AI Is Automating the Literature Review While Creating New Labs to Run.

Research and development is one of the few sectors where AI is clearly doing two things at once.

It is removing parts of the existing workforce, especially where the work is information-dense, repetitive, and structurally similar from case to case. At the same time, it is creating entirely new demand in AI drug discovery, scientific foundation models, self-driving laboratories, and computational research infrastructure.

That is why R&D cannot be described with a single “automation” story.

In the underlying March 25, 2026 assessment, the study covers 66 roles and arrives at a simple average AI replacement rate of 35.5%, with a weighted average of 38.2%. That puts the industry in the limited-assistance band overall. But the average hides a much sharper divide:

information-processing roles are under real pressure,
many wet-lab and leadership roles remain resilient,
and AI-native research roles are expanding rather than shrinking.

R&D is not being automated evenly. It is being restructured by function.

The Market Is Vast, but the AI Layer Is Growing Faster Than the Human Layer

The source file places global R&D spending at roughly $2.6 trillion in 2025, with a professional R&D services market around $931 billion once CROs, contract research, and related technical services are included.

Inside that larger system, the fastest-moving segments are the AI-linked ones:

AI for scientific discovery at roughly $4.8 billion in 2025, rising to about $5.85 billion in 2026 and projected toward $34.78 billion by 2035,
AI drug discovery around $2.9 billion in 2025 and $5.1 billion in 2026,
AI in regulatory affairs on a path from $1.71 billion in 2025 to $3.9 billion by 2029,
and a growing self-driving laboratory market that the source places near $0.8 billion in 2025.

This is the right way to read AI in R&D. The industry is not being replaced by a single tool category. A new stack is being built around discovery, experimentation, regulation, and automation.

The Adoption Wave Is Already Real

The source assessment highlights a set of milestones that make the shift difficult to dismiss:

more than 80% of large pharmaceutical companies have already deployed or piloted AI,
about 26% of newly initiated sponsor-led trials are AI-assisted,
more than 173 AI-discovered or AI-designed pipeline programs are cited for 2026,
AlphaFold has been used by more than 2 million researchers,
GNoME has predicted 2.2 million new compounds, with 736 already experimentally validated,
and more than 50 self-driving laboratories were running globally by the end of 2025.

This is not speculative adoption. AI is already embedded inside real research workflows.

The Industry Splits Cleanly Between Information Work and Physical Research Work

The strongest insight in the source file is that R&D is being hit hardest where the job is mostly information handling.

That includes:

literature search,
document drafting,
knowledge management,
routine statistical analysis,
patent analysis,
and large parts of data management.

The lowest-risk work, by contrast, tends to involve:

setting research direction,
running teams,
making decisions under uncertainty,
conducting complex physical experiments,
or building the AI infrastructure itself.

That is why the industry feels contradictory. AI is strongest in precisely the areas many people once saw as “knowledge work,” while many embodied and frontier-science tasks remain much harder to compress.

The Most Exposed Jobs Sit in Search, Drafting, and Structured Analysis

The highest-risk roles in the study are overwhelmingly information-processing roles.

The highest-exposure roles in the study

Role	Estimated AI replacement rate	Why exposure is high
Literature Search Specialist	80%	Elicit, Rayyan, and related tools already automate large parts of screening and search
Scientific Documentation Writer	70%	SOPs, protocols, CSR drafts, and report scaffolding are increasingly machine-generated
Patent Analyst	65%	Prior-art search, patent mapping, and FTO-style pattern analysis are highly automatable
Data Management Manager	65%	EDC build, cleaning, and structured data workflow management are being automated
Knowledge Management Specialist	65%	Enterprise RAG and knowledge graphs replace large parts of traditional KM work
Junior Research Assistant	60%	Entry-level analysis, literature support, and routine workflow tasks are increasingly absorbable by AI

The common pattern is obvious. These roles revolve around searching, organizing, summarizing, structuring, and drafting from large corpora of information. That is exactly where modern AI systems are already strongest.

