AI Was Meant To Fix Drug Discovery – So Why Do Nine In Ten Drugs Still Fail?

AI was supposed to transform drug discovery, or at least that’s what pharma spent the last decade promising.

The pitch made plenty of sense: condense years of molecular hunting into a few months, find hidden candidates traditional approaches missed and put an end to the brutal failure rates that make developing medicine so expensive. Investors bought into the story with billions of dollars, thousands of startups popped up overnight and the tech itself made massive strides.

The numbers that matter most at late stage, however, have barely moved. Nine out of ten candidates entering Phase I clinical trials still fail before getting approved. Developing a drug still takes an average of 9.1 years, and more than half of trial protocols need at least one amendment, with a single Phase III modification costing anywhere from $140,000 to $535,000. These issues seem to come from downstream operational breakdowns rather than initial molecular discovery, occurring right where the bulk of the money, time and human stakes are.

This raises a tough question about where all that money went. Pharma spent a decade aiming AI at the prettiest part of the pipeline instead of the most challenging part that holds everything back. Lab discovery is high-profile and makes for a great story indeed. Clinical execution, which includes handling protocols, targeting the right patients, keeping up with regulations and tracking side effects, isn’t as easy to market. But if we want AI to genuinely fix how drugs are made, the messy operational work is where it actually needs to happen.

The widespread availability of advanced AI from frontier model providers ensures that having the underlying models is no longer a distinctive advantage. The businesses building lasting value are those armed with proprietary data, deep integration into essential workflows and validated, auditable results that satisfy regulators. Virtually none of these firms are public today.

The overwhelming lack of AI-native public companies dedicated to clinical trial execution, regulatory submission or pharmacovigilance says a great deal about past capital allocation and points to where future investments should be directed.

 

Locating The Real Problem

 

Look across the big studies and you’ll find the same trend every time.

Early-stage Phase I trials for AI-discovered molecules deliver success rates between 80 and 90 per cent, landing right around the usual industry average. Push those same candidates into Phase II, however, and success rates sink back to the standard 40 per cent mark. There’s no doubt that AI is sharpening the early search, but it hasn’t solved the wall these drugs hit later on.

That’s because Phase II and III failures aren’t primarily computational failures. They’re biological complexity failures, patient stratification failures, trial design failures and, in some cases, failures to stop programmes early enough when the evidence is already pointing the wrong way. These choices require nuanced judgement calls, which is an area AI is just starting to touch despite getting a lot less funding and attention.

Running pivotal trials costs upwards of $40,000 per patient, while one major Phase III protocol tweak can set a sponsor back $535,000. On top of that, trial schedules have stretched by over a third in the last ten years alone. Most of the famous $2.6 billion price tag to launch a drug goes straight into clinical development, not initial discovery. Up until now, the standard play has been to throw AI at discovery while leaving clinical execution to traditional CROs.

We asked a group of investors, founders and operators whether that is about to change.

 

Our Experts

 

 

• Andre Magrini, Global CRO and AI Revenue Architect, OGI Systems
• Rachael Sparks, Co-founder and CMO, Phase Advance AI
• Muhammad Arif, HealthTech Delivery Lead, TechnBrains
• Dr. Saravanan Thangarajan, Visiting Scientist, Harvard T.H. Chan School of Public Health
• Brett Jansen, CEO and Managing Partner, brettjansen.ai
• David H. Crean, Managing Partner, Cardiff Advisory
• Dr. Harsha Moole, Founder, PhysicianEstate
• Riken Shah, Founder and CEO, OSP Labs
• Katelyn Fitzgerald, Founder and CEO, Luminous Skin Lab

 

 

Andre Magrini, Global CRO and AI Revenue Architect, OGI Systems

 

 

“The industry may not be pointing AI at the wrong bottleneck so much as rewarding the easiest story to finance. Discovery produces dramatic demonstrations; clinical execution produces operational decisions. The larger near-term value is likely in reducing decision latency: improving protocol design, patient and site selection, safety-signal triage, submission preparation, and stop-or-go decisions. Those uses should be measured by avoided cycle time, prevented amendments, earlier termination of weak programmes, and auditable improvements in quality, not by model accuracy alone.

“As frontier models become broadly available, access to a model is not a moat. The durable advantage is proprietary longitudinal data, integration into regulated workflows, human accountability, and evidence that an intervention changes an economic or clinical outcome. The absence of major AI-native public winners downstream suggests that validation costs, fragmented data, conservative procurement, and unclear ownership of outcomes still limit scale. The next investment cycle should favour companies that can prove closed-loop value inside existing clinical and regulatory operations, not merely add another prediction layer.”

