“The real question is not ‘is the model accurate?’ It is: ‘If a clinician acts on this prediction, does the patient do better?’ Those are completely different questions, and the gap between them is where most of today’s AI lives. Correlation is for generating hypotheses. Guiding an intervention requires interventional evidence.” – Deepak Mittal, Founder & CEO, NextGen Invent, A Straive Company

The Evidence Bar for AI in Clinical Development

AI in clinical development earns trust one decision at a time. The evidence a model needs depends on what it influences, how much weight its output carries, and what happens to patients if it is wrong. This interview covers where that evidence already holds, how sponsors can validate AI across FDA, EMA, and ICH expectations, how to keep adaptive trials and third-party models auditable, and whether AI-guided decisions improve patient outcomes.

Michael Kaminaka: Across Predictive, Generative, and Agentic AI, Where is the Evidence Holding Up, and Where Has Industry Excitement Moved Ahead of Proven Results?

Deepak Mittal: We believe the right way to answer this is to stop focusing on the type of AI and start looking at the decision it touches. Regulators like the FDA look at three things:

• which decision the model influences,
• how much weight its output carries, and
• what happens to the patient if it is wrong?

When it comes to where the evidence is holding up, the pattern is consistent. AI in clinical development works best when it informs a decision that a qualified person still owns and can verify. That is where predictive AI in drug development has earned its place: trial design and feasibility, site selection, patient recruitment, protocol-deviation and data-quality monitoring, and target identification backed by human genetics.

Feasibility assessment answers a simple question before a trial opens: can we enroll enough of the right patients, at the right sites and in the right countries, within the timeline?

Enthusiasm is running ahead of evidence anywhere a human no longer reviews and owns the decision. We see the dangerous middle as generative AI in clinical trials moving into regulated writing, including protocol authoring, safety narratives, and endpoint adjudication, on assumed reliability instead of demonstrated reliability. In that setting, hallucination is not a typo. It is a mischaracterized adverse event.

• FDA’s April 2026 warning letter to Purolea made this concrete. The firm lets AI agents write GMP specifications and records without Quality Unit review. When asked why it skipped process validation, it said the AI never mentioned it was required. The FDA did not need a new rule. It applied one that is forty years old: the quality unit cannot outsource its judgment to a computer.

The same caution applies to agentic AI in life sciences when it takes the form of autonomous clinical agents, and to AI used as a therapeutic, where there are no agreed efficacy standards yet. Continuously learning models in confirmatory trials face a structural problem, not a paperwork one.

Michael Kaminaka: How Do Sponsors Bring Their Validation Approach in Line with FDA’s Demand for Transparency and Reproducibility Without Drifting from EMA and ICH Principles on Precision Medicine?

Deepak Mittal: Think about two countries that both want their citizens to drive safely. One option is to write an identical rulebook with the same speed limits, the same stopping distances, and the same rules for every situation. The other is to agree on shared principles, such as don’t endanger others, stay in control, and adjust to conditions, and then apply them case by case. Regulators chose the second path for AI. A fixed rulebook can’t keep up with AI. Any detailed checklist would be outdated within a year, and it would either block good new ideas or lock in bad ones.

The same logic applies to FDA and EMA. They have not created a single validation approach that clears both agencies, so sponsors shouldn’t try to build one. The question regulators are asking is narrower and harder: is the evidence credible enough to support the specific decision the model is being used to make? AI changes how we generate evidence. It does not change that bar. That is the starting point for AI validation in clinical trials. You define the decision first, then build the evidence around it.

Sponsors also need to be clear about what each framework covers. Clarifying the EU AI Act works the same way. It is product-safety law, not a drug-evidentiary standard, so it does not get your evidence accepted by the FDA. ICH M15 is the right backbone for model credibility, but it was written for pharmacometrics models, not AI. Sponsors still must fill the gaps for AI themselves.

