The state of AI in biotech in 2026

Artificial intelligence in biotech covers machine learning and large language models alongside predictive computational methods, applied across drug discovery and laboratory operations as well as clinical development. The AI in drug discovery market was estimated at approximately $6.93 billion in 2025 and is projected to reach $17.81 billion by 2034 at a CAGR of 9.90%. 2026 is the year the industry stops debating whether AI works and starts asking the harder operational questions, about where it works under what conditions and what stands between early-stage results and production deployment.

The applications with the highest and most consistent adoption all operate on data that is well-structured and integrated into existing scientific workflows, with results that can be verified against established knowledge. The Benchling 2026 Biotech AI Report, based on a November 2025 survey of approximately 100 biotech and biopharma organizations, reports these adoption rates:

• Literature review at 76% adoption
• Protein structure prediction at 71% adoption
• Scientific reporting at 66% adoption
• Target identification at 58% adoption

These use cases succeed because their outputs can be validated against established knowledge and they require no special data infrastructure beyond what research teams already maintain.

DeepMind's AlphaFold 2 was followed by AlphaFold-Multimer and then AlphaFold 3, the latter co-developed by Google DeepMind and Isomorphic Labs. AlphaFold 3 extended prediction to biomolecular complexes, including protein-DNA interactions and small-molecule binding. Structure prediction is now an operational tool in production research.

In antibody design specifically, Chai Discovery's Chai-2 model, described in a July 2025 preprint, achieved a 16% experimental hit rate in zero-shot de novo antibody design across 52 novel targets. At least one binder was identified for 50% of targets, an over 100-fold improvement on prior computational methods. Even so, accurate structure prediction does not automatically produce druggable targets. Current models still show limitations with conformational flexibility and systematic biases toward particular receptor states.

AI is compressing early-stage pipeline timelines, with analysis of AI-assisted programs indicating a reduction in preclinical candidate development from the traditional three to four years to approximately 13 to 18 months. The Benchling survey found that 50% of organizations using AI report faster time-to-target identification, and 42% report improvements in hit rates when using scientific models.

One often-cited clinical example is Insilico Medicine's rentosertib, a TNIK inhibitor for idiopathic pulmonary fibrosis. Insilico reports that its AI platform identified the disease-associated target through PandaOmics and the candidate compound through Chemistry42. Preclinical nomination was completed in approximately 18 months and the program reached Phase 1 in under 30 months. Phase IIa results published in Nature Medicine confirmed general safety and tolerability, though the trial had limitations including a 12-week treatment duration and a small cohort (n=17 to 18 per arm).

The gap between AI adoption and measurable impact has emerged as a prominent theme in life sciences technology. The MIT NANDA initiative's July 2025 report "The GenAI Divide" analyzed over 300 public AI deployments alongside 52 executive interviews and 153 survey responses. It reported that 95% of enterprise generative AI pilots delivered no measurable P&L impact despite an estimated $30 to $40 billion in enterprise spending. A February 2025 Gartner analysis separately predicted that through 2026, organizations would abandon 60% of AI projects not supported by AI-ready data. These findings apply broadly across enterprise sectors. Life sciences organizations are not exempt and face additional constraints from regulated environments and the complexity of biological data.

In our work with biotech and life sciences clients, we have observed a recurring pattern across stalled implementations:

• AI models are deployed in controlled pilot settings using curated datasets, often yielding promising initial results
• Those results fail to transfer to production environments, where data is fragmented and inconsistently formatted
• Post-mortems trace the failure to data infrastructure limits, with model capability further down the list

In our own Q4 2025 CodePhusion survey of over 100 biotech professionals, 73% reported relying heavily on Excel or Google Sheets for data management, and 55% reported that existing tools do not reflect actual workflows. In such conditions, validating AI outputs is difficult because the underlying environment lacks a consistent baseline that can be traced and validated.

Across the biopharma sector, data quality is consistently identified as a primary constraint on AI effectiveness, often outweighing limits in the models themselves. The Benchling 2026 report identifies data quality and availability among the top reasons AI pilots fail. Concerns around IP and security and compliance also rank high.

As explored in our analysis of laboratory digitalisation and technology stacks, laboratory environments were not originally designed for integrated, data-intensive workflows. ELN and LIMS systems often run alongside MES and analytical instrument software with limited interoperability between them. Deloitte's 2025 "Pharma's R&D lab of the future" report, based on a survey of 104 R&D executives, found that 31% of laboratories remain digitally siloed and rely on multiple ELNs and LIMS with limited integration and automation. Effective AI deployment depends on the opposite condition, with continuous and well-governed data environments where inputs are traceable and outputs are auditable.

To date, AI has not conclusively demonstrated improvements in clinical success rates. Its most consistent contribution has been in accelerating early discovery stages. Downstream processes like clinical trials and manufacturing scale-up remain constrained by biological complexity and patient recruitment logistics. Regulatory requirements add further constraints that computational advances do not directly address.

