Free-wheeling analysis agents won't cut it for regulatory work

01Introduction

In regulated research and development, an analytical result is rarely just a number. Behind a clinical endpoint, a manufacturing trend, or a molecular assay are decisions about which records to include, how to clean the data, which statistical method to use, and whether the assumptions hold. Those decisions can determine the conclusion, and they can become part of the evidence that scientists, engineers, quality teams, and regulators must review.

Now imagine an AI agent performing the analysis in minutes. It opens a dataset, writes code, executes a statistical workflow, and returns a polished answer. The speed is compelling. But what happens if it silently drops duplicate records, changes a transformation, or chooses between two plausible statistical procedures? A plausible result is not necessarily a defensible one.

That is the challenge for AI-assisted regulatory R&D. The analytical workflow must be steerable while it is happening, not merely inspectable after it ends. At Frontier Instruments, we are building Eureka around that principle: agents can do the computational work, while qualified scientists and engineers retain authority over consequential analytical decisions.

02Why a correct answer is not enough

Eureka is an agentic platform for regulatory R&D. Given an analysis task, its agent can write and run code in a persistent Python or R environment, inspect intermediate results, and use scientific skills, domain-specific libraries, and research tools.

But a computationally successful run can conceal choices that deserve scrutiny. Data selection, cleaning, sample exclusion, modeling decisions, hyperparameters, and statistical tests may all change the answer. The right choice depends on the dataset, scientific intent, intended use, and applicable quality requirements. Those are not decisions an agent should quietly make on behalf of the accountable practitioner.

Our design goal is to let Eureka advance routine work while surfacing consequential forks in the analysis. At those points, the scientist or engineer can approve a proposed approach, request comparisons, revise the instructions, or stop the workflow. The choice and its rationale remain visible for later review.

03A benchmark result raised a more important question

In early internal tests on BixBench Verified 50, Eureka scored 80.0% on average across three runs (82%, 78%, and 80%), compared with reported scores of 88.7% for Phylo, 78% for Edison, 65.3% for Claude, and 61.3% for OpenAI (Figure 1). Eureka performed competitively with the leading reported agents, although it did not achieve the highest score. That result was encouraging, but we did not interpret it as validation of regulatory evidence generation. A benchmark primarily tells us whether an agent returns an accepted answer. It does not necessarily tell us whether the analysis was scientifically justified, whether material choices were disclosed, or whether a reviewer could reconstruct them.

We therefore examined more than final scores. We reviewed agent logs and generated notebooks against reference notebooks, independently reproduced selected analyses, and reran tasks with different models, temperatures, and prompts. We wanted to know not only whether the agent found an accepted answer, but whether it reached that answer for a defensible reason.

One seemingly straightforward question exposed the difference.

04The curious case of Bix-27-q5

Bix-27-q5 asks:

After performing principal component analysis (PCA) with 100 components on the log10-transformed gene expression matrix (with pseudocount of 1, samples as rows and genes as columns), what percentage of the total variance in the dataset is explained by the first principal component (reported as a percentage between 0 and 100)?

The apparent recipe is simple: orient the count matrix with samples as rows and genes as columns, add a pseudocount of one, apply a log10 transformation, compute 100 principal components, and report the variance explained by the first component.

Following that recipe, our analysis produced 56.47%. The accepted answer range was 55% to 56%. That discrepancy was not explained by an ordinary eigensolver precision difference. In our investigation, it traced to an unstated decision about how to clean the input data.

Some sample records and their metadata disagreed, including duplicate samples with uncertain lineage. After identifying the duplicates, an analyst still had to make a choice: keep all records, retain one copy, remove an original, or collapse duplicates using a mean or median. These are not interchangeable preprocessing steps. In our checks, alternative cleaning decisions shifted the answer by approximately 0.5 to 1 percentage point (Figure 2).

The result is revealing because the question never tells the analyst how to handle disputed duplicate records. The missing instruction is not a minor detail. It changes the numerical answer and may determine whether that answer falls inside the benchmark’s accepted range.

For regulatory evidence, such a choice would need to be assessed on its scientific merits. A reviewer should be able to see exactly which records were affected, what made them questionable, which treatment was chosen, why it was chosen, and whether a reasonable alternative would materially alter the inference.