The most crucial step is to engage regulators early and engage them jointly. Bringing FDA and EMA into the conversation before validation plans are locked is how sponsors avoid building evidence for one agency and then rebuilding it for the other.

Michael Kaminaka: When Continuously Learning Models Drive Adaptive Trials, What Practical Guardrails Keep Every Decision Auditable Across Precision-Medicine Subpopulations?

Deepak Mittal: When an AI model learns during a trial, the protocol your IRB approved may no longer describe the model that is treating patients. That is the core risk in adaptive clinical trials with AI, and the guardrails must be built around it.

• The first guardrail is knowing where FDA draws the line. FDA separates a locked model from a continuously learning one, and it will not let a model that rewrites itself sit in a decision-making seat unchecked. Sponsors need to define in advance what the model is permitted to become, how each change will be validated, and how the model will be monitored once it is running.
• The second is auditability, which means being able to rebuild any patient’s decision back to the exact model state that produced it. Every model version that touches a patient decision should be frozen, hashed, timestamped, and stored like a controlled regulatory record, with its training snapshot, thresholds, and validation results attached. If an investigator asks six months later why a patient received a particular dose, you should be able to reconstruct the full path: biomarker result, source data, preprocessing, model version, score, clinician review, and final decision. If you can’t do that, you are not ready to run the trial.
• The third guardrail is the one FDA worries about most, because it is easy to miss. A model that adapts to accumulating data is quietly looking at your trial data again. That inflates the error rate and biases the treatment-effect estimate. So, the guardrail is not only to watch for unfairness across subpopulations. Sponsors need to pre-specify the adaptation rules and prove through simulation that the trial’s operating characteristics hold, including type I error, power, and bias.

Bayesian adaptive designs can accommodate learning only when those operating characteristics are agreed upon beforehand. FDA’s Adaptive Design guidance rests on the same foundation: prospective planning, simulation, and trial integrity.

Check Our Case Study: Clinical Trial Supply Chain Intelligence to Improve Forecasting, Drug Availability & Operational Efficiency

Michael Kaminaka: When a Third-Party Foundation Model Provides the Core Capability, Which Contractual Rights, Such as Model Cards, Audit Rights, Version Freezes, and Change-Control Provisions, Do Sponsors Need to Protect Their Scientific and Regulatory Credibility?

Deepak Mittal: You can outsource the model. You cannot outsource accountability. When the FDA receives the evidence, the sponsor is responsible for it. So, the real question is not what the vendor owes the FDA. It is how much visibility and control we need to stay accountable for something we don’t own.

Before building a hundred-clause contract for every API call, sponsors should apply the lens FDA uses: context of use and model risk. If a foundation model drafts protocol boilerplate that a qualified person fully reviews and owns, the credibility bar is modest, because the person is the control. The full weight applies once the model is stratifying patients, extracting adverse events that feed a safety analysis, or producing anything that influences a regulated decision. Rigor should match the model’s role in the decision. It shouldn’t be spread evenly, and it shouldn’t be spread thin.

Where the model influences a regulated decision, the contract must secure a few things.

• Reproducibility. If a sponsor cannot tell the FDA which model version produced a result, and cannot reproduce that result, the evidence isn’t credible.
• Detection. A contract only works as well as the sponsor’s ability to detect a breach. A “no silent updates” clause with no system that could ever detect a silent update is a compliance fiction. The burden of proving that nothing changed sits with the sponsor, not the vendor.
• Transparency. If a vendor won’t share enough for the sponsor to verify what it is claiming to the FDA, that opacity becomes the sponsor’s credibility risk to carry.
• Continuity. If the vendor retires the model or goes out of business, the sponsor must still be able to reproduce every inference it submitted. An escrow holding model weights that no one can run is worthless.
• Inspection Flow-down. “The vendor won’t cooperate” is not a defense. Cooperation with inspections must be written into the contract, all the way down to sub-vendors.

Michael Kaminaka: Is Predictive AI Truly Raising the Odds of Clinical Success, or Is It Mostly Helping Us Fail Faster and More Efficiently?