The source file cites Elicit as searching 138 million+ papers, and notes that Rayyan can raise screening efficiency by about 90%. Once that is true, the classic literature-review labor model changes immediately.

Scientific Writing Is Being Rebuilt Faster Than Many Labs Are

One of the most exposed roles in the report is the scientific documentation writer at 70%.

That number makes sense. Generative models are already strong at:

creating first drafts of SOPs,
writing study documents,
generating structured reports,
producing protocol scaffolds,
and formatting submission-ready content from templates and prior material.

The source material also cites claims that clinical study report writing can be accelerated by roughly 40% with AI-enabled workflows.

This does not eliminate the need for expert review. But it does mean the old model of building large drafting-heavy support teams becomes much harder to defend.

Wet Labs Still Have a Strong Human Buffer

The industry’s physical side remains far more resistant.

Roles such as:

cell therapy scientist,
nanotechnology researcher,
principal investigator,
lab head,
prototype development engineer,
and many field-specific experimental scientists

remain in the low-exposure band because the work still depends on:

physical manipulation,
tacit lab skill,
troubleshooting in messy real environments,
material variability,
safety constraints,
and scientific judgment under incomplete knowledge.

This is the practical meaning of what the source calls the “physical-world barrier.”

AI can accelerate the reasoning around experiments. It still struggles to replace the person who has to run a delicate cell protocol, adapt a chemical synthesis in real time, or diagnose a failed instrument during a complex assay.

Self-Driving Laboratories Matter, but They Do Not Remove Human Science Yet

The rise of self-driving laboratories is one of the most important themes in the source assessment.

These systems can already:

design experiments,
run automated cycles,
analyze outputs,
and propose next-step iterations.

The source cites:

more than 50 active SDLs by late 2025,
material-discovery workflows that can move 10x faster,
and autonomous enzyme-engineering systems that can complete protein-engineering loops without direct human intervention.

This is real progress. It is also easy to overread.

Today’s SDLs are still concentrated in specific domains where:

experiments can be formalized,
instrumentation can be robotically controlled,
and inputs/outputs are structured enough for closed-loop optimization.

That still leaves a very large portion of R&D outside the fully autonomous zone. So SDLs are best understood as a major growth vector and an important labor reshaper, not as a universal replacement engine.

Biotech and Materials Science Show AI at Full Power and Full Limit

Biotech and materials are two of the most transformed subfields in the report.

In biotech, the most important signals include:

AlphaFold 3,
IsoDDE reportedly outperforming AlphaFold 3 on harder drug-design cases,
large-scale alliances such as those involving Isomorphic Labs, Eli Lilly, and Novartis,
and autonomous protein-engineering workflows.

That is why roles like protein engineer and bioinformatician move much higher on the exposure curve than many classic lab roles.

In materials science, the same pattern appears through:

GNoME and Energy-GNoME,
rapid screening of battery materials,
and AI-accelerated discovery pipelines.

Still, the source file does not make the mistake of equating better prediction with total replacement. Materials scientists and biologists are not disappearing. Their work is shifting from brute-force search toward:

framing the right discovery problem,
validating AI-generated candidates,
interpreting unusual outcomes,
and integrating findings into a broader scientific or commercial program.

Clinical Research Is Being Compressed Through Data and Monitoring

Clinical R&D is one of the most commercially mature AI subfields in the report.

The source points to:

Medidata supporting 36,000+ clinical trials,
more than 120 AI-assisted trials launched in 2025,
AI-enhanced patient recruitment,
faster site activation,
and remote monitoring plus anomaly detection reducing the need for traditional monitoring routines.

That is why data management manager and other operational support roles move into the high-exposure band, while roles like medical monitor physician and senior clinical leadership stay low.

The pattern is the same as elsewhere:

structured coordination and data handling get automated,
while clinical judgment, safety decisions, and high-stakes accountability remain human.

R&D Leadership Remains Hard to Replace for the Same Reason CEOs Are Hard to Replace

The least exposed roles in the entire study are leadership and research-direction roles.

