Mapping AI across the biotech and medtech value chain

Real players: Recursion Pharmaceuticals runs automated cell-imaging experiments and trains models on the images. Isomorphic Labs (an Alphabet company) sells structure-based design. Insilico Medicine has moved AI-designed molecules into human trials.

Where value is real: narrowing millions of possible molecules to hundreds worth synthesizing. This compresses timelines and reduces wet-lab spend.

Where it is theater: claims that AI alone "discovers drugs." No AI-designed drug has yet completed Phase 3 and reached approval as of early 2026. The biology bottleneck remains.

Stage 2: Preclinical and lead optimization

Here AI predicts how a candidate molecule will behave: will it bind the target, will it be toxic, will the body absorb it? This is called ADMET prediction (absorption, distribution, metabolism, excretion, toxicity).

A simple lens on why this matters: if a model can flag toxic candidates early, you avoid expensive animal studies.

Value test: does the model reduce the number of candidates that fail later? If your late-stage failure rate does not drop, the model is decoration.

Stage 3: Clinical trials

Trials are the most expensive stage. A single Phase 3 study can cost tens to hundreds of millions of dollars (industry estimates vary widely). AI targets three real pain points:

• Patient recruitment: matching eligible patients to trials by scanning electronic health records. Slow recruitment is a leading cause of trial delay.
• Site selection: predicting which hospitals will actually enroll patients.
• Trial design and synthetic control arms: using historical patient data to reduce the number of people who must receive placebo.

Regulators are engaged here. The US Food and Drug Administration (FDA) has published guidance on the use of AI to support regulatory decision-making for drugs and biologics. The European Medicines Agency (EMA) has issued a reflection paper on AI across the medicinal product lifecycle. Neither has approved fully autonomous trial decisions.

Where value is real: recruitment and document automation. Writing a clinical study report (CSR) draft with a large language modellarge language modelA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.Voir la définition complète → can save weeks.

Where it is theater: promising AI will "predict trial success." Outcomes depend on biology and chance the model has never seen.

Stage 4: Manufacturing and quality control

This is medtech and biomanufacturing's most underrated AI zone.

Computer vision for QC: inspecting vials, syringes, and tablets for defects at line speed. A camera plus a trained model catches cracked glass or fill errors humans miss on a fast line.

Process control for biologics: growing therapeutic proteins in living cells (a bioreactor) is sensitive to temperature, pH, and oxygen. Models predict when a batch is drifting out of spec so operators intervene before a batch worth millions is lost.

Regulatory anchor: manufacturing changes fall under Good Manufacturing Practice (GMP) and, in the US, 21 CFR Part 11 (rules for electronic records and signatures). Any AI touching a batch record must be validated and auditable.

A worked ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → example (illustrative)

Assume a vision QC line inspects 1,000,000 vials per year. Historic manual defect escape rate: 0.2 percent. Each escaped defect costs an estimated 500 dollars in recall and rework. AI cuts escapes to 0.05 percent.

• Defects escaping before: 1,000,000 x 0.002 = 2,000
• Defects escaping after: 1,000,000 x 0.0005 = 500
• Defects prevented: 1,500
• Annual benefit: 1,500 x 500 dollars = 750,000 dollars

If the system costs 300,000 dollars to build and 100,000 dollars per year to run, first-year net is roughly 350,000 dollars. (Figures are illustrative, not a market benchmark.)

That is a concrete, defensible case. Notice it does not require any breakthrough biology, just a well-scoped vision task with clean labels.

🎬 [VIDEO: "How AI is transforming drug discovery" - youtube.com - an accessible overview of where machine learning fits in the pharma pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.Voir la définition complète →]

Stage 5: Post-market surveillance

Once a product is approved, companies must monitor safety. This is pharmacovigilance for drugs and post-market surveillance for devices.

AI reads adverse event reports, social media, and medical literature to flag safety signals faster. Natural language processing can triage the flood of case reports that regulators like the FDA (via its FAERS adverse event system) and EMA (via EudraVigilance) receive.

Where value is real: automating case intake and duplicate detection. High volume, repetitive, and well understood.

Where it is theater: claiming AI "detects new side effects no human could." It surfaces candidates. Humans and statistics confirm them.

How to tell value from theater

Across all five stages, the same test applies. Ask four questions:

1. Is the task narrow and repetitive? Vision QC and case triage qualify. "Cure cancer" does not.

2. Is there clean, labeled data? Manufacturing lines generate it. Rare-disease trials often do not.

3. What replaces the AI if it fails? If the answer is "a human doing the same task today," the ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → is measurable. If the answer is "nothing, this never worked before," be skeptical.

4. Is the output auditable for regulators? In this sector, a model you cannot explain to the FDA or EMA is a liability, not an asset.

The regulatory reality for AI devices

Medtech has a specific category: Software as a Medical Device (SaMD), software that performs a medical function without being part of a hardware device. The FDA maintains a public list of AI-enabled medical devices it has authorized, which numbers in the hundreds and is dominated by radiology and cardiology imaging tools.

In Europe, these fall under the Medical Device Regulation (MDR) and, from 2026 onward, interact with the EU AI Act, which classifies most medical AI as "high risk" and imposes documentation, transparency, and human-oversight duties. Budget for compliance as part of any AI ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → in Europe.

The realistic balance of power

Big pharma (Roche, Novartis, AstraZeneca, Pfizer) buys or partners rather than builds from scratch. AI-native biotechs (Recursion, Insilico, Isomorphic) supply discovery platforms. Big tech (Alphabet, Microsoft, NVIDIA) supplies compute and foundation models. In medtech imaging, GE HealthCare, Siemens Healthineers, and Philips embed AI into scanners.

No single player owns the chain. The durable advantage is proprietary data plus regulatory track record, not the model itself. Models commoditize; a validated clinical dataset does not.

Key takeaways

• AI value concentrates at the ends of the chain: discovery (structure prediction, candidate screening) and operations (vision QC, pharmacovigilance triage). The messy middle (predicting clinical success) is where hype outruns evidence.
• Demand a measurable counterfactual. If AI replaces a human doing a task today, you can calculate ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète →, as in the vial-inspection example. If it promises something never done before, treat it as research spend, not a business case.
• No AI-designed drug has completed the full approval path as of early 2026. Discovery speed is real; end-to-end "AI drugs" are still a claim, not a track record.
• In this sector, auditability is not optional. FDA, EMA, GMP, MDR, and the EU AI Act mean an unexplainable model carries regulatory risk that can erase its efficiency gains.
• The moat is data and regulatory experience, not the algorithm. Evaluate vendors on their validated datasets and clearance history, not their model architecture.