Recursion's REC-994 was discontinued in May 2025 after Phase 2 long-term data did not confirm earlier efficacy trends. It was one of several AI-designed programs that have been deprioritized or wound down following the Recursion-Exscientia merger. Deep Genomics founder Brendan Frey, quoted in The Globe and Mail about the prior generation of AI drug discovery, said, "AI has really let us all down in the last decade when it comes to drug discovery. We've just seen failure after failure." Frey has since argued that newer foundation-model approaches change the trajectory, but the historical record underscores the gap between performance in computational or preclinical settings and outcomes observed in clinical development.

Organizations that have moved AI into production typically encounter challenges that pilots do not surface. In operational use, AI systems behave like infrastructure, needing monitoring and version control alongside change control and clear ownership. Data drift and integration debt become recurring costs, and ongoing maintenance is a continuous line item. Teams structured for pilot delivery alone are often not resourced to absorb these requirements, which is part of why deployment and sustained value diverge.

95% of enterprise generative AI pilots delivered no measurable P&L impact, according to MIT NANDA's 2025 analysis of over 300 public deployments. In biotech specifically, data quality and availability are consistently named among the leading reasons pilots stall, with IP and security and compliance constraints close behind.

On January 14, 2026, the FDA and EMA jointly published Guiding Principles of Good AI Practice in Drug Development, the first coordinated regulatory statement on AI from the two authorities. The principles cover four areas:

• Explainability and transparency of AI outputs
• Human oversight in regulatory-critical applications
• Risk-based and context-of-use approaches across the AI lifecycle
• Data governance and documentation of AI limitations

The document does not create new specific compliance obligations, but it establishes a shared language between the two jurisdictions and reduces the risk of divergent requirements that would have forced separate validation approaches for US and EU markets.

The EU AI Act applies on a staggered timeline. Prohibited-practices rules began applying in February 2025, followed by obligations on general-purpose AI models in August 2025. For biotech software, three milestones matter most:

• On August 2, 2026, most remaining provisions become applicable, including transparency rules under Article 50 and broader governance requirements, plus Annex III high-risk system obligations
• On August 2, 2027, high-risk AI obligations extend to AI that is itself, or is a safety component of, a product under EU harmonization legislation. This includes medical devices and IVDs under MDR and IVDR
• Penalties reach up to €15 million or 3% of global annual turnover for high-risk AI violations (the most relevant tier for biotech), and up to €35 million or 7% of turnover for prohibited-practice violations

For biotech software, this regulatory timeline has architectural consequences. Audit trails must be built into systems from the outset, along with traceability and risk management documentation, since adding them later creates compliance gaps and costs more. The CodePhusion guide to 21 CFR Part 11 and what your lab software needs covers how compliance requirements translate into specific software architecture decisions.

The FDA's January 2025 draft guidance, "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products," introduces a seven-step risk-based credibility assessment framework. Sponsors are expected to develop a credibility assessment plan tied to the model's context of use and risk, and to document the underlying data and model performance. Public comment closed in April 2025, and the timing of final guidance has not been formally announced. The guidance focuses specifically on AI that supports regulatory decision-making, so most current early discovery applications fall outside its scope, though this distinction is not always clearly understood in practice.

The Benchling report found that 80% of organizations surveyed plan to increase AI budgets over the next 12 months, with 23% expecting to double their spend. Notably, much of that capital is moving into data engineering and workflow integration, while model development takes a smaller share. The organizations we see generating durable value from AI tend to build connected and auditable data systems first, then layer AI on top. Deploying AI on fragmented infrastructure and trying to fix data quality afterward consistently produces poor results.

Large pharmaceutical and biotech companies are pursuing platform-oriented approaches. Eli Lilly's TuneLab, launched in September 2025, lets external biotech partners fine-tune AI and ML models built on Lilly's proprietary datasets through a federated learning architecture, so partners gain model access without raw data leaving Lilly's environment. A year earlier, in September 2024, Ginkgo Bioworks launched Datapoints, a fee-for-service offering of curated biological datasets including protein characterization and functional genomics, designed for AI pretraining and fine-tuning. Both approaches reflect a shift in where competitive advantage sits in AI for drug discovery, with data and platform infrastructure now mattering more than algorithmic novelty.

A consequential near-term development is Phase III clinical data from AI-designed drug programs. Multiple candidates are entering or approaching pivotal trials in 2026, and their results will provide the first large-scale test of whether AI-driven discovery translates into improved clinical outcomes. According to Drug Target Review's 2026 predictions, the first regulatory approval of an AI-designed therapeutic is most likely in 2027 or 2028. That approval will validate AI as a legitimate discovery tool, but it will not by itself resolve whether AI systematically improves clinical success rates across therapeutic areas and target classes, a question that will take considerably longer to answer.