Factory AI Stays in Pilot While New Drug Plants Rise

GlobalData says pharmaceutical manufacturing AI will pay first by squeezing plants drugmakers already run. Digital twins, predictive maintenance, and live quality checks are the tools named in that June 11, 2026 note. Most of them are still in pilots.

The same companies are still funding new oral-dose lines that will not finish until 2028, because a live GMP line is a poor place to test a model a quality unit cannot explain to an inspector.

GlobalData Points at Plants That Already Exist

Edita Hamzic, a healthcare analyst at GlobalData, wrote the factory case into a monthly Bio/Pharmaceutical Outsourcing note rather than a discovery white paper. AI is already a common lab tool. The new argument is that the next gain sits on the shop floor, in downtime, waste, and batch consistency, and that the first job is to improve existing facilities without new infrastructure.

That line is easy to applaud in a slide deck and hard to run on a filling line that cannot stop. A twin that models a bottleneck before anyone touches steel is cheap next to a greenfield plant. It is also a change to the validated state of a process, which means the quality system, not the data-science team, owns the go-live.

Hamzic’s own test is blunt. Firms that treat AI as part of how the plant already works, rather than as a side project with a demo date, are the ones she expects to see a return. The rest can buy software and still ship the same number of batches.

While many pharmaceutical companies are investing in AI, implementation remains the biggest challenge. Many companies face problems with outdated systems, uneven data quality, and difficulties in moving from pilot projects to routine use in highly regulated environments.

Edita Hamzic, Healthcare Analyst, GlobalData, June 11, 2026 note

Success, she added, depends on combining manufacturing skill with digital plumbing in day-to-day operations. That is a plant problem, not a model problem.

Why a Twin Rarely Leaves the Pilot Line

A digital twin in this setting is a live model of equipment and process, updated with plant data, used to test a change before the physical line moves. A static 3D drawing of a room does not qualify. Predictive maintenance watches sensors for failure. Real-time quality monitoring flags a batch that is drifting while it is still in the tank. All three show up in vendor pitches. All three stall at the same three points Hamzic listed.

THE THREE BLOCKS ON A LIVE LINE

• Old plant systems: Many shops still run mixed MES, SCADA, and paper records that were never built to feed a model every minute.
• Uneven data: Batch records that are complete enough for release are often too sparse, too late, or too inconsistent to train a twin that quality will trust.
• The GMP jump: A pilot can sit beside the line. Routine use means change control, validation, and a human who can override the model when it is wrong.

Quality people who work inspections keep making a simpler point than the vendor decks. The European rule set that will govern these tools is still being rewritten, and a lot of the people buying twins have not read it. Until a qualified person will sign a model change the way they sign a batch, the software stays in the bay next to the line it is supposed to run.

Sanofi has already used process twins in a narrower way that shows why the prize is real. When a partner site hit trouble on a first set of batches, a scientist fed the partner’s data into a digital model, found the cause, and tested settings without repeating the failure on the floor. Physical experiments could have taken months. The company said digital methods produced a fix in just a few days. That was troubleshooting, with people still in the loop. It was not an unsupervised model releasing commercial product.

Lilly Used Models to Make More Product

Eli Lilly is the exception GlobalData is describing, not the industry mean. Diogo Rau, Lilly’s chief information and digital officer, said in early 2026, “We literally made more product last year than we possibly could have without AI,” and that the extra volume “would’ve been material in our earnings reports.” Scot Lindsey, senior vice president and information officer for manufacturing and quality, has been putting twins on actual lines to simulate scale-up, find bottlenecks, and tune settings before anyone changes the physical process.

In one bottleneck, a twin predicted output under different settings, the change was tested in a pilot, then it went into manufacturing. Lindsey called the time cut in that step tremendous. He also described a shift from one-off data-science tools to a predictive engine that operations, engineering, and manufacturing could use in near real time. The rules he named were human-in-the-loop, explainability, and transparency.

We don’t want any black box. We want to fully understand why our AI solution is recommending what it’s recommending.

Scot Lindsey, Senior Vice President and Information Officer for Manufacturing and Quality, Eli Lilly

That sentence is why most twins stay in pilot. A filling line making a GLP-1 dose cannot take a recommendation that the quality unit cannot reconstruct. In October 2025 Lilly also partnered with NVIDIA on a company-run supercomputer that, among other jobs, is meant to support manufacturing twins and robotics. Discovery still gets the louder budget fight, and AI drug discovery still waiting on Phase II has soaked up years of that spend. Lilly’s manufacturing comments are the rare case of a shop-floor model being credited with extra saleable product.

Lindsey’s closer is the same as Hamzic’s, in plainer words. “AI, in and of itself, isn’t the solution,” he said. “It’s an expert organization that embraces the technology and works with it to make them even better at what they do.”

A €432 Million Bet on New Oral Capacity

If extra output from existing plants were easy to book, fewer companies would be buying steel for 2028. On March 2, 2026, Novo Nordisk announced a €432 million investment (the company put that at about DKK 3.2 billion) to upgrade and retrofit its tabletting site in Monksland, Athlone, Ireland. The 45-acre project is meant to add oral GLP-1 capacity for markets outside the United States. Construction had already started. Novo said it will finish in phases from the end of 2027 through 2028.

Kasper Bødker Mejlvang, executive vice president for CMC and product supply, said the Athlone spend expands oral production so the company can meet current and future demand outside the U.S. The site’s 260 employees stay on oral products. Novo said the build can create up to 500 construction jobs. That is a capacity answer, with a three-year clock, to the same demand that digital twins are supposed to meet on lines that already exist.

WHERE FACTORY AI ACTUALLY SITS

The table is the irony in one view. The agency that inspects plants is already using models to choose its targets. The companies that need the extra batches are, with a few exceptions, still running twins as projects. New oral plants remain the spend they can defend to a board, because a building has a completion date and a validated process does not have to argue with a neural net.

