In life sciences, production AI rarely looks like a chatbot. Hansa's deal with Cradle shows the alternative

Ask a pharmaceutical executive where AI is in production and the first answer is often an employee assistant: summarizing documents, drafting emails, searching SOPs. Useful, but rarely where the money or the risk is. The higher-value AI in life sciences tends to sit somewhere less visible: inside the scientific workflow itself, proposing what to make and test next.

Hansa Biopharma’s October 1 announcement with Cradle is a clean example of that second kind.

What was announced

Hansa, a Swedish biopharmaceutical company, is deploying Cradle’s AI-driven protein design and engineering platform across its discovery pipeline, to design and optimize candidates for new immunomodulatory therapies in autoimmune disease. According to Cradle, its platform combines candidate generation, optimization and continuous learning from experimental data in one workflow, from design to candidate.

Context matters here. Hansa is not a startup running its first AI pilot. Its lead product, imlifidase, sold as IDEFIRIX, is an IgG-cleaving enzyme that is commercially available in Europe and, per the announcement, under FDA review with a PDUFA date in December 2026 for highly sensitized kidney transplant patients. A next-generation molecule is in development for Guillain-Barré syndrome. Hansa’s head of research and early development said the platform lets Hansa scientists design better candidates faster and with a greater probability of success. That is the company’s expectation, not a measured result yet.

Equally important is what the announcement does not say. The AI does not validate therapies, replace laboratory experiments or produce clinical conclusions. It proposes designs; scientists and experiments decide.

1. Scientists set the target profile
2. Model proposes candidates
3. Lab makes and tests them
4. Results update the model
5. Scientists select candidates

AI stepHuman and laboratory step

Why this is the pattern to watch

The model learns from the company’s own experiments. Continuous learning from experimental data means every assay result makes the next round of designs better for that organization’s targets. That is a compounding asset, and it raises the first enterprise question: who owns the improved model, and the weights or parameters derived from your data? Contracts with AI design platforms need to answer that precisely, including what happens at the end of the relationship.

Traceability becomes scientific hygiene. When a candidate reaches development, someone will ask why that sequence was chosen. The answer should be reconstructable: which model version proposed it, from which data, under which constraints, and which experiments confirmed it. The same discipline that makes an IT system auditable makes a discovery decision defensible.

Integration is where the time goes. The platform has to exchange data with the lab’s electronic notebooks, LIMS and assay systems. Formats, identifiers and data quality decide how fast the loop turns, far more than the model does.

Domain experts stay in charge. The design brief, the choice of what to synthesize and the interpretation of results remain scientific judgments. This is the same principle we described in our analysis of Snowflake’s accounts payable system: AI proposes inside a process whose decisions stay with accountable people and verifiable checks.

Where regulation starts to bite

Discovery-stage AI is lightly regulated today, and that is by design. FDA’s January 2025 draft guidance on AI used to support regulatory decision-making for drugs and biologics explicitly leaves out AI used in drug discovery when it does not affect patient safety, product quality or the reliability of study results. The scrutiny starts later, when AI-generated data or analyses enter the evidence a sponsor submits.

That gives discovery teams room to move. It also creates a trap: records that nobody kept during discovery are hard to recreate when a program reaches the stage where they matter. Keeping lineage from the start costs little and saves a scramble later. Access governance matters too; as we noted in our piece on Anthropic’s life sciences verification program, model access in this sector is becoming a governed asset in its own right.

What life sciences teams can take from this

1. Separate the two AI portfolios. Employee productivity tools and scientific workflow AI have different owners, risks and value. Govern and fund them separately.
2. Negotiate data and model rights first. Experimental data, derived models and improvements: who owns what, during and after the contract.
3. Log design lineage from day one: model version, inputs, constraints, selected candidates, confirming experiments.
4. Budget the integration: ELN, LIMS and assay data pipelines are the real project.
5. Keep claims honest. Report probability-of-success gains only when your own data shows them.

The bottom line

The most consequential AI in life sciences is not a chatbot that talks about science. It is a model embedded in the cycle of design, experiment and learning, owned by scientists and fed by the company’s own data. Hansa and Cradle are one case. The architecture questions, data rights, lineage, integration and human judgment, apply to every organization trying to put AI to work at the bench.