Recently there is loads of discussion in scientific community on targeting Quantitative Systems Pharmacology (QSP) with AI (LLM driven) for MIDD (model informed drug development). Recently there is also an interesting turn in the pharma landscape and the latest WSJ article covers this quest.
Models scale.
Judgment does not.
Scientific AI has a firepower problem. The models are fast. The outputs are dense. The decisions — the ones that actually matter — are still made in heads, emails, and meetings. That is the bottleneck.
Modern scientific engines generate extraordinary volumes of output with minimal effort. In drug discovery alone, AI has demonstrated the potential to accelerate candidate identification, predict efficacy and toxicity, and optimize interactions far beyond traditional methods. Yet only a tiny fraction of this output ever becomes actionable evidence suitable for internal governance, partner reviews, or regulatory filings.
Between raw output and defensible insight lies an invisible layer of human judgment: Which scenarios align with protocol intent? Which results are representative rather than artifacts? Which plots genuinely support a conclusion under scrutiny? Today, these choices live in heads, meetings, emails, and tacit institutional norms — not in auditable systems. AI floods the pipeline with results. In regulated science, humans still decide what counts.
Why unguided automation erodes trust
The instinctive response has been more automation. If AI generates the outputs, why not let it select, summarize, and report them? In consumer domains this works. In pharma, it creates a dangerous failure mode: plausible but ungrounded decisions.
When an AI system includes a scenario or plot without a clear rationale, the questions are immediate: Why this result and not another? What assumptions drove the selection? Could an alternative choice have altered the conclusion? How would we defend this to reviewers or regulators?
These are not interface problems. They are governance and epistemology problems.
A 2024 scoping review of AI adoption in healthcare identified trust as the single most significant catalyst for implementation — and the factor most vulnerable to erosion by opacity, absent explainability, and regulatory ambiguity. Clinicians and scientists hesitate when they cannot trace decisions back to explicit intent, data provenance, or protocol constraints. The finding appears consistently across the systematic review literature.
Regulators have reached the same conclusion. The FDA's 2025 draft guidance on AI for regulatory decision-making, and the joint FDA–EMA Guiding Principles for Good AI Practice in Drug Development issued in January 2026, both center on a risk-based credibility framework: context of use, data provenance, documentation of analytical decisions, traceability. Credibility cannot be assumed from model performance alone. It must be demonstrated through structured, auditable evidence that the system operated within defined scientific intent.
Recent frameworks in QSP and computational toxicology show what this looks like in practice. MAPLE (Eliason & Popel, 2026) uses structured validation schemas for traceable LLM–human collaboration in QSP data extraction and calibration. AI-QSP (Goryanin et al., 2026) integrates LLMs with SBML standards for expert-in-the-loop model reconstruction. QSP-Copilot (Saini & Farnoud, 2025) employs multi-agent orchestration with provenance metadata across end-to-end workflows. ToxMCP (Djidrovski, 2026) introduces guardrailed agentic pipelines via the Model Context Protocol, wrapping toxicology and PBPK tools with explicit provenance bundles and policy hooks designed for regulatory-grade auditability. Each of these encodes the same core insight: AI proposes; accountable experts authorize.
Decisions are the missing data layer
Scientific AI excels at transforming structured inputs into structured outputs. It struggles when relevance, priority, and intent remain implicit.
Modelers know which analytes truly matter. Project leads know which scenarios serve clinical goals. Regulatory experts know which outputs must be surfaced and which must not. Because this context is rarely encoded as formal input, AI is forced to either surface everything — creating overload — or guess via heuristics, which is unacceptable in regulated settings.
These frameworks fix that by encoding the missing layer. Structured schemas, markup standards for model exchange, provenance metadata, and guardrails turn tacit judgment into explicit, auditable decision pipelines. The intelligence of the model is no longer doing the work that should belong to scientific intent.
Why reporting reveals the bottleneck
No part of the workflow exposes the decision layer more clearly than reporting.
A report is not a passive document. It is a structured argument: about what was explored, what was prioritized, and what evidence supports a conclusion. Every report embeds dozens of decisions — usually unspoken — that determine its credibility. When those decisions lack provenance, the entire argument becomes indefensible under review. The frameworks above demonstrate how structured provenance and human-in-the-loop checkpoints transform reporting from a manual bottleneck into a reproducible, regulator-ready pipeline.
From documents to decision pipelines
The industry needs to reframe what a report actually is.
Not a static artifact generated at workflow's end. A decision pipeline: an explicit, structured process that defines required evidence types, specifies inclusion and exclusion rules tied to scientific intent, preserves full provenance at every step, and produces outputs that are reproducible, inspectable, and reviewable.
Within this framework, AI is no longer asked to "write a report." It is asked to operate inside a pre-defined representation of scientific authority — exactly as MAPLE, AI-QSP, QSP-Copilot, and ToxMCP do.
Human judgment is not removed. It is made explicit through metadata, validators, and approval gates.
This directly aligns with what regulators are asking for. The FDA credibility framework and FDA–EMA principles call explicitly for data governance, documentation of processing steps, risk-based validation, and clear separation of model intelligence from human oversight.
What trustworthy scientific AI actually looks like
Systems that earn adoption in regulated biopharma share five properties — ones these frameworks operationalize in practice:
- Traceability — Every reported result links back to inputs, assumptions, and intermediate decisions.
- Reproducibility — The same encoded intent reliably yields the same output under independent review.
- Constraint awareness — Invalid, out-of-scope, or biologically implausible inferences are rejected transparently, not silently suppressed.
- Decision visibility — Selection logic is inspectable and reviewable, never buried in opaque heuristics.
- Human accountability — AI assists. Humans remain the final authority.
These properties transform AI from a source of plausible outputs into an agent that strengthens governance.
The strategic opportunity
Organizations that succeed with scientific AI will not be those with the largest models or the most aggressive automation. However that is just a place to start with AI and integration.
Winning organization will be those that treat decision-making as the primary engineering challenge.
By encoding scientific intent explicitly — they will move faster without sacrificing trust. They will turn reporting from a bottleneck into a competitive advantage.
This is not a usability challenge. It is a governance and epistemology challenge — one that directly addresses the barriers documented across peer-reviewed literature and regulatory guidance.
The path forward is already being prototyped. The organizations building decision pipelines today will lead tomorrow — not because their AI is the most fluent, but because their decisions are the most trustworthy. Until decision provenance is engineered with the same seriousness as model output, the true bottleneck in scientific AI will remain exactly where it has always been: not in the models, but in the decisions that surround them.
