Every artificial intelligence-assisted conclusion that makes it outside the walls of an R&D organisation is a scientific claim that eventually has to stand on its own

Most of the recent discourse around artificial intelligence (AI) in pharma and life sciences has focused on how AI can help the industry work faster, such as screening more compounds, surfacing more targets and moving candidates through pipelines with fewer manual steps. The questions getting less attention are the upstream ones. What is the AI reasoning from? And can its outputs be verified back to a primary scientific record?

These questions matter more as AI moves from human-in-the-loop assistance towards agentic workflows that chain decisions across multiple steps. Whatever the motivation – speed, scale, cost and/or complexity – agentic systems have fewer points where a human is positioned to catch what the system missed. That makes the origin and verifiability of the underlying knowledge more consequential.

The field is investing in AI to deliver more correct candidates, fewer dead ends and faster progression through pipelines. Whether those goals are realised depends less on where humans sit in the workflow than on what the data and the system underneath are actually built to support.

The challenge for R&D leaders deploying agentic AI is not the output that’s obviously wrong; it’s the accurate-sounding result that passes through several automated steps before anyone has reason to question it. The patterns recur across drug discovery, clinical development and manufacturing, and the consequences amplify as each of these stages move towards increasingly AI-mediated workflows. The most visible failure mode is outright hallucination. AI systems built on uncurated or insufficiently verified sources can return compounds, protein targets, clinical trial identifiers or manufacturing process references that look credible but don’t exist in any primary record. In an agentic workflow, a fabricated identifier upstream can become a real entry in a downstream document.

Less visible – and more difficult to trace – is misattribution. AI drawing from inconsistently sourced data can attribute activities to the wrong compound, apply efficacy or safety signals to the wrong patient subgroup, or assign yield and impurity profiles to the wrong process condition, often with a real-looking citation that doesn’t actually support the claim.

Subtler still is what happens when a number survives aggregation, but its context does not. A bioactivity value without its assay conditions, a clinical endpoint without its patient population, or a process yield without its equipment and material lot is a decontextualised value that looks structured on the surface, and models trained on inputs like these can learn associations that don’t hold up in the conditions that matter.

There’s also the problem of details that were never captured in the first place. Information that determines whether something works, fails, or harms is routinely lost in unstructured data: stereochemistry that separates an active drug from a toxic one; dosing schedules that distinguish a positive trial result from a null one; process conditions that separate a usable batch from a failed one. The AI can’t ask about what isn’t there.

These failures share a common downstream effect. In agentic workflows built on unverified scientific data, outputs can carry every hallmark of a credible answer while telling the user nothing about the underlying reliability. In chained automated steps, no domain expert is positioned to catch the error.

The same data quality problems that cause inefficiency at the bench become an integrity risk at the programme and submission level. The difference is in where in the workflow the error lives and how long it compounds before anyone catches it.

In agentic and automated workflows that run on uncurated scientific data, errors migrate. A wrong target annotation upstream can shape a hit list, then a lead series, then a candidate selection before anyone notices it was wrong. Without a human review step at each handoff, the workflow doesn't flag the original error; it inherits it, builds on it and carries it forward as if it were fact. This matters most at the decision points where conclusions get locked in: go/no-go calls, hit-to-lead progression, validation cycles, and the competitive landscape work that feeds licensing and due diligence decisions. These are the moments where an unverified data point moves from a research artifact to a commitment.

Stakes will continue to rise as AI moves closer to regulatory submissions and clinical decisions. Organisations will increasingly need to demonstrate how their conclusions were reached and which data informed them. Traceability has to be built into the system from the start. It’s far easier to design for now than to retrofit later under the pressure of an agency question or an audit.

Data quality is foundational to any scientific claim an AI-assisted workflow produces. Treating it as an IT concern adjacent to the science understates what is actually at stake, which is the organisation's ability to stand behind its own conclusions when it counts.

