Pharma AI Stays Stuck in Pilot While Obesity Drug Demand Soars

Pharmaceutical companies are racing to put AI on the factory floor, but most of the work is still stuck in pilot. A report on AI in pharma manufacturing from GlobalData published this month finds that the technology remains an emerging capability in drug manufacturing, with many companies still testing AI through limited programs. That gap is now the binding constraint on supply for obesity and diabetes drugs, where demand is rising faster than factory capacity.

The latest edition of GlobalData’s monthly Bio/Pharmaceutical Outsourcing report frames the moment as a shift from discovery to execution. Edita Hamzic, Healthcare Analyst at GlobalData, told reporters that outdated systems, uneven data quality, and the difficulty of running AI in highly regulated environments are the chief obstacles. The report’s case is straightforward: drugmakers need to extract more output from existing plants, because building new ones cannot keep pace with demand in high-value therapy areas. The risk is that companies run AI as a standalone technology project, isolated from the manufacturing operations it is meant to support. A related Abridge’s deal with Eli Lilly and platform expansion shows how AI-pharma deals are also extending past the lab.

AI in Pharma Manufacturing Is Still Stuck in Pilot

The report, drawn from GlobalData’s monthly Bio/Pharmaceutical Outsourcing series, lands a single conclusion. The technology remains an emerging capability in pharmaceutical manufacturing, with many companies still testing these tools through pilot programs. Hamzic put the obstacle list in plain terms, naming outdated systems, uneven data quality, and the difficulty of running AI in a regulated environment as the binding constraints.

Drugmakers have run pilots across digital twins, predictive maintenance, and real-time quality monitoring. The pilots work. The scale-up does not. Success, the report says, depends on execution and the ability to combine manufacturing expertise with digital infrastructure in day-to-day operations.

The report organizes itself around that distinction. AI in drug discovery has been a well-established tool for years, while AI on the production line is a different kind of problem with different failure modes. The first wave of pilots is teaching the industry what those failure modes look like, and the second wave will need a different approach to clear them.

The Tools Drugmakers Are Testing

Three tool families dominate the pilots. Digital twins let manufacturers model and test production changes in a virtual environment before any change is made on a live line. The point is to surface failure modes and yield trade-offs in software before they show up in a batch of medication.

Predictive maintenance pulls sensor data off pumps, mixers, and bioreactors to flag parts that are about to fail and to schedule interventions before they trigger an unplanned stop. Real-time quality monitoring runs statistical process control on continuous measurements, including temperature, pH, and impurity profiles, so that a batch drift is caught within minutes, well before the end of the run.

Drugmakers are testing these tools to identify and test production changes in real time before implementing them in live environments. The result, when the tools work, is lower downtime, less wasted material, and tighter batch consistency. When the tools stay in pilot, the result is a long list of use cases and a short list of production lines actually using them at scale.

Why the Pressure Has Suddenly Intensified

The capacity pressure is concentrated in two therapy areas: obesity and diabetes. Demand for GLP-1 drugs and adjacent therapies has outrun the supply of manufacturing capacity, and the fastest path to more doses is to get more out of the plants that already exist. That is where AI enters the picture, and that is why the report frames AI as a tool for getting more from existing assets, with greenfield expansion off the table for the near term.

New factories take years to build, validate, and license. AI in production can be deployed in quarters, and the capex is small relative to a new plant. The trade-off is that the AI has to work inside a regulated environment where every output is auditable. That is a slower problem than the demand curve, and it is the constraint the report keeps coming back to.

Hamzic’s framing is that the primary AI opportunity in pharma manufacturing is to improve the performance of existing facilities without the need to build new infrastructure. For drugmakers still running pilots, the question is whether the technology can be carried on a production line. The report’s own finding is that many companies are still in that phase.

What Keeps Pilots From Going Live

The obstacles Hamzic names sit in operations, with the technology itself largely settled. The integration is what fails, with each obstacle a separate fix that rarely lines up with the others. The obstacles are operational, and the report treats them that way.

The list runs to three problems. The third is specific to pharma’s regulated environment, where most pharma AI projects stall.

• Outdated systems. Legacy MES, batch records, and historian databases that were never designed to feed a model.
• Uneven data quality. Sensors and lab instruments that drift, are not calibrated on the same schedule, or are missing entirely on older lines.
• Pilot to production in cGMP. Moving a working pilot to a controlled, validated production line is where the bulk of pharma AI projects stall, because every model output becomes a quality-relevant decision.

Hamzic’s prescription is to see AI as a layer across the operational model. Success, the report says, depends on execution and the ability to combine manufacturing expertise with digital infrastructure in day-to-day operations. The companies that treat AI as a layer across operations are the ones most likely to clear that bar.

The full Hamzic line is in the report. That sentence is the operational model argument in one line.

