The Real Bet in Pharma AI: Not "Faster," But "Wrong Less Often"

Bringing a single drug to market takes more than a decade on average — and costs hundreds of millions of dollars. But the greatest losses in that journey don't come from being slow. They come from selecting the wrong target, or designing a trial around the wrong patient population. The brutal reality: those mistakes often don't surface until a Phase 2 or Phase 3 failure, years later.

Even after clearing Phase 1, a drug's probability of reaching market sits at just 5%. Industry analysis suggests AI can compress early discovery timelines by 30–40% — yet there is still no meaningful evidence that it has improved pharma's roughly 90% clinical failure rate.

The central question for 2026, the report concludes, is not whether AI can accelerate preclinical timelines — it's whether AI can actually improve clinical success rates.

The real challenge for pharma AI, in other words, is not efficiency. It's judgment. What drug developers need from AI isn't faster candidate generation — it's better selection. Better bets on which candidates and which strategies are most likely to succeed in the clinic.

Four Ways AI Can Improve Clinical Success Rates

Pharma AI approaches the challenge of improving clinical success from four distinct directions.

The first is reducing uncertainty around drug targets — integrating omics, pathology, and clinical data to more rigorously validate targets that look compelling in animal models but may prove insufficiently relevant in human disease.

The second is defining the patient populations most likely to respond. By combining genomic, transcriptomic, and clinical outcome data, AI can help narrow in on the patients where a drug's effect is most likely to manifest. The goal isn't manufacturing an effect that doesn't exist — it's preventing a real effect from being diluted across too broad a population.

The third is improving go/no-go decisions at Phase 1 and Phase 2 — helping teams make sharper calls about whether and how to advance based on early clinical readouts.

The fourth is operational: optimizing site selection, patient matching, and trial execution so that a good drug doesn't fail for avoidable logistical reasons.

Of these four, the third — using early clinical data to inform development strategy — represents the most critical frontier for pharma AI today.

After Phase 1: The Dilemma Every Drug Developer Faces

For decades, Phase 1 oncology trials were understood as a single-purpose exercise: establish safety, tolerability, dosing, and pharmacokinetics. That framing no longer holds.

For complex modalities — immuno-oncology agents, targeted therapies, antibody-drug conjugates (ADCs) — patient selection is as consequential as the molecule itself. By the time Phase 1 data is in hand, teams are already facing critical strategic decisions. The question isn't just "is this drug safe?" It's "is this drug worth advancing — and if so, for whom?"

The questions that follow a Phase 1 readout compound quickly. Do we advance to Phase 2 — and if so, in which tumor type, against which patient population? Do we pursue monotherapy or a combination regimen? What endpoints should anchor the Phase 2 design? And can we defend every one of these choices to an internal portfolio committee, to investors, to partners, and to regulators?

The problem is that Phase 1 data rarely provides clean answers. Patient numbers are small. Tumor types and treatment histories are heterogeneous. Efficacy signals are, in most cases, ambiguous. When a handful of patients show a response, it's often impossible to tell whether that's noise or a reproducible signal in a definable subgroup. Conversely, a drug that looks inert by conventional response criteria may already be driving meaningful biological change in certain lesions or patient subsets — changes that standard assessments simply don't capture.

This is the fundamental dilemma of early clinical development: you are making high-stakes, largely irreversible decisions on the basis of incomplete information.

Advance a drug that doesn't work, and you lose years and hundreds of millions of dollars. Walk away from a drug with a real signal in a narrow population, and you've abandoned both a potential therapy and a significant commercial opportunity.

Three Capabilities AI Must Deliver

So how does AI address this dilemma? For AI tools supporting post-Phase 1 clinical strategy, three functional capabilities define the floor.

(1) Quantifying Early Efficacy Signals

Oncology response is not reducible to tumor shrinkage. In a single patient, some lesions may regress while others remain stable or progress. Tumor volume may change little while internal necrosis, contrast enhancement, density, metabolic activity, or tissue heterogeneity shift substantially. Standard response criteria are essential clinical benchmarks — but they don't capture everything about whether a drug is biologically active in a given patient.

AI can track change at the lesion level, not just the patient level. It can detect whether a consistent pattern of regression is emerging in specific lesions even when the overall assessment reads as stable disease (SD). It can flag whether what looks like early progression might instead be the precursor to a delayed response. Radiomics plays a central role here — extracting hundreds of quantitative features from standard clinical imaging (CT, MRI, PET) to characterize tumor morphology, texture, heterogeneity, and change over time.

(2) Identifying Responder Subpopulations

In Phase 1, who responded matters as much as how many responded. An overall response rate that looks equivocal across the full trial population may mask a concentrated, reproducible signal in patients defined by a specific biomarker, prior treatment history, metastatic pattern, or imaging profile.

That distinction changes everything about Phase 2 design. A trial that would fail in an unselected population may succeed once the target subgroup is properly defined. AI can synthesize genomic, transcriptomic, and longitudinal clinical data to sketch the contours of that responder population — and give development teams a principled basis for narrowing their next trial design.

(3) Prioritizing Indications and Development Pathways

In a Phase 1 basket trial — where small numbers of patients across multiple tumor types are enrolled simultaneously — teams must eventually decide which indication to expand into a larger cohort or advance to Phase 2. Picking the tumor type with the highest observed response rate is not sufficient.

Response rate is one input among many. Sample size, biological consistency across the enrolled population, unmet medical need, competitive landscape, regulatory pathway, and commercial viability all factor in. A rigorous AI tool should move beyond "where do the numbers look best" to answer a harder question: where does the most compelling biological, clinical, and commercial case converge?

How AI Proves Its Value in Pharma

Ultimately, what's being asked of AI in this space is not a binary answer. In practice, the decision tree has three branches: stop, advance, or redesign. And in early oncology development, the most valuable answer is usually the third — narrow the hypothesis, revise the design, and proceed.

There is real demand for AI after Phase 1 oncology trials. That demand doesn't stem from technological novelty. It stems from a structural problem: irreversible decisions, made on incomplete data, at enormous cost.

In that context, AI earns its place by making clearer which signals to trust and which to interrogate, which patient populations to pursue, how to design the next trial, and — critically — when to stop.

In an industry where 95% of drugs that enter Phase 1 never reach patients, AI won't be remembered as the tool that made better calls on incomplete data.

About Kakao Ventures

Founded in 2012 and backed by Kakao — Korea's leading tech platform — Kakao Ventures is one of Korea's most active Seed-stage venture capital firms, with approximately $280M USD in AUM. We partner with founders before the path is fully defined, when conviction in people matters more than proof in numbers.

Our portfolio includes Lunit (AI cancer diagnostics), Rebellions (AI semiconductors), and Dunamu (operator of Upbit, one of Asia's largest crypto exchanges).

If you're building at the edge of what's possible — we'd like to hear from you.