Pharma's ultimate goal for AI investment is to develop better drugs more quickly and affordably. Reducing manual steps, eliminating rework and applying AI to more parts of the pipeline are how organisations get there. The AI investments that actually deliver are the ones where the underlying data and the system around it are built for scientific rigour from the start. AI that gets the science right reduces manual steps and rework because its outputs hold up downstream. AI that doesn’t get the science right adds steps rather than removing them, because every plausible-but-wrong output eventually has to be unwound somewhere downstream, or worse, is never noticed.

Every AI system inherits the quality of the data it’s built on. This is a structural property of how these systems work – not a problem unique to any one tool or vendor. The path to automation that holds up under scrutiny runs through data verified before the model ever saw it and through systems that can show where any specific output came from. Whether it’s acknowledged or not, tools that can’t expose those things are making a provenance choice on the organisation's behalf.

In mature AI adoption, scrutiny is a property of the system, not a step in the process. It shows up in how the data was prepared, how outputs are sourced, and how the system handles disagreement and uncertainty. The question shifts from ‘has anyone checked this?’ to ‘could this have been wrong without the system telling us?’

Structure isn’t the same thing as rigour. An organisation can have well-designed schemas, mature ontologies, and clean

canonical identifiers and still be feeding its AI tools data that doesn’t hold up scientifically. The questions below are designed to call attention to that gap so that R&D leaders can tell whether their AI is drawing on genuinely sound data or data that only looks organised on the surface.

The first is a question about provenance: Where did this data come from and what happened to it before the model saw it? Every data set has a chain that runs from a primary observation – an experiment, a clinical record, a process measurement – to whatever form the AI is reasoning over. The question is whether that chain is known, traceable and intact. Was the data extracted from the primary source by people who understood the science or scraped from somewhere downstream? Were entities reconciled so that the same compound, target or process appears as one thing rather than several? Has anything been transformed, normalised, or inferred between the primary record and the form the model sees? Data with an unknown or unstable chain of custody isn’t a safer input just because it sits in a structured schema.

The second is a question about context: What was preserved alongside the data and what was stripped away? Numbers in a table without their experimental, clinical or process context are not scientific data; they’re decontextualised values that happen to appear structured. The question is whether bioactivity values retain their assay conditions, whether clinical endpoints retain their patient population, and whether process parameters retain the equipment and material context in which they were measured. Aggregation that strips context is the most common way structured-looking data loses its scientific meaning.

An AI that can’t surface conflict or uncertainty isn’t safe to remove humans from, no matter how clean its data looks.

Every AI-assisted conclusion that makes it outside the walls of an R&D organisation is a scientific claim that eventually has to stand on its own. When a regulator reviews a submission, a partner conducts due diligence, or an internal review board assesses a candidate, the agent that produced the conclusion will not be in the room. Only the conclusion itself is there and the organisation is left to defend it.

This is what makes the data underneath AI systems an issue of scientific integrity. The decisions an organisation makes now about provenance, context and rigour are the ones that will determine whether its AI-supported science will be defensible in regulatory submissions, licensing deals and the programme decisions made every day.

Agentic AI is moving faster in this industry than the conventions for how to evaluate it. The norms for what a scientific AI system should be able to show, what claims it should be allowed to make and what evidence should sit behind an automated conclusion are still being established. The organisations pressing for better answers now are the ones shaping what the next generation of scientific AI will be expected to deliver.

The third is the question that cuts hardest at the distinction between structure and rigour: By what methods are data quality, consistency and scientific rigour actually maintained against defined standards – by humans, by automation or by both?

A system that runs automated consistency checks against well-defined standards is doing real work, but so is one with domain experts adjudicating where automation cannot. A system that has neither, but has tidy schemas, is producing output that only appears rigorous. The real test is what the system does at the edges when its sources disagree and when it doesn’t know.

Andrea Jacobs is director of Artificial Intelligence at CAS. She has been with the company for over 16 years and has received a BA in Chemistry and Computer Science from Wellesley College, MA, US. Andrea then went on to achieve an MBA from the Fisher College of Business at Ohio State University, OH,US.