The deeper problem is not that an agent might return 56.47% instead of a number inside the accepted interval. It is that a single score cannot reveal whether the agent noticed the ambiguity, concealed it, or made a justified choice.

05What regulators already expect

Regulatory frameworks already require scientific analyses to be justified, traceable, and reviewable. AI-assisted analysis is no exception. When an AI system makes decisions that affect regulatory evidence, those decisions need appropriate scientific justification and documentation.

AI-specific guidance reinforces human oversight. The FDA’s January 2025 draft guidance for AI supporting drug and biological product regulatory decisions describes a risk-based framework for establishing model credibility. The EMA’s September 2024 reflection paper similarly emphasizes transparency, documentation, and human oversight, particularly when AI could influence regulatory decisions. The joint FDA, Health Canada, and MHRA Good Machine Learning Practice principles also emphasize the performance of the human-AI team, not just the model alone.

Existing data-integrity and statistical standards establish the underlying expectations. FDA guidance on out-of-specification investigations and data integrity emphasizes scientifically justified data handling, accurate records, and accountability in their respective pharmaceutical manufacturing contexts. Where applicable, 21 CFR 11.10(e) requires audit trails for electronic records, while 21 CFR 211.68 establishes controls for automated equipment. ICH E9 and E9(R1) address prespecified statistical methods and sensitivity analyses, while ICH E6(R3) addresses risk-based data governance and fitness for purpose in clinical trials.

These frameworks differ in scope and legal status, but they point to a common principle: consequential analytical decisions should be scientifically defensible, documented, and subject to appropriate human review. They do not prescribe Eureka’s particular review interface, and not all apply directly to the BixBench example.

For Bix-27-q5, that principle means duplicate-sample handling cannot be an invisible implementation detail. A reviewer needs to know which samples were affected, why a cleaning strategy was selected, and whether reasonable alternatives would materially change the PCA result.

The concern is not that AI performs the analysis. It is that AI might make consequential choices without giving qualified experts a meaningful opportunity to evaluate and direct them.

06Keeping scientists and engineers in control

Bix-27-q5 illustrates why review must happen when a consequential choice arises, rather than only after the agent has produced its final result.

In Eureka, the intended workflow is straightforward. The agent can inspect the data and identify an issue, such as duplicate sample identifiers paired with conflicting metadata. Before deciding how to handle the records, it presents the issue and the relevant options to a scientist or engineer (Figure 3). The reviewer can inspect the affected records, direct the agent to compare plausible treatments, select an approach and document the rationale, or decline to proceed.

The important shift is from reviewing a finished answer to directing a consequential decision. The agent remains useful because it can do the computational work, identify alternatives, run sensitivity checks, and document the resulting analysis. The scientist or engineer remains accountable because the choice is visible and subject to approval.

A reviewable record should preserve the question posed, input dataset and version, relevant code and outputs, affected samples, alternatives considered, reviewer instruction, rationale, and downstream effects. The level of control and formal validation required will depend on the intended use and regulated context. Human approval is a critical safeguard, but it is not by itself a substitute for qualified personnel, validated systems where required, or a fit-for-purpose quality process.

07Beyond answer accuracy

Benchmarks such as BixBench remain valuable for measuring whether agents can perform challenging scientific tasks. But a regulatory-facing evaluation needs an additional dimension: how agents behave when the problem specification leaves room for scientific judgment.

We should ask whether an agent detects ambiguous records, distinguishes facts from assumptions, surfaces reasonable alternatives, seeks review when a choice could materially change a result, and retains enough evidence for an independent reviewer to reconstruct the analysis. A high answer-accuracy score cannot establish those capabilities by itself.

That motivates how we are designing Eureka. We want AI systems that make analysis faster without obscuring the decisions on which its credibility depends.

08Conclusion

The promise of AI in regulatory R&D is not simply that it can produce more analyses, faster. It is that it can help experts work through complex analyses while keeping the important decisions explicit.

Bix-27-q5 shows how a seemingly routine PCA question can turn on an unstated choice about duplicate samples. In a regulated workflow, that choice is part of the evidence. It should be surfaced, evaluated, justified, and documented, not silently resolved because an agent needs to return a number.

Eureka is designed to keep consequential analytical decisions visible and under human direction. The objective is not to automate away scientific accountability. It is to give scientists and engineers more leverage while preserving the judgment and traceability on which credible regulatory evidence depends.