 

Rachael Sparks, Co-founder and CMO, Phase Advance AI

 

 

“Pointing efforts at the wrong bottleneck is half of the challenge here; in fact, we are pointing the wrong AI. The phases where most drugs fail are post-discovery to first-in human, when we’re establishing integral doses, dose timing, efficacy, and safety. We need to solve for those points of failure.

“Models of complex human biological systems cannot be derived by LLMs, which think in a mathematically linear fashion, because biological systems are nonlinear. The AI that biotech needs is one that drives models of complex systems that make up the human lifetime in a nonlinear mathematical environment. Our in silico clinical models run from birth to death, through disease progression, and can deliver day-to-day clinical patient endpoints outpacing clinical trials by orders of magnitude. Phase Advance’s custom ML modules and agentic ML systems seek out patterns: adverse reactions, trends in genetic groups, the ideal disease stage to initiate treatment, and novel biomarkers which mark therapeutic progress. Custom ML can’t hallucinate, and we know exactly how it works.

“These older, less resource-intensive AI systems can take the limited information available post-discovery to predict human clinical trial results with startling accuracy. They have predicted complete clinical trial datasets of 10,000 patients per trial, in various diseases and drug classes, and when the human trials followed in time, predicted results matched human results with over 99% overlap. LLMs cannot achieve this. These predictions can and have prevented patient harm, prevented millions of wasted efforts, and saved years of development time. AI will save pharma trillions. But it probably won’t be the AI you were looking at.”

 

Muhammad Arif, HealthTech Delivery Lead, TechnBrains

 

 

“Drug discovery has attracted most of the attention because it is an obvious place to apply AI. But once a drug moves into clinical development, a different set of problems starts to dominate: finding the right sites and patients, keeping data usable across systems, reviewing safety information and getting decisions made quickly enough to keep a programme moving.

“I have seen a similar problem in healthcare software projects. In one engagement, our team had to integrate an insurance clearinghouse, practice-management software, banking systems, and internal workflows so that eligibility checks, claims processing, and payment posting could be automated. Most of the work went into the parts people rarely talk about. What happens when two systems disagree? Where does an exception go? Who reviews a failed match? Can someone trace what happened later? If those questions are not answered properly, automation creates more work rather than less.

“That is why the recent pharma examples are interesting to me. Novartis used AI in its 14,000-patient Leqvio trial and reportedly reduced site selection from several weeks to a two-hour meeting. Boehringer Ingelheim is also applying generative AI to pharmacovigilance case intake, where large volumes of safety information have to be structured and reviewed. I think the next investment opportunity sits in companies that solve these operational problems well. The technology matters, but so do the integrations, review steps, exception handling, and audit trail around it.”

 

Dr. Saravanan Thangarajan, Visiting Scientist, Harvard T.H. Chan School of Public Health

 

 

“The next durable advantage in pharmaceutical AI will not be the model. It will be a validated place in the decision chain.

“Across statewide public-health implementation and hospital-based feasibility work, I kept seeing the same pattern: having the technology or guidance was not enough. The harder part was getting the right decision to happen reliably, at the right point in the workflow, in a way the people responsible for it could trust.

“That is how I would look at AI in clinical development. The value is not “AI for trials” as another software category. It is reducing decision latency: identifying an eligibility criterion likely to choke recruitment before sites stall, surfacing a safety pattern sooner, or giving a sponsor enough confidence to enrich a subgroup, redesign a protocol, or stop a weak programme earlier. The scale of the execution problem is real. Tufts CSDD found that 76% of protocols had at least one amendment, with an average 260 days from identifying the need for an amendment to final oversight approval. Sites then operated with different protocol versions for an average of 215 days.

“So the question I would ask of any AI product in trials, regulatory work or pharmacovigilance is simple: does it change an important decision earlier, with evidence that can be audited and defended? Model access is becoming easier. Trusted workflow position is not. That is where I would expect durable value to concentrate.”

 

Brett Jansen, CEO and Managing Partner, brettjansen.ai

 

 

“Discovery attracted the capital because discovery is the part of the pipeline that demonstrates well. A better molecule is a clean story for a board. A protocol amendment that costs $535,000 is an operations problem, and operations problems do not raise rounds at the same multiples.

“I have spent sixteen years selling technology into healthcare organisations, and I now sit on the capital side as an AI council advisory for private equity firms. The pattern holds across the industry. Money flows to the visible part of the workflow, and value accumulates in the unglamorous part where the failures concentrate.