Two Regulators, Two Different Bets

The regulatory frame is hardening around AI in pharma manufacturing, but the FDA and the EMA are taking different paths. The FDA is already using AI to determine where inspections are carried out under its one-day inspection pilot. The agency is exploring the use of AI to identify lower-risk sites so that inspectors can focus on facilities where compliance concerns are most likely to arise, and the criteria behind that risk model are opaque.

The EMA has taken a different route, focused on safeguards around AI use. The European agency sees AI as a useful tool across the whole medicine lifecycle, but only if it is used in a transparent and human-centered way. The two approaches are not contradictory, but they pull in different directions. The FDA is using AI to triage compliance, while the EMA is using regulation to constrain how AI gets used at all.

For drugmakers running pilots, the practical question is which regulator’s frame will govern the production line. The FDA’s overall federal AI guidance for drug development and the EMA’s EU regulator’s AI guidance for medicines are the primary documents a global supply chain has to clear. A pilot that satisfies one may not satisfy the other, and the company has to be ready to clear both.

Rather than replacing established manufacturing practices, AI is being harnessed to strengthen them. AI is therefore becoming increasingly important in drug manufacturing as the sector moves towards systems that link production, quality, and regulation more closely than before.

That is Hamzic’s reading, in the report’s closing line.

The Pricing Squeeze Beneath the AI Push

The manufacturing pressure sits on top of a separate commercial squeeze. A separate survey on P&R delays in pharma conducted from 28 September to 8 November 2025 found pricing and reimbursement delays ranking as the third most negative factor facing the pharmaceutical sector. P&R constraints were cited by 22% of respondents, behind only Trump administration actions and trade wars at 36% each.

Milena Izmirlieva, senior director and head of health economics and market access research and analysis at GlobalData, framed the pricing pressure as a structural shift. Behind P&R in the ranking, four of the five trends identified by the respondents are also drug-pricing related: inflation, the Inflation Reduction Act, the Most Favored Nation policy, and International Referencing Pricing. The US MFN policy effectively implements IRP in the US, where IRP is used in more than 75 countries.

The link to manufacturing is direct. If pricing pressure forces drugmakers to defend margins by squeezing capex, AI in production is one of the first things to get cut or delayed, since it is easier to defer a software project than to idle a plant. The two pressures run in opposite directions, and the companies that survive the squeeze are the ones that can fund the AI deployment while the unit price is falling.

• P&R delays impact: 3.7 out of 5, just behind AI, immuno-oncology, clinical trial complexity, and personalized medicine
• International Referencing Pricing: used in more than 75 countries
• Survey dates: 28 September to 8 November 2025
• P&R constraints: 22% of pharma respondents; Trump actions and trade wars: 36% each

What Has to Change for Pilots to Become Production

Hamzic’s prescription is direct. “Companies that see AI as part of their operational model, not as a standalone technology project, are most likely to benefit,” the report concludes. That sentence is the report’s organizing claim, and the rest of the report is an argument for it.

The harder work in AI in pharma manufacturing is integrating the technology into operations, with integration decided by operations leadership. The companies that close the gap will look different from the ones that do not. They will treat pilot results as production requirements, and staff the validation and quality teams for AI work. Most of the industry, the report suggests, is still in the early stages of that work, with the companies that are not setting the pace for the rest.

Frequently Asked Questions

What does AI in pharmaceutical manufacturing actually mean?

It means using AI on the factory floor, not in the lab. Drugmakers are testing AI for production planning, real-time quality monitoring, predictive maintenance, and digital twins that model plant behavior in software. The goal is to get more output from existing facilities, since building new plants takes years.

Why is AI in pharma still stuck in pilot?

Most of the technology works in a pilot. The hard part is moving it to a regulated production line. GlobalData’s Edita Hamzic points to outdated systems, uneven data quality, and the difficulty of running AI inside cGMP as the main reasons pilots do not become production.

What specific AI tools are drugmakers testing?

Three families dominate: digital twins, which model and test production changes in software; predictive maintenance, which uses sensor data to flag equipment at risk of failure; and real-time quality monitoring, which runs continuous statistical process control on temperature, pH, and impurity. Together they target downtime, waste, and batch consistency.

How are the FDA and EMA approaching AI in pharma manufacturing?

The FDA is using AI to triage inspections under its one-day inspection pilot, identifying lower-risk sites so inspectors can focus elsewhere. The EMA is focused on safeguards around AI use, requiring transparency and human oversight. A global supply chain has to clear both frames.

How big is the demand pressure driving AI investment?

GlobalData frames the pressure in obesity and diabetes therapies, where demand for GLP-1 drugs and adjacent treatments has outrun manufacturing capacity. The fastest path to more doses is more output per plant, which is what AI in production is meant to deliver.

What is the role of pricing and reimbursement in slowing AI scale-up?

A separate GlobalData survey from late 2025 found pricing and reimbursement delays ranked as the third most negative factor facing pharma, cited by 22% of respondents, behind only Trump administration actions and trade wars at 36% each. Squeezed margins can delay the capex a controlled AI deployment requires.