Deepak Mittal: Predictive AI in drug development is changing the economics of the business far more than it is changing the bar for approval. A drug still must show substantial evidence of effectiveness and a benefit that outweighs its risk. AI does not lower that standard.

Failing faster still has real value, and FDA welcomes it. The reason is patients, not margins. Every low-probability program stopped early means fewer patients enrolled in a trial of a drug that was never going to work, and more capital available for one that might.

Where AI falls short is in confusing plausible with right. AI is very good at inventing molecules, proposing new chemical structures that look like they could work. It is like a writer who can produce endless, smooth, confident sentences that may or may not be true. The output is impressive, but a plausible molecule is different from one that is right about biology.

The real payoff comes when AI is pointed at what predicts success: targets backed by human genetic evidence. This is causal human biology, meaning there is proof the gene drives the disease and does not simply appear alongside it. AI that designs molecules for target human genetics shows that it truly drives the disease is where it stops being a faster way to fail and starts improving the odds of success.

There is one caution. The same tools that help a sponsor fail faster can also help it design an apparent win through optimized endpoints or enriched populations. Regulators will look at whether that win holds up in patients who will take the drug, and sponsors should hold themselves to the same test.

Michael Kaminaka: What Evidence Do Predictive AI Models Need to Show That Their Predictions Can Reliably Guide Clinical Interventions and Not Just Point to Correlations?

Deepak Mittal: The real question is not whether the model is accurate. It is whether the patient does better when a clinician acts on the prediction. Those are two quite different questions, and most of today’s AI sits in the gap between them.

We think of the evidence in three tiers.

• Discrimination. This tier asks whether the model can rank patients by risk and separate signals from noise. Every model stops here, and the performance numbers often look impressive for adverse events or treatment-effect prediction. But a model can rank patients well and still give the wrong probabilities. It also tells you nothing about whether acting on its output helps anyone. Strong results at this tier are where the conversation starts, not where it ends.
• Calibration and Clinical Utility. Calibration means when the model says seventy percent, about seventy percent of patients respond. Utility means the model delivers a net benefit over the simple strategies of treating everyone or treating no one, which is what decision curve analysis tests. The field keeps skipping this tier.
• Interventional Proof. This is the tier we give the most weight to. It requires a prospective study, ideally randomized, that compares model-guided care with usual care. The results so far are humbling. The first rigorous randomized trial of LLM decision support covered more than 9,600 patients across sixteen clinics in Kenya. It found the tool was safe and improved documentation and prescribing practices, but it did not significantly reduce treatment failure at fourteen days. The process got better. Patient outcomes did not change measurably. That is the trap: mistaking a better-looking workflow for a better outcome.

Our standard for AI validation in clinical trials is simple. The model is not the product. The product is the model plus the action it triggers. That full pathway is what needs to be evaluated, in the population that will receive it, with performance broken out by subgroup. Correlation is useful for generating hypotheses. Guiding an intervention requires interventional evidence.

Wrapping Up

Deepak’s answers keep coming back to one test for AI in clinical development: when a clinician or scientist acts on the model’s output, do patients do better? The evidence holds up where AI supports a decision that a qualified person still owns, and it thins out once that person leaves the loop. How much proof a sponsor needs depends on the decision the model influences and the cost of getting it wrong, not on how advanced the technology is.

The practical steps follow that test. Define the decision before building the evidence and bring FDA and EMA in early. Keep every model version that touches a patient decision traceable, prove through simulation that adaptive designs protect trial integrity, and write reproducibility, continuity, and inspection rights into vendor contracts. Sponsors can license a model from a vendor. The accountability for what it produces stays with them.

Predictive AI already helps sponsors stop weak programs sooner and point discovery at targets backed by human genetics. The organizations that pull ahead will be the ones that can show their AI-guided decisions improve outcomes for the patients who receive the treatment, not just the workflow around it.

Speaker