AI in Biologics Manufacturing 2026: Plasma and Cell Therapy Compared
Plasma fractionation and cell therapy both get called biologics manufacturing, and in a plant they have almost nothing in common. Plasma has been processed at industrial scale since the 1940s. You put litres in, you get grams of immunoglobulin out, and the certificate says how pure it is. Cell therapy has no equivalent number. Nobody has agreed how to measure whether a batch of cells will do anything. That gap decides what AI is worth in each one. This report puts the two side by side using what the FDA has actually published and our own count of the clinical-trial registry.
Why the two processes are not comparable
Plasma fractionation separates donated plasma into albumin, immunoglobulins and clotting factors through precipitation and chromatography. Every stage has a yield. Every batch gets an assay. The finished product is defined by concentration and purity, and you can tell whether a run went well before it ships. It is an ordinary industrial process, and the market size reflects that maturity: roughly USD 40.3 billion in 2025, growing near 9% a year. Cell therapy works differently. The product is living cells, and what they are believed to do comes from what they secrete rather than from what they become. No validated assay predicts whether a given batch will help a patient. That is not our reading of the literature — it is what the FDA says. Its August 2026 draft guidance on potency for active immunotherapy products states that potency is hard to assess because it depends on the host immune response, and recommends quantitative bioassays on living cells as the best method currently available.
What the trial registry shows
We counted every study registered on ClinicalTrials.gov for nine cell types on 4 September 2026. Of 2,721 studies, 766 have no trial phase assigned at all. A study without a phase is outside the route that leads to an approved medicine, so over a quarter of registered activity in this field is not heading for approval. The split by cell type is sharper. Mesenchymal stromal cells have 1,837 registered studies and 75 of them reach Phase 3 — one in twenty-four. Exosomes have 573 studies and five reach Phase 3. Chondrocytes have 90 studies with 21 at Phase 3, and they are the only cell type in the set with an authorised medicine in the EU. The reason is not mysterious. Chondrocytes build cartilage, which is a structure you can specify and test against. Where the product is a secretion profile instead, there is nothing firm to write on a release specification, and the phase distribution shows it.
Two thirds of exosome studies have no trial phase
Of the 573 registered exosome studies, 378 carry no phase. That is 66%. Most of that work sits in aesthetics and wellness, outside the regulatory route entirely. This matters if you are deciding where to put process AI. A model can hold a process to its release specification, but only the specification decides whether that means anything. If the spec is a particle count rather than a measure of what the product does, holding it precisely achieves nothing useful.
What the FDA asks of an AI model
| Context of use | Example | Credibility burden |
|---|---|---|
| Utilities and environment | Clean-steam and HVAC load prediction | Low — no product impact |
| Equipment reliability | Centrifuge and chromatography-skid monitoring | Low to moderate — availability, not quality |
| Process control | Chromatography step yield optimisation | Moderate — affects yield and purity |
| Product quality decision | Predicting batch potency or release | High — the model is part of the specification |
In January 2025 the FDA published draft guidance called Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products. It does not approve or forbid particular techniques. It asks you to state the context of use — which question the model answers and which decision depends on the answer — and then sets the evidence you need in proportion to that. In practice this means the same algorithm carries very different obligations depending on where you point it. A model scheduling a chiller needs very little. A model that influences batch release becomes part of your specification and needs a great deal. Teams that roll one model across both without separating the two end up with a regulatory problem they did not plan for.
Where AI actually pays in a biologics plant
Leave the product-quality claims aside and what remains is a list any brewery or chemical site would recognise. Clean steam, water-for-injection and HVAC are large continuous loads under strict validation. Cold chain runs without pause and fails expensively. Autoclaves, centrifuges and chromatography skids are rotating and pressurised equipment with ordinary failure modes. Environmental monitoring produces far more data than anyone reads. None of this touches potency, which is why it can be deployed now. No product specification depends on the output, so the regulatory burden stays low. Starting here also builds the data discipline that a quality-affecting model would demand later anyway. Surface temperature belongs on the same list. Clean-steam distribution, autoclave bodies and sterilisation lines run hot in rooms where people work all shift, and uninsulated valves and flanges cost energy and burn skin.
What to do with this if you are planning a project
In plasma fractionation, AI improves a process that is already measured. The gains are real and incremental: yield, energy, uptime. In cell therapy the measurement problem comes first. Without a validated potency assay there is no stable target for a model to learn, so a model aimed at predicting batch quality has nothing reliable to aim at. Sequence the work accordingly: get the measurement right, then automate it. Judge any vendor claim by whether it respects that order.
FAQ
Can AI predict whether a cell-therapy batch will work?
Not reliably, and the limit is not computing power. Most cell products have no validated potency assay that predicts clinical effect, so there is no dependable label for a model to learn from. The FDA's August 2026 draft guidance on potency says assessment is difficult because it depends on the host immune response. A model is only as good as the measurement behind it.
Is AI in pharmaceutical manufacturing regulated?
The technique is not, the use is. The FDA's January 2025 draft guidance asks you to define the context of use — what the model answers and what decision rests on it — and scales the required evidence to that. Predicting HVAC load and predicting batch potency sit at opposite ends of that scale.
Where does AI pay off fastest in a biologics plant?
Utilities and equipment reliability: clean steam, water-for-injection, HVAC, cold chain and rotating equipment. These are large continuous loads that are already measured, no product specification depends on them, and savings show up within a year.
Why is plasma fractionation easier to optimise than cell therapy?
Its output is specified and measured. Plasma processing yields defined proteins at defined concentrations and purities, so every step has a target to train against. Cell-therapy products are living cells whose effect is attributed to secretion, and the field has not settled on how to measure that.
How many cell-therapy studies reach late-phase trials?
Of 2,721 studies registered for nine cell types on ClinicalTrials.gov as of 4 September 2026, 766 have no phase assigned. Mesenchymal stromal cells reach Phase 3 in about one study in twenty-four, exosomes in fewer than one in a hundred. Chondrocytes, the only cell type here with an authorised EU medicine, reach Phase 3 in 23% of theirs.
Sources
- U.S. Food and Drug Administration — Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products (draft, January 2025)
- Regulatory Affairs Professionals Society — FDA drafts guidance on assessing potency of immunotherapy products (August 2026)
- Fortune Business Insights — Plasma Fractionation Market Size, Share & Industry Report
- StemCellAtlas — Cell-type registry profiles — reproducible ClinicalTrials.gov extraction, 4 September 2026
- ClinicalTrials.gov — API v2 — study records used for the counts in this report
Related
Predictive Maintenance: A Practical Guide for Plants · Industrial Heat Loss and Insulation: How Much You're Losing · Condition-Based Maintenance (CBM) · Digital Twin
Charts: How widely manufacturers have adopted AI
Sectors: Pharmaceuticals · Chemicals · Food Processing