Usage of Artificial Intelligence in Pharmaceutical Manufacturing

With the rapid growth of AI, the race is no longer about who has the "best and strongest AI" but rather about who can harness its features to create actual meaningful and long-lasting benefits. In the pharmaceutical industry, a considerable amount of attention has been given to the potential of AI in regulatory affairs and clinical trials. While I agree that these are valuable areas where AI could make a difference, I believe that AI’s most impactful role in the future of pharmaceuticals will be in the manufacturing sphere.

Pharmaceutical manufacturing, an industry rooted in well-established practices and traditional methods, tends to shy away from changing its ways. The usage of AI is no different in this regard. It represents a departure from the traditional model of human interaction and intellect. In my experience, this hesitancy stems from a lack of understanding of its capabilities and its general usage rather than a fear of a new system. At its core, AI models and systems are tools that can reinforce and fortify the existing human-centric practices.

The FDA Form 483 is a report issued to firm management at the conclusion of an inspection when the investigators observe violations or mistakes in conditions or practices that could potentially adulterate the product [1]. In 2024, the FDA issued 561 483 observations to drug companies and 4056 to all others under the FDA umbrella [2]. Since they largely come from various examples of “mistakes”, it presents a clear opportunity for AI to shine and prove its use cases.

Below are three of the top reasons that 483s are issued to drug companies:

1. Inadequate Written Procedures

2. Deficiencies in Investigations of Deviations or Failures

3. Data Integrity Issues

Inadequate Written Procedures

"Inadequate written procedures" refers to the failure to establish, implement, or maintain detailed, clear, and accurate written protocols that govern production and process control systems, as required under 21 CFR 211.100(a).

Given the sheer volume of procedures, protocols, and documentation generated in the pharmaceutical environment, it's not surprising that some procedures are improperly followed or forgotten entirely. For those with manufacturing experience, how many times have you heard someone ask, "What does the SOP say about this? Wait, do we even have a procedure for this?"

By leveraging AI, these issues can be significantly reduced, if not entirely eliminated. Imagine an AI assistant integrated into the company’s servers with access to all facility documents. It could provide real-time answers about the state of the documentation and offer further guidance to Quality Assurance (QA) teams.

For example, let’s consider a scenario involving cost analysis. Suppose it's discovered that a specific, expensive filter is being replaced frequently, creating a drain on resources that could be allocated elsewhere. To address this, you would first need to determine whether the equipment and filters were validated and whether the replacement schedule was outlined in the qualifications or SOPs. For some companies, this process would require significant time spent combing through numerous validation documents and SOPs. Multiple SOPs referencing the same filter could create an even more complex, time-consuming scenario.

In contrast, an AI assistant could instantaneously search all documents, identify every instance where the filter is mentioned, and report on its qualification status, replacement schedule, and other relevant details. With this information in hand, the QA team could make an informed determination on the best path forward.

The AI could even be expanded to act as a proactive reviewer. The large language models (LLM) underpinning AI algorithms, such as ChatGPT, are also natural language processors (NLP). NLP abilities enable AI to assess clarity, tone, and structure across multiple documents, ensuring consistency. An AI-driven review process could flag sections where the tone is inconsistent or where the structure diverges from established formats. This ensures that documentation not only meets regulatory standards but also remains readable and easy to follow for staff.

Additionally, AI could play a crucial role in maintaining alignment between related SOPs. During change control processes, for instance, AI could ensure that all procedures impacted by a specific change are properly updated. If a change was made to filters in an engineering SOP, but that same filter was referenced in a small quadrant of a production SOP, the AI tool could identify the inconsistency and flag it during the review process. This would prevent scenarios where outdated or contradictory information remains hidden, reducing regulatory risk.

AI can also be programmed to conduct predictive gap analyses. For example, if regulations evolve, AI could compare the updated standards with the company's current SOPs and suggest areas requiring immediate revision. This proactive approach ensures compliance without the need for exhaustive manual reviews each time regulations are updated.

Finally, AI can enhance the way procedures are written in the first place, by analyzing publicly available information, current guidance, and new research provided by the various pharmaceutical organizations and even the requirements set forth by the FDA and the company. The model could collate all the information and assist teams in drafting clear, concise, and robust documentation. 

While AI cannot prevent bad actors from ignoring established protocols, it can help ensure SOPs are robust, up-to-date, and compliant. If utilized effectively, AI has the potential to eliminate 483 observations related to inadequate written procedures altogether.

Deficiencies in Investigations of Deviations or Failures

"Deficiencies in the investigation of deviations or failures" refer to the failure to thoroughly investigate, document, and address the root cause of deviations, non-conformances, or failures in manufacturing processes, testing, or quality systems, as required under 21 CFR 211.192. Such investigations must include timely identification of the issue, root cause analysis, evaluation of potential product impact, and implementation of corrective and preventive actions (CAPAs).

Investigations in the pharmaceutical sphere are challenging due to their detailed yet exacting nature. Striking the right balance between investigating appropriate causes and rigorously challenging potential root causes is essential. The most common issues these investigations face include failing to identify the root cause, lacking adequate documentation of the investigation process, and neglecting to implement the proper corrective actions to prevent a recurrence of the deviation.