“Clinical execution is the larger opportunity, and the maths is not subtle. With nine years of average development time and ninety percent attrition after Phase I, the most valuable decision in a portfolio is the decision to stop early. AI is good at precisely that, reading fragmented enrolment and safety data faster than a committee can convene, and surfacing the judgement call while it still has value. The near-total absence of AI-native public companies in trials, regulatory submission, and pharmacovigilance tells you the constraint was never model quality. Proof is the constraint. Regulated buyers require validated, auditable, reproducible output, and that cannot be demonstrated in a ninety-day pilot.

“Model access is becoming a commodity, so the next cycle should fund the companies holding proprietary trial data, sitting inside workflows a sponsor cannot rip out, and able to hand a regulator a complete audit trail.”

 

David H. Crean, Managing Partner, Cardiff Advisory

 

 

“The industry pointed AI at the part of the pipeline that photographs well. Discovery was never the constraint.

“Look at what happened to poster children of discovery. In May 2025, Recursion halted its cerebral cavernous malformation and NF2 programmes after mid-stage data disappointed, then trimmed further to stretch runway into 2027. AI found those molecules. Clinical biology killed them. No amount of compute upstream rescues a Phase II readout.

“Clinical execution is a larger opportunity by an order of magnitude. Site selection, protocol design, enrolment prediction, patient stratification, and the discipline to kill a programme 18 months earlier. Every one of those moves risk-adjusted NPV directly. Better molecules do not, if the trial is designed to fail.

“I would challenge the premise of absence, though. Certara and Simulations Plus are public companies that have been putting model-informed evidence into regulatory submissions for years. The market prices them as software vendors, not platforms. That is the tell. Investors paid a premium for the narrative of discovery and a discount for the work that actually moves approvals. The next cycle of investing belongs to whoever owns the operational data: trial records, safety signals, site performance, payer outcomes. Frontier model access is commoditised. Proprietary longitudinal data and auditable, regulator-ready validation are not.”

 

Dr. Harsha Moole, Founder, PhysicianEstate

 

 

“Short answer: yes, the industry has been pointing AI at the wrong constraint. Discovery got the glamour and the capital because it tells a simple story: find a better molecule. But the failure economics live downstream. Nine in ten Phase I candidates still fail, programmes still run over nine years, and protocol amendments still burn six figures each. That is where the money actually dies, and it is where AI changes decisions, not molecules.

“The gap in AI-native public companies in trials, regulatory and pharmacovigilance is not an accident. These domains are hard to demo, slow to sell, and the buyers are risk-averse. Discovery is easier to pitch to investors, even when the return is worse. The next investment cycle will go to companies with three things: proprietary data, embedded positions in critical workflows, and auditable, regulator-validated results. Model access is commoditising fast. Data and workflow position are the moat.”

 

Riken Shah, Founder and CEO, OSP Labs

 

 

“The argument is right, but I’d push it one step further back. Trial design, patient matching, and pharmacovigilance all depend on a problem that is still badly underestimated: the data feeding those decisions lives across fragmented systems. Clinical records, call-centre transcripts, patient-portal data, and other sources rarely arrive in a consistent, decision-ready format. You can build an excellent predictive model and still spend a disproportionate amount of the implementation effort establishing whether the underlying data is complete, reliable, and fit for purpose.

“Discovery attracted capital because a better molecule is a clean, fundable story. Clinical execution is messier with fragmented ownership, legacy infrastructure, data governance, and workflows that have to survive regulatory scrutiny. The FDA’s current AI principles explicitly emphasise the context of use, data provenance, documentation, validation, performance assessment, and lifecycle management. That changes where I’d look for durable value. Model access is becoming less differentiated; trusted data, workflow integration, and demonstrable evidence are harder to replicate. Tempus is already showing that AI, multimodal data, and clinical-development workflows can converge in a public company, so the opportunity is not an empty category.”

 

Katelyn Fitzgerald, Founder and CEO, Luminous Skin Lab

 

 

“Focusing so heavily on AI for drug discovery without addressing downstream bottlenecks like clinical trial execution feels like optimising the wrong end of the process. If nine out of ten candidates entering Phase I trials fail, the constraints likely lie further along the chain, in trial design and execution, patient stratification, or the regulatory and safety follow-up phases. AI has immense potential in clinical execution, where data-driven insights could make trials faster, more adaptive, and better at identifying viable candidates while also stopping dead ends earlier.

“The near-absence of AI-driven public companies in the clinical execution space suggests an untapped opportunity. The next investment cycle may benefit investors who shift focus from molecule discovery to embedding AI in workflows that directly impact timelines, budgets, and approval rates. The companies that can leverage proprietary data and validate real-world outcomes will likely outperform those relying solely on algorithmic potential.”

 

For any questions, comments or features, please contact us directly.