The least exposed roles in the study

Role	Estimated AI replacement rate	Why exposure stays low
Chief Scientific Officer	8%	Vision, portfolio judgment, and leadership remain human
VP of R&D	10%	Resource allocation, organizational politics, and stakeholder management remain human
CTO	10%	Technology bets still require industry judgment
Principal Investigator	12%	Defining the research question and securing funding remain deeply human
Research Director	12%	Team building and cross-disciplinary direction remain human-led

These roles survive for the same reason top leadership survives elsewhere. The value is not in processing information faster. It is in choosing which uncertain path is worth committing resources to.

AI can sharpen the analysis. It does not remove responsibility for the decision.

The One Category Where AI Creates More Jobs Than It Destroys

The most unusual part of the report is the AI and computational science category.

Here, the roles are mostly low exposure not because AI is irrelevant, but because they are the people building the next layer:

AI research scientist,
machine learning researcher,
AI drug discovery specialist,
scientific LLM specialist,
laboratory automation engineer.

These are not “AI victims.” They are AI-native growth roles.

The source correctly frames this as one of the few categories in the wider labor market where AI is creating obvious new employment while destroying older support work nearby.

That is why R&D differs from, say, clerical operations. In R&D, the same force that removes literature-screening staff can simultaneously raise demand for scientists who design AI experiments, automate labs, or translate model outputs into real programs.

Regulation Slows the Final Step of Automation

R&D also has a built-in brake that some tech forecasts underestimate: regulation.

The source highlights:

FDA’s launch of the generative AI tool Elsa,
the emergence of joint FDA-EMA guidance,
growing AI-specific regulatory frameworks,
and the effect of EU AI Act high-risk provisions.

These developments accelerate AI adoption in some ways because regulators themselves are modernizing. But they also ensure that human accountability remains central, especially in regulated life sciences, GMP environments, clinical trials, and pharmacovigilance.

So AI helps the system move faster while still keeping final responsibility human.

What This Means

R&D is not heading toward uniform automation. It is heading toward a harsher division of labor.

The most exposed work is:

literature review,
routine statistical analysis,
scientific drafting,
structured data management,
and classic knowledge-management support.

The more durable work is:

experimental science in the physical world,
leadership and portfolio judgment,
regulatory responsibility,
and AI-native roles that build or supervise the new research stack.

That has concrete implications.

For research organizations:

automate literature, knowledge, document, and routine analytical workflows aggressively,
invest in self-driving lab infrastructure where experiments are structured enough,
and build stronger bridges between wet-lab talent and AI/computational talent.

For researchers:

the safest path is no longer domain knowledge alone,
it is domain knowledge plus the ability to operate, interpret, and govern AI-enhanced workflows.

For early-career professionals:

the biggest risk is building a career around tasks that are mostly search, synthesis, or first-draft production,
because those are exactly the tasks the new stack is swallowing first.

The Structural Conclusion

R&D is one of the clearest cases where AI does not simply replace labor. It changes the shape of scientific work.

It removes large parts of the informational scaffolding around research. It strengthens the leverage of the scientists who can run frontier programs. And it creates a new layer of roles around lab automation, scientific models, AI drug discovery, and computational experimentation.

So the real story is not that AI is replacing science.

It is that AI is automating the literature review while creating new labs to run.