An AI assistant could transform how deviations are investigated by providing standardized templates and immediate feedback, which can significantly expedite the root cause analysis process. For example, consider a scenario where your water system produces action-level microbial results. With an AI assistant trained in investigation techniques and equipped with access to the facility’s documentation, investigators could ask questions to retrieve relevant records. In this case, the AI could present data in a structured manner, enabling investigators to identify trends or threads to explore further. Since the AI retrieves information quickly, investigators save time previously spent on manual searches. Additionally, the AI could verify the presence of necessary attachments and supporting documentation, flag missing evidence, and provide insights based on its training. Over time, the AI could even conduct entire investigations with minimal human input.

Once the investigation is complete, the AI could propose potential corrective actions and help design CAPAs to address and eliminate the deviation effectively. By ensuring that CAPAs are robust and properly implemented, the AI would help prevent recurrence and improve overall compliance.

With AI’s ability to streamline investigations, ensure complete documentation, and propose effective CAPAs, it has the potential to eliminate common deficiencies like inadequate root cause analysis and poor corrective actions.

Data Integrity Issues

"Data integrity issues" refer to failures in ensuring the accuracy, completeness, consistency, and reliability of data throughout its lifecycle, including its generation, recording, processing, retention, retrieval, and reporting, as required under 21 CFR Part 11 and 21 CFR 211.68. Such issues often arise when records are incomplete, manipulated, falsified, or lack proper controls to ensure data authenticity and traceability. These failures can compromise the entire quality assurance framework, making it difficult to ensure that drugs are manufactured safely and according to regulatory standards.

Ensuring confidence in the accuracy and reliability of the documentation generated by pharmaceutical sites is one of the most important functions of Quality Assurance (QA). When individuals in a company leave or are indisposed, the only record of an activity or rationale is the documents that they created and approved. Maintaining accuracy ensures that accurate information is maintained. However, given the vast amount of information produced, efficiently evaluating and analyzing the accuracy and legitimacy of each document can be challenging.

Let’s create a scenario where AI could be used for data integrity. A new piece of equipment was just introduced to the site, and all of the installation and equipment validation was just completed, so the executed protocols and a significant amount of supporting data were submitted to QA for review. Normally, reviewing documents this size could take days, maybe weeks, depending on how many mistakes are found or resources are available. With AI implementation, the AI could review the entire document within minutes and provide the location for every missing entry, illegible signature, mistake, or missing document.

Taken further, AI could even be used to implement a proactive data integrity computer model. Instead of relying solely on periodic reviews and audits that could overwhelm or create larger gaps, the system could run continuous checks on the reliability and accuracy of the data. For example, paper logbooks or other documentation could be scanned using a phone and entered into the model, while computer-based logbooks or documents could be processed directly. The AI could take the scanned document, and both review the document stand-alone and cross-verify it with other systems. Let’s use a water system logbook to illustrate this. An alarm occurs, but the responding tech either knowingly or unknowingly forgets to capture the alarm on the logbook. It may take time before QA is able to identify the missed entry and respond to the situation. The AI, on the other hand, could cross-verify the logbook page with the auto-generated audit report stored on the servers and provide a notification that entries were missed.

AI has the potential to revolutionize data integrity efforts by rapidly reviewing extensive documentation, identifying errors, and ensuring compliance with regulatory standards. This shouldn’t signify that AI should be a replacement for QA. The use of AI will be that of a tool that can help accelerate the timeline, reduce mistakes that would otherwise be missed, and help to maintain confidence in the accuracy of the records.

Challenges

The implementation of AI will undoubtedly be incredibly useful in reducing errors. However, it is important to acknowledge the challenges that come with such advancements. Currently, an AI capable of solving all these issues does not yet exist. While tools like ChatGPT are close in nature, they cannot fully achieve the desired outcomes. Training such an AI would require significant time and resources, including an immense amount of data for pre-training. Additionally, fine-tuning the system to maximize its potential would demand substantial monetary and time investments. Despite these challenges, the benefits far outweigh the obstacles. Training AI would also take an enormous amount of time but the time spent in development will allow a big decrease in time spent on these types of activities in the future to be able to focus on other important tasks.

The successful implementation of AI assistants will need to be tailored to individual companies and designed for “unidirectional” learning and training. A major concern for many organizations is the risk of data breaches and the sharing of sensitive material when AI is directly connected to servers. Additionally, every pharmaceutical site operates differently, necessitating site-specific AI models that address unique needs. Once deployed, these AI assistants must undergo rigorous training to adapt to the specific environment of their respective sites.

When it comes to regulatory and standard practices, such as validation and testing techniques, AI models should leverage open-source examples during training. This is where the concept of “uni-directional” learning becomes critical. A central model could be developed to provide standardized training to various localized AI systems within companies. However, to ensure data security and site confidentiality, these localized systems should not communicate back with the central model. A novel LLM that has recently begun gaining traction is DeepSeek. The strength of DeepSeek lies in its fully open-source model. A site-specific model could be developed to achieve the goals listed above.

Incorporating AI into the pharmaceutical industry has the potential to reshape current processes by enhancing efficiency and reducing errors. This, in turn, can improve regulatory compliance. While challenges such as training requirements and security concerns exist, the long-term benefits far outweigh these obstacles. By embracing AI as a powerful tool, the industry can streamline operations and reduce the occurrence of 483 forms, ultimately improving patient outcomes and paving the way for a more reliable and efficient future in pharmaceuticals.