Distinguish a useful prediction from a viable program
A model can rank a molecule highly without establishing that it can become a medicine. The ranking may help decide which experiment to run next, but the experiment and subsequent development supply different evidence. Treating those stages as equivalent makes both technical and financial expectations difficult to assess.
This guide focuses on platforms that influence candidate selection and may participate in the resulting drug programs. The separate AI discovery and research-services guide addresses paid research delivery. A platform that owns program risk and a provider paid for work can have very different economics despite using related methods.
The context of use determines the claim
FDA's January 2025 draft guidance discusses credibility assessment for AI used to support drug-related regulatory decisions. The linked document is draft, non-binding guidance, not a general authorization for AI-designed medicines.
FDA's development overview distinguishes discovery, preclinical work, clinical research and review. Those stages provide a useful boundary for evaluating claims. A successful prediction or laboratory result should not be described as proof of clinical benefit or as a completed regulatory outcome.
The bottleneck can move to the experiment
Generating more candidates helps only if the organization can evaluate them with relevant experiments and make better decisions. Assay quality, data provenance and the relationship between the measured endpoint and the development objective can limit the usefulness of a sophisticated model.
Evaluation should also distinguish new problems from repetitions of familiar ones. A favorable result on a dataset closely related to training can overstate utility on a genuinely new program. We would ask which decision improved, what alternative process was used for comparison and what evidence remained unavailable after the model's contribution.
Keep the development-risk model explicit
Suppose an illustrative platform reduces one screening activity from $2 million to $1 million. That is a saving in the defined activity if the assumptions hold. It does not mean that the total cost of developing an approved medicine has fallen by half, because other work and failure risks remain.
These figures are invented, not a company result or an estimate of clinical-development costs. A financial model should identify which costs move, whether the decision changes program success and who funds the later stages. More candidates can increase downstream spending if selection quality does not improve.
Read partnership economics by contract stage
A collaboration announcement can include research funding, possible milestones and contingent royalties. Those amounts need separate treatment. A maximum potential contract value does not establish near-term recognized revenue or a probability-weighted value for every program covered by the agreement.
Platform ownership also changes exposure. A company retaining a medicine's development risk may need substantial capital after the discovery stage. A service provider can receive payment without owning the same downstream upside or downside. We would avoid assigning both business models one universal software-like margin or valuation assumption.
What would change our judgment
The case strengthens with reproducible decision improvements, credible program progression and disclosures that distinguish earned revenue from contingent possibilities. Failed or discontinued programs also belong in the record. A persuasive account explains what the platform changed and what still depended on experimental and clinical work.
It weakens if favorable benchmarks do not translate into better decisions or if downstream costs absorb the claimed discovery savings. Our conclusion is conditional: AI can be useful research infrastructure, while medicine development remains an evidence-generating process.