Sources

Precedence Research, AI for Scientific Discovery Market
https://www.precedenceresearch.com/ai-for-scientific-discovery-market
Isomorphic Labs, Drug Design Engine
https://www.isomorphiclabs.com/articles/the-isomorphic-labs-drug-design-engine-unlocks-a-new-frontier
WinBuzzer, Isomorphic Labs AI Doubles AlphaFold 3 Accuracy
https://winbuzzer.com/2026/02/11/isomorphic-labs-ai-doubles-alphafold-3-drug-discovery-accuracy-xcxwbn/
Fortune, Isomorphic Labs First Human Trials
https://fortune.com/2025/07/06/deepmind-isomorphic-labs-cure-all-diseases-ai-now-first-human-trials/
Axis Intelligence, AI Drug Discovery 2026
https://axis-intelligence.com/ai-drug-discovery-2026-complete-analysis/
Fortune Business Insights, AI in Pharma and Biotech
https://www.fortunebusinessinsights.com/ai-in-pharma-and-biotech-market-110287
IQVIA, Global Trends in R&D 2025
https://www.iqvia.com/insights/the-iqvia-institute/reports-and-publications/reports/global-trends-in-r-and-d-2025
Nature, AI Is Threatening Science Jobs
https://www.nature.com/articles/d41586-026-00444-9
Science / AAAS, AI Has Supercharged Scientists but May Have Shrunk Science
https://www.science.org/content/article/ai-has-supercharged-scientists-may-have-shrunk-science
Royal Society Open Science, Self-Driving Laboratories Review
https://royalsocietypublishing.org/rsos/article/12/7/250646/235354/Autonomous-self-driving-laboratories-a-review-of
ScienceDaily, AI-Powered Lab Discovers Materials 10x Faster
https://www.sciencedaily.com/releases/2025/07/250714052105.htm
Google DeepMind, GNoME
https://deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/
ScienceDirect, Energy-GNoME Database
https://www.sciencedirect.com/science/article/pii/S2666546825001375
Cypris, AI-Accelerated Materials Discovery
https://www.cypris.ai/insights/ai-accelerated-materials-discovery-in-2025-how-generative-models-graph-neural-networks-and-autonomous-labs-are-transforming-r-d
Clinical Leader, CRO Industry Outlook 2026
https://www.clinicalleader.com/doc/cro-industry-outlook-the-next-stage-of-clinical-trial-transformation-0001
Dassault Systemes, Medidata AI Leadership
https://www.3ds.com/newsroom/press-releases/medidata-delivers-decade-ai-leadership-500-clinical-studies-and-growing
Clinical Trials Arena, 2026: Time for AI to Deliver
https://www.clinicaltrialsarena.com/comment/2026-the-time-for-ai-to-really-deliver/
GlobeNewswire, AI in Regulatory Affairs
https://www.globenewswire.com/news-release/2026/01/06/3213923/0/en/AI-in-Regulatory-Affairs-Global-Research-Report-2025-A-3-9-Billion-Market-by-2029-Driven-by-Automating-Processes-Expediting-Drug-Approvals-Minimizing-Human-Error-and-and-Data-Accur.html
McKinsey, Rewiring Pharma Regulatory Submissions with AI
https://www.mckinsey.com/industries/life-sciences/our-insights/rewiring-pharmas-regulatory-submissions-with-ai-and-zero-based-design
Nature Communications, Autonomous Enzyme Engineering Platform
https://www.nature.com/articles/s41467-025-61209-y
StatRanker, Top Global Companies by R&D Spending
https://statranker.org/digital-innovation/top-100-global-companies-by-rd-spending/
PharmaShots, Top 20 R&D-Spending Biopharma Companies 2025
https://pharmashots.com/27331/top-20-rd-spending-biopharma-companies-of-2025/
Cypris, AI Tools for Scientific Literature Review
https://www.cypris.ai/insights/11-best-ai-tools-for-scientific-literature-review-in-2026
Elicit
https://elicit.com/
Grand View Research, AI Drug Discovery Market
https://www.grandviewresearch.com/industry-analysis/artificial-intelligence-drug-discovery-market
Stanford HAI, 2025 AI Index Report
https://hai.stanford.edu/ai-index/2025-ai-index-report
UNESCO, 2025 Survey on R&D Statistics
https://www.unesco.org/en/articles/launch-2025-survey-research-and-development-rd-statistics-sdg-95
Nature, Sky-High Pay for AI Researchers
https://www.nature.com/articles/d41586-026-00474-3
Oxford Academic, AlphaFold 3 and Drug Development
https://academic.oup.com/pcm/article/8/3/pbaf015/8180385
PMC, AI Regulatory Perspectives for Pharma GMP
https://pmc.ncbi.nlm.nih.gov/articles/PMC12195787/