked operators what would make their day easier, and the answer was touch screens and iPads along the lines: information now goes in within 2 or 3 seconds, where it used to take several minutes at a keyboard. As a UX writer, this is the part I like best: the technology was adopted because it starts from the real gesture of the people using it, and steady training did the rest.

The results: 3% gains in batch efficiency

Two figures have been shared. Work-cycle efficiency rose by 3%, and the time needed to complete a 1,000-unit batch and release it to market dropped by 3% as well. That may sound modest, but on pharmaceutical lines that run all year, every percentage point means boxes reaching pharmacies sooner and fewer hours of idle equipment.

Why operators remain the heart of the system

The detail that best supports my view came from Cardi, answering a question from the audience. Angelini runs machines from different manufacturers, including IMA and Partena, and very distant technology generations work side by side: a recent IMA machine provides detailed, structured PLC and sensor data, while one from twenty years ago only raises generic alarms. Precisely where the sensors were poorest, the notes operators had piled up over time turned out to be the most valuable source for training the models.

In other words, the AI learned from the people who have spent years getting the lines moving again. Without that experience, written down day after day, the system would have had very little to suggest. The same group has also just wrapped up Smart Pharma, a project backed by the Italian Ministry of Enterprises and Made in Italy that tested predictive maintenance and augmented reality smart glasses at its Ancona, Casella and Aprilia sites, to help staff with format changeovers and training. There too, the worker stays at the center and technology works alongside.

The second industrial revolution, AI edition

Between the late nineteenth and early twentieth centuries, electricity, steel and the assembly line broke work down into repetitive operations, but they also created new roles: maintenance crews, toolmakers, quality technicians. Today’s forecasts point to a similar pattern. According to the World Economic Forum’s Future of Jobs Report 2025, by 2030 about 170 million jobs will be created and 92 million will disappear, for a net gain of 78 million, while 39% of current skills will be transformed or become outdated.

Italy is feeling the push as well. The Artificial Intelligence Observatory at Politecnico di Milano estimated the Italian AI market at €1.8 billion in 2025, up 50%, with 71% of large companies running at least one project compared with just 8% of small and medium-sized businesses. On average, 47% of workers already use AI tools on the job. The gap between large and small firms remains the real sticking point, and cases like Angelini’s show that what makes the difference is the quality of data gathered over the years, far more than the size of the budget.

I touched on this when writing about the skills mismatch and workers over 50: the real challenge AI sets us is staying in the saddle. And it is no coincidence that Gen Z is rediscovering skilled trades, as I discuss in High school or a wrench?. For a broader picture, there is also AI and the labor market.

The rules: from clean rooms to Annex 22

So far the application runs on the packaging lines. Once it reaches the so-called white areas, the controlled-contamination environments where the drug itself is made, the rules become stricter, and Angelini will start the formal validation required by pharmaceutical quality standards, with reference to ISO 9001 and the GAMP and Annex 11 guidelines.

Meanwhile, the European Union is drafting Annex 22 of its GMP guidelines, the first framework dedicated to AI in medicine manufacturing. The draft allows only static models in critical applications, rules out models that keep learning during use, and keeps GenAI and LLMs away from any critical decision. It does leave room for supporting tasks, such as summarizing reports or searching procedures, as long as a qualified person reviews the output and takes responsibility for it. Lawmakers, in short, also want a human being next to the machine.

Faster and, hopefully, better

The Angelini Pharma case confirms my belief that generative AI on the factory floor works when it builds on the people already there: thirty years of notes have become an assistant that answers in seconds, but the operators wrote them, and they are still the ones confirming every diagnosis. Like the assembly lines of a century ago, this revolution will take away repetitive gestures and leave more room for experience. Everything will get done faster and, hopefully, better.

Related sources

Chiara Bonifazi