AI can collapse discovery to a single day. The clock to the clinic barely moves.
Generative and agentic AI are compressing target‑to‑candidate from years toward hours. But the phases that follow — safety, efficacy, regulatory review and manufacturing — are gated by biology and statute, not compute. Drag the control to see how much, and how little, end‑to‑end time this actually saves.
Illustrative central estimates; cited ranges shown below. Figures are for a novel small‑molecule programme (see sources).
How long does the discovery phase last?Discovery phase: 1.3 years
Years from “I want a drug for disease X” to patient access
Time to patient
11.4 years
of 14.5 years status quo
Time saved by AI
3.1 years
vs 14.5 years baseline
Total reduction
21%
of end‑to‑end timeline
Invariant floor
9 years
clinical + regulatory core
Why the floor holds
Discovery is the compressible fraction — roughly a quarter to a third of the clock. Everything downstream is bounded by processes AI cannot presently shorten by more than the margins.
Phase
Typical range
Why time stays (largely) constant
Preclinical / IND‑enabling
1–2 yr
GLP toxicology, safety pharmacology, genotoxicity and CMC packages are gated by protocol‑defined observation periods and dosing durations. NAMs/NATs — organ‑on‑chip, QSP models — can inform but do not yet substitute for the duration‑bound in vivo studies most regulators require for an IND/CTA.3,4
Phase 1
1–2.3 yr
Sequential dose escalation with mandated safety‑observation windows, PK sampling and DLT assessment. The rate‑limiter is dosing cadence and cohort sequencing in humans, not candidate availability.1
Phase 2
2–3.6 yr
Dose‑ranging and proof‑of‑concept require sufficient drug exposure and endpoint maturation across hundreds of patients. Recruitment and follow‑up dominate the timeline.1
Phase 3
2.5–3.3 yr
Statistically powered confirmatory efficacy and safety in thousands of participants. Enrolment, event accrual and endpoint follow‑up set the floor — a function of the disease’s biology, not of compute.1
Regulatory review
0.3–2 yr
Statutory assessment windows plus GMP inspection and label negotiation: FDA ~10–12 months standard (~6 priority); EMA ~210 active days plus clock‑stops; MHRA International Recognition Procedure as fast as 60–110 days.7,8
Manufacturing & scale‑up
+1–3 yr
Process development, tech transfer and validation batches; for an established platform much runs in parallel with the clinic, but a slice sits on the critical path. AI‑assisted process development and in silico bioprocess optimisation can trim it at the margins (shown as a slightly shorter block in the AI row). For novel modalities a bespoke, purpose‑built GMP facility can be multi‑year and squarely rate‑limiting — toggle it above.9
An Amdahl’s law for drug development
Amdahl’s law is a rule from computing: the overall speed‑up you gain by making one part of a task faster is capped by the fraction of total time that part occupies. Optimise a component that is only a small slice of the whole, and the end‑to‑end gain stays small — however dramatically that slice itself is accelerated.
End‑to‑end time behaves like a system with a small compressible fraction (discovery) and a large serial fraction (clinical, regulatory, manufacturing) bounded by biology and statute. As with parallel computing, speeding only the compressible part yields diminishing total returns: if the invariant floor is f ≈ 0.7 of the timeline, the theoretical ceiling on speed‑up from perfect discovery AI is just 1 / f ≈ 1.4× — a ~30% reduction, no matter how close discovery gets to zero.
The honest caveat. The floor is not literally immovable. AI is already trimming it at the edges — better trial design and site selection, faster recruitment, synthetic/external control arms, adaptive and decentralised designs, in silico evidence gaining regulatory traction, and AI‑assisted process development shaving time from manufacturing scale‑up. But these are incremental gains today, not the order‑of‑magnitude compression seen in discovery. The structural point stands: accelerating discovery alone cannot, on its own, get a genuinely novel medicine to patients dramatically faster.
A real‑world check
The most‑cited AI drug‑discovery success illustrates the thesis rather than refuting it.
Case · INS018_055 (rentosertib) · Insilico Medicine
Rentosertib is a TNIK inhibitor for idiopathic pulmonary fibrosis (IPF) and the first drug with both target and molecule generated by AI. Insilico’s platform took it from therapeutic hypothesis to a nominated preclinical candidate in ~18 months — the candidate was nominated in December 2020, roughly half the traditional 2.5–4‑year discovery/preclinical timeline.5
~2019
AI‑compressed
Target discovery begins (PandaOmics → TNIK)
Dec 2020
~18 months
Preclinical candidate nominated — the fast part ends here
2021–2023
not compressed
IND‑enabling work & Phase 1 (healthy volunteers, New Zealand & China)
The point is the gap. From the preclinical candidate (December 2020) to a single positive Phase IIa readout (3 June 2025) was about 4.5 years — none of it compressible by discovery AI — and that only reaches a mid‑stage efficacy signal, not approval. Phase III began only in 2026.6 AI roughly halved the ~18‑month front end; the clinical and regulatory back end proceeded on its usual, biology‑bound clock. That is precisely the shape of the timeline above.
Sources
BIO / Informa analysis of 9,704 clinical programmes (2011–2020): mean phase durations — Phase 1 2.3 yr, Phase 2 3.6 yr, Phase 3 3.3 yr, submission‑to‑approval 1.3 yr. Summarised at n-side.com.