A deviation gets flagged on a 2,000-liter batch. The product sits on hold until someone can explain what happened and whether it is still sound, and the batch is worth several million dollars. Three engineers are pulled onto the investigation. Over the next two weeks, most of their hours will not go to reasoning about the science. They will go to finding things.
The chromatography data lives in one system. Environmental monitoring lives in another. The batch record is part electronic and part paper. The handful of deviations that looked like this one over the past two years sit in a quality system that talks to neither, and in the memory of an engineer who has since changed teams. Once the data is in front of them, the reasoning takes an afternoon. Assembling it takes the rest of the two weeks.
This is precisely the work AI is supposed to take off their hands. Reading across thousands of pages, correlating weak signals, drafting the first pass of an investigation report: machines are good at this now. And yet most AI sold into pharma manufacturing stalls before it ever touches a batch on hold. The reason is rarely the model. It is that the model has nothing to work from. An agent cannot reason over data it cannot reach, and in most plants that data was never gathered in one place to begin with.
The second wall is higher than the first
Suppose you clear the first wall and get the data into one place. A harder one waits behind it. In regulated work, a fast answer is worth nothing if it cannot be trusted and traced. An investigation has to point to the exact records its conclusion rests on. It has to hold up when an auditor reads it a year later. A model that writes a confident summary it cannot source has not helped the quality unit; it has handed them a new liability. This is why a chat window laid over the existing tools rarely survives a pilot. It can describe the work. It cannot stand behind it.
The teams getting value fixed the foundation first
The pharma teams seeing real results from AI on the floor have usually done the unglamorous thing first. They fixed what sits underneath it. This is the premise Katalyze AI was built on. The company, which recently raised a $10.5 million seed round led by Bonfire Ventures, makes an agentic operating system for pharma. Before it runs a single agent, it pulls the data scattered across the MES, the LIMS, the historian, and the quality system into one operational record, with every fact anchored to its original source.
The agents that run on that record are trained on the work itself, investigating deviations, tracking CAPAs, drafting APQRs, in the language a process engineer or a reviewer actually uses. Their skills are written with a community of more than 100 tenured scientists and engineers from companies including Pfizer, Sanofi, and Lilly, the people who have spent careers doing this work themselves. Because the record and its traceability are built into the architecture, the agents operate inside GxP and a company’s data rules rather than around them. Every answer arrives with its evidence attached.
The effect on that Friday-afternoon investigation is the point. In one early deployment, an analysis that would have taken a year and $4 to $6 million was finished in 45 minutes. Five of the 20 largest pharmaceutical companies now run on the platform, and it has helped their teams get 10 million doses to patients.
None of this is speed for its own sake. A batch released two weeks sooner is a therapy that reaches a patient two weeks sooner. A deviation closed correctly the first time is a recall that never happens. The work between a molecule and the patient waiting for it has always carried that weight. What is new is that the teams who put AI to work on it, carefully and on a foundation they can defend, are the ones who will feel it first.
Learn more at katalyzeai.com.
About Katalyze AI. Katalyze AI is the agentic operating system for pharma, from molecule to patient. It unifies fragmented production data into one source-grounded record and runs domain-trained agents across the regulated span from process development to release. Based in San Francisco, Katalyze is backed by Bonfire Ventures, with iNovia Capital, Ripple Ventures, Alumni Ventures, and leading angel investors. Five of the 20 largest pharmaceutical companies build on it to get medicine, faster and more reliably, to the patient.
The editorial staff had no role in this post's creation.