Fujifilm uses AI to help investigate manufacturing deviations that can hold up medicines
Mike Tomasco described how the contract manufacturer draws on past investigations to help quality teams decide where to look, with people responsible for checking the evidence and closing the case.
FUJIFILM Biotechnologies is using generative AI to help staff investigate manufacturing deviations, a category of problems that can prevent a batch of medicine from being released until the quality work is complete. Mike Tomasco, the company’s chief digital and information officer, said the software helps write more consistent reports, search historical records and propose possible causes for people to investigate.
Tomasco described the use in a discussion led by Phillip Duffy, managing director in DSG Global’s life sciences practice, during The CDMO Forum at HLTH USA 2026 in Las Vegas. CDMOs are contract development and manufacturing organizations. Their customers can include a small biotech developing its first candidate or an established drug company seeking capacity for commercial production. In both cases, the ability to make a batch and the ability to release it are separate operational questions.
The example Tomasco gave was an investigation aid. He described a model helping people decide where to start and what evidence to review. He gave no measured reduction in investigation time, no error rate for its suggestions and no estimate of the resulting change in medicine availability. Those are outcomes a manufacturer would need to measure to show that the approach reliably shortens the wait for a batch.
How records enter the investigation
“You have to do an investigation,” Tomasco said of a departure from the manufacturing process. A quality team has to establish what happened, determine what the event means for the batch and record the conclusion. If the team identifies a cause it can address, the next decisions include what corrective action is appropriate and whether the work has addressed the underlying problem.
One role for a language model is to help staff record the event consistently enough for another investigator to find and compare it. Tomasco said the system can search roughly two decades of historical data, “suggest root causes and put you on an investigative path.” Records of previous cases can give a team useful leads, especially if different shifts or facilities use different terms for similar events. Investigators still need to check whether the earlier cases match the conditions of the batch they are reviewing.
The FDA’s guidance on contract manufacturing quality agreements says the agreement should lay out how deviations are reported, investigated, documented and resolved. That allocation matters for drug sponsors and their manufacturing partners. If a software tool supplies a lead, the agreement still needs to identify the people or organizations responsible for the work and the decisions that follow.
The design also makes it possible to ask a specific performance question. A model could help investigators find an applicable record faster. It could also rank similar-looking events ahead of one with a more relevant cause. Evaluating its use requires a way to inspect the records returned, see whether a person could reproduce the reasoning and check whether a suggested cause was eventually supported. The time it takes to close a case becomes useful alongside measures of investigation quality.
Two different jobs for AI in a factory
Tomasco drew a distinction between the language model used in an investigation and the mathematical models manufacturers have long used to monitor a process. He described staff using equipment data and prior experience to anticipate how production is progressing. The purpose is to help teams see changes early enough to decide whether to intervene. The input and the decision differ from a system that searches written investigations after something has departed from plan.
Those differences affect how a company can test the software. A monitoring tool can be assessed against observed results from production, and the organization can set limits on the recommendations staff are allowed to act on. An investigation assistant needs a different set of checks, including the completeness of its source records and whether it keeps a plausible suggestion distinct from a verified finding. Tomasco’s examples give quality and data leaders two uses they can evaluate on their own merits.
Capacity decisions come before commercial launch
Duffy also asked how biotech companies with small teams can prepare for production, and what rising numbers of potential therapies mean for manufacturing capacity. Tomasco said a development-stage company may know its science well and still need outside expertise for clinical manufacturing and later production. He described planning for demand over years while working with candidates that may never make it through clinical development.
That leaves a sponsor and its manufacturer making decisions on different clocks. A sponsor may want to wait for more evidence before committing to production. A manufacturer needs enough warning to provide the people, equipment and production time needed if the program succeeds. The conversation made clear why partnerships can begin well before commercial demand is certain, and why a reserved slot only helps if the teams can complete the development and quality work in time to use it.
Duffy put a practical version of the timing problem to Tomasco. He described calls from leaders nearing approval who realized they lacked a manufacturing plan. Tomasco said a partner could help with clinical development and manufacturing steps well before the sponsor needed commercial supply. The exchange puts a deadline on decisions that can look distant during discovery. A sponsor that waits until a late-stage outcome is clear may have fewer choices about where and when a product can be made.
There is also a handoff between companies when a process moves to a different site or from development to commercial supply. Tomasco described that work as an area where paper-based processes can make it harder for teams to use what they already know. If a project moves, an incoming team needs the development record, the decisions behind it and a common understanding of what the receiving site will check. Digitizing documents gives both sides a record they can work from. The end result depends on whether the information remains interpretable at the receiving site.
In that context, an AI-assisted investigation addresses one part of the supply problem. It may help a team spend less time looking for old cases and more time determining what happened in the current one. The broader test for a sponsor or manufacturer is whether the tool consistently improves the investigation and release process while leaving an auditable record of how the conclusion was reached.