From Central Plants to Smart Hubs: How AI and Inline Analytics Will Decentralize CGT Manufacturing
Cell therapy cannot scale to thousands of patients with paper, manual sampling, and siloed platforms. This session explored a practical path: digitize the process, measure what matters in real time, and use AI to predict outcomes and guide release. The goal is not fewer people, but better, faster decisions that make decentralized networks safe and affordable.
Captured live at Advanced Therapies Europe 2025, this session distills how digital batch records, inline analytics, and pragmatic AI can shift CGT from centralized plants to smart regional hubs without losing quality.
TL;DR: In a hurry? Here are the essentials at a glance:
Decentralization is coming fast, but most teams see a hybrid model first: decentralized manufacturing with centralized analytics and oversight.
Inline telemetry on 14 process parameters turns black-box cultures into controllable systems, enabling earlier interventions and fewer failed batches.
AI and machine learning already predict yield from day 2 signals like dissolved oxygen slope, lactate trends, and pump-rate recovery, helping optimize activation and feeding.
Digital EBRs and automated data capture unlock process characterization in months, not years, by surfacing true CPPs and CQAs across runs and sites.
QPs do not get replaced. Digital tools reduce paperwork, surface risks, and accelerate safe, documented release.
Europe’s access gap is as much pricing and commissioning as science. Decentralized academic manufacturing can help widen availability for rare indications.
Interoperability matters. Platforms that “talk” to each other make tech transfer, filling, and release faster and less error-prone.
Secure ownership
Supply chains and digital programs work best when one person is accountable. Appoint a named owner for digitization and data governance who convenes manufacturing, analytics, QA, IT, and the QP to design for release from day one. Move early to electronic batch records with automated data ingestion from instruments and platforms so information is right the first time and retrievable during characterization and BLA prep.
“Pen and paper is my enemy. Pen and paper is my competitor.” — Alexander Seyf, Autolomous
Make ownership visible. Publish a RACI for data flow from apheresis to disposition. Put QA and the QP at the table as you define rapid-release criteria, escalation paths, and what the regulator will see.
Design the route
Decentralization requires more than an extra cleanroom. The route needs harmonized SOPs, qualified people, and interoperable systems at every node. Several panelists argued for regional GMP hubs adjacent to hospitals or cancer centers rather than every hospital running its own suite. Others highlighted familiar networks like blood banks as potential hosts if quality systems align.
“The challenge is how you get the data.” — Steven Feldman, Stanford
Standardize interfaces. Choose platforms that expose process data and integrate scheduling, inventory, and fill-finish so tech transfer looks like configuration, not a re-write.
Autologous vs allogeneic
Autologous remains unforgiving on time and variability. Shorten culture to 3 days where biology and selection allow, then rely on inline analytics to adjust activation and feeding without opening the system. Allogeneic can batch and stock, but still needs deep understanding of donor-to-donor variability and genuine depot qualification.
“I strongly believe in shorter processes… shorten it to three days… with centralized analytics and decentralized manufacturing.” — Edwin Bremer, Groningen UMC
Use AI to learn what makes a “super donor,” and when exhausted starting material needs a different recipe.
Release with speed and safety
Fresh products need hours-scale decisions that still stand up to scrutiny. Build a two-phase release: rapid, data-rich checks for disposition today, followed by confirmatory assays. Autolomous reported with a national program that digital EBRs and QP assistants saved roughly two thirds of QA effort, reclaimed about 100 days per month for QA teams, and cut documentation errors dramatically.
“At day two you know what is going to come out at day nine or ten.” — Ohad Karnieli, Adva Biotechnology
Design release criteria around real-time signals. Document who can release, what data they need, and how to act if later results disagree.
Navigate EU, UK, and customs
The regulatory question is shifting from “if” to “how” for decentralized models. Europe’s barrier is often access and pricing rather than science; academic manufacture can help fill gaps while commercial models evolve.
“Every center that has FACT accreditation is absolutely capable of doing CAR T therapy.” — Mounzer Agha, University of Pittsburgh
“As far as I know only in Germany all the EMA-approved CARs are actually available… in the rest of Europe people cannot be treated because of pricing.” — Edwin Bremer, Groningen UMC
Engage regulators early with correlated hard data and digital provenance that shows control is consistent across sites.
Involve QP early
QP insight belongs at program start: vendor selection, IMPD planning, inline metrics, and rules for rapid release. AI will not replace QPs; it reduces paperwork, finds gaps, and makes risk more visible.
“Humans will never be replaced by machines when it comes to QPs and certain roles. Treat technology as your colleague.” — Alexander Seyf, Autolomous
Codify when shipment under quarantine is appropriate, what real-time results are required, and how the QP documents decisions.
Standardize methods and orchestration
Clinics and CDMOs slow down when every sponsor turns up with a new portal. Agree a single orchestration workflow, consolidate logins, and use premium couriers with live telemetry. Most importantly, insist that manufacturing, filling, and QC systems interoperate so tech transfer is click-through.
“You are going to need platforms that talk to one another.” — Edwin Beale, Cellipoint
Instrument identity and tracking
Make what matters visible in real time: dissolved oxygen, pH, lactate, glucose, temperature, gas mix, and pump activity. The team showed how DO slope after activation, lactate trend breaks, and pump-rate recovery predict yield with high accuracy, allowing same-donor optimization by adjusting activation and feed timing. Fewer manual samples, fewer surprises.
Treat alerts like process, not exceptions. Rehearse loss-of-signal and excursion playbooks and record actions in your deviation system.
Centralize, decentralize, or hybrid
The consensus path is hybrid. Centralize analytics, method governance, release authority, and data science. Decentralize execution where it removes transport risk and shortens time to dose. Use shared SOPs, reference assays, and a common data layer so the QP sees the same evidence regardless of where product is made.
“Decentralized manufacturing will democratize the process and allow everyone to learn from shared data.” — Mounzer Agha, University of Pittsburgh
Final word: Reduce uncertainty at every handoff: own the data, measure what matters inline, standardize what repeats, and involve QPs early. Do this well and ATMPs move from fragile pilots to reliable patient access.
It involves producing cell therapies closer to patients, often at regional facilities or hospital-affiliated centers, to improve accessibility and reduce costs.
AI analyzes real-time manufacturing data to predict outcomes, optimize process parameters, and reduce batch failures, enabling faster and more consistent production.
Regulators need to accept new manufacturing approaches and digital evidence for product quality, requiring early and ongoing collaboration with industry.
Sharing de-identified process data helps the industry learn collectively, improve processes, and accelerate therapy development.
No, AI is a tool to assist humans, improving efficiency and decision-making, but human oversight remains essential.
)