An AI Shortlist for One-Day Plant Visits

On May 6, 2026, the U.S. Food and Drug Administration said it is piloting one-day inspectional assessments that started in April. The screens are meant to complement standard inspections, not replace them, and they do not apply to higher-risk or more complex sites. Commissioner Marty Makary said the short visits let the agency cover more ground without easing the rules, and that they cut disruption for lower-risk establishments.

As of late April 2026 the agency had completed approximately 46 one-day assessments. Most confirmed compliance and were classed as No Action Indicated. Some ran longer when investigators found significant issues. They keep the right to widen the visit. The pilot is scheduled to continue through fiscal year 2026, with metrics on duration, escalation, and whether the findings help later risk scoring.

Makary also told a Food and Drug Law Institute conference that the one-day visits are screening inspections at lower-risk facilities “that our AI is identifying as low risk,” and that the point is to do more inspections. The written FDA release does not name the model. The targeting claim sits on the commissioner’s own words. Associate Commissioner Elizabeth Miller said the agency is studying outcomes, risk signals, and investigator feedback before it decides how far to push the approach.

THE 2026 CALENDAR AROUND FACTORY AI

1. March 2, 2026: Novo Nordisk announces the €432 million Athlone oral-dose retrofit, with phased completion from the end of 2027 through 2028.
2. April 2026: FDA begins one-day inspectional assessments at selected lower-risk facilities.
3. May 6, 2026: FDA publishes the pilot, reporting approximately 46 assessments by late April, most of them No Action Indicated.
4. June 11, 2026: GlobalData publishes the manufacturing note that puts extra output from existing plants at the center of the AI case.
5. June 30 to July 1, 2026: EMA holds its Annex 22 workshop on AI in medicines manufacturing, after about 1,300 comments on the draft.

Inspectors using a model to pick a one-day visit is not the same as a plant using a model to release a batch. It does show that the regulator is willing to put AI into its own workflow while it still asks manufacturers for human oversight on theirs.

Annex 22 Keeps Humans on Critical Steps

In Europe the gate has a number. EU GMP Annex 22 is the planned annex on artificial intelligence in medicines manufacturing. A 2025 consultation on the draft drew about 1,300 comments. The draft had said dynamic, adaptive, and probabilistic models, including generative AI and large language models, should not be used in critical GMP applications. After industry pushback, EMA’s GMP inspectors’ group held a two-day Annex 22 workshop on manufacturing AI on June 30 and July 1, 2026, to gather evidence on guardrails rather than a flat ban.

The first day was open. The second was closed, for the drafting group. EMA has said it is still weighing the consultation. A February 2026 network meeting on AI put a final Annex 22 at the end of 2026. Until that text lands, a quality unit has to defend shop-floor AI under older GMP rules that never named a neural net.

WHAT WE KNOW

• The draft fence: The consulted text told firms not to put generative or other adaptive models on critical GMP steps.
• The comment pile: About 1,300 responses came in, with support in the consultation for allowing generative tools if controls exist.
• The workshop: EMA asked experts for mitigation measures such as guardrails, logging, and human override.

WHAT IS UNCONFIRMED

• The final rule: Whether Annex 22 will keep a hard bar on generative models in critical use, or allow them with named controls, is not published.
• The date: End-of-2026 publication was the working expectation in February 2026, not a statute.
• U.S. alignment: FDA’s inspection pilot does not rewrite 21 CFR 211, and it is not a green light for unsupervised release on a U.S. line.

Hamzic’s June note already framed the European position as useful across the medicine lifecycle only if the tools stay transparent and human-centered. That is the same human-in-the-loop test Lindsey described at Lilly. It is also why a twin that works in a sandbox still needs a year of validation theater before it can change a setpoint on a commercial peptide line.

The Output Chase Starts After Approval

The reason plants are being asked for more batches, while models sit in pilot, is not only obesity-drug demand. A separate GlobalData survey, fielded from 28 September to 8 November 2025 for the State of the Biopharmaceutical Industry report, found that delays in turning approved medicines into revenue through pricing and reimbursement already sit beside AI and trial cost as a board-level headache. Respondents scored that delay at 3.7 on a 1 to 5 scale, only 0.1 to 0.2 below the four leading trends in the same survey.

On a different question about regulatory and macroeconomic hits over the next 12 months, pricing and reimbursement constraints ranked third, cited by 22% of respondents. Actions of the Trump administration and trade wars and tariffs each drew 36%. Milena Izmirlieva, GlobalData’s senior director for health economics and market access, noted that the top three negative items all hit price, and that four of the next five were also price tools, including U.S. Inflation Reduction Act negotiation and most-favored-nation policy.

SURVEY FLAGS ON THE NEXT YEAR

• Revenue delay score: Pricing and reimbursement delay after approval scored 3.7 on a 1 to 5 impact scale.
• Third-ranked drag: Pricing and reimbursement constraints were cited by 22% of respondents as a top negative factor.
• Tied first: Trump administration actions and trade wars and tariffs were each cited by 36%.

When launch revenue slips because a payer has not finished its process, the plant becomes the only lever that still answers to the company. That is the setting in which a twin that could free a few percent of line time looks valuable, and in which nobody wants to be the person who changed a validated process with a model the inspector has never seen. New tablet capacity in Athlone has a budget line and a 2028 date. A shop-floor model has a comment period.

Hamzic’s June close still stands as the operating rule, not a slogan. AI is being used to strengthen established manufacturing practice, she said, not to replace it. Until Annex 22 is final and a quality unit will sign the change control, extra obesity and diabetes doses will keep arriving on construction schedules, and the twins will keep running beside the line instead of on it.