Why Bolting AI onto Legacy CMC Operating Models Won’t Work — And What Structural Redesign Really Looks Like

As a student of innovation, I found a new analysis by Georgi Tancev, published in IEEE Transactions on Engineering Management (2026), compelling. Drawing on Paul David’s seminal work on electrification, it applies general purpose technology (GPT) theory to argue that AI in pharma is following the same trajectory as electricity and computing before it.

The pharmaceutical industry is in the early stages of AI adoption across Chemistry, Manufacturing, and Controls (CMC). Companies are deploying large language models (LLMs) and machine learning tools for formulation screening, process optimization, analytical method development, deviation investigation, and even regulatory dossier drafting. Yet a growing body of evidence suggests that without fundamental changes to operating models, these investments risk repeating a familiar historical pattern: the productivity paradox.

The core insight: technology alone does not transform productivity. Structural redesign of the operating model around the new “power source” does.

For CMC organizations operating under strict GxP, ALCOA+ data integrity, and regulatory submission requirements, this isn’t an academic debate — it’s an existential one.

The “Unit Drive Question” Hits CMC Hardest

Tancev frames the central challenge as the “unit drive question”: What should the operating model look like if every process could draw on its own dedicated power source?

In the electrification era, factories initially bolted electric motors onto existing belt-and-pulley systems powered by a central steam engine. Productivity gains were modest until the entire factory layout, workflow, supervision logic, and building design were reimagined around distributed electric power.

The same logic applies to AI in CMC today. Legacy CMC operating models are built for centralized human-mediated information processing. Data moves through human layers for routing, transformation, analysis, judgment, and decision-making. This architecture made sense when humans were the only reliable “processors” capable of handling ambiguous signals, regulatory interpretation, and defensible documentation.

AI changes the physics. LLMs and other models can now handle large volumes of structured and unstructured data, draft reports, flag deviations, and even propose process adjustments. But when these tools are simply inserted into unchanged workflows — query a model here, paste output there, human reviews everything — the organization captures only incremental, easily replicable gains. The deeper, cumulative value remains locked behind structural constraints.

Tancev notes that CMC is actually an ideal domain to observe (and solve) this problem because:

• It sits downstream in the value chain where regulatory and quality constraints are most explicit.
• Knowledge is already relatively structured (batch records, analytical data, deviation histories, prior knowledge reports).
• The potential for cumulative, cross-campaign learning is enormous — yet currently episodic and poorly captured.

Four Interdependent Pillars of AI-Driven CMC Transformation

Tancev synthesizes the required response into four interdependent “works” that together constitute a transformation programme rather than a technology implementation:

1. Operating Model Redesign Stop treating AI as a productivity tool layered on top of existing processes. Instead, redesign workflows so that AI becomes the new “central shaft.” This means distinguishing between:

• Routing and automation functions (can increasingly be AI-driven or AI-augmented with clear rules).
• Judgment functions (interpretation of ambiguous signals, weighing competing considerations, making decisions that must be documented and defensible) — where AI supports rather than replaces human oversight.

In practice, this looks like re-sequencing CMC activities around query modes, drafting modes, and retrieval-augmented generation, with explicit human-AI collaboration protocols at each step.

2. Knowledge Architecture for Cumulative Intelligence The real prize in CMC is not faster individual campaigns but compounding organizational learning across campaigns, sites, and modalities. Current systems often produce “productivity everywhere, learning nowhere” because knowledge remains siloed in documents, emails, and individual expertise.

A redesigned knowledge architecture makes experimental learning cumulative, standardized, digitizable, and retrievable. It turns every campaign into both an output and an input for the next. This is the foundation that allows AI to move from episodic assistance to genuine organizational capability building.

3. Measurement Infrastructure Beyond the Productivity Paradox Existing CMC key performance indicators (cycle times, right-first-time rates, submission timelines, deviation closure speed) were designed for the pre-AI regime. They systematically undervalue — or render invisible — the long-term returns from cumulative knowledge building.

Tancev argues for new measurement frameworks that make learning legible from the first operation and provide successive leadership with an evidence base to sustain investment across organizational tenure cycles. Leading indicators of knowledge maturity, retrievability, and cross-campaign reuse become as important as lagging operational metrics.

4. Governance Architecture Tailored to AI Realities in Regulated Environments This is where many current initiatives are most exposed.

Existing frameworks — ISPE GAMP AI Guide, 21 CFR Part 11, EU GMP Annex 11, and the emerging FDA/EMA Guiding Principles of Good AI Practice in Drug Development (January 2026) — represent genuine progress. They emphasize human-centric design, risk-based approaches, data governance, and lifecycle management. However, Tancev highlights a critical limitation: these frameworks largely assume that LLM failures are, in principle, remediable through sufficient engineering effort and traditional validation thinking.

Three failure modes of LLMs are formally and provably irreproducible in the classical sense:

• Hallucination (fluent, confident, but factually false outputs).
• Generation of plausible-sounding but unverified claims.
• Inability to provide independent human-verifiable “universal claims” about quality without trusted authorship.

These failures occur precisely at judgment functions — the areas where traditional governance most needs augmentation. The implication is profound: you cannot govern an LLM by governing the model alone. Governance must be use-case specific.

The same model and technical configuration can require entirely different governance, tier classification, failure-mode analysis (including irreducibility flags), human review protocols, and boundary specifications depending on whether it is used for:

• Early formulation screening (lower criticality).
• Drafting sections of a regulatory submission (high criticality).
• Real-time process control decisions.

Governance burden therefore scales with the number of use cases, not the number of models deployed. Each use case needs its own explicit boundary documentation — what the model is not authorized to do — and a clear delegation protocol between AI and qualified humans.

Aligning with (and Going Beyond) Current Regulatory Signals

The January 2026 FDA/EMA joint principles stress human oversight, risk-based credibility assessment, data provenance, and lifecycle controls. FDA’s draft guidance on AI for regulatory decision-making and its earlier thinking on AI in drug manufacturing further reinforce the need for context-of-use (COU) evaluation and transparency.

Tancev’s framework is highly complementary: it explains why these regulatory expectations are necessary and provides the operating model logic required to meet them sustainably. Organizations that treat AI governance as a bolt-on compliance exercise will struggle. Those that redesign the underlying operating model, knowledge architecture, and measurement systems will find regulatory alignment becomes a natural outcome rather than a constant friction.

What CMC Leaders Should Do Now

1. Run the “unit drive question” exercise on your top 5–10 CMC workflows. Ask: If we could redesign this process from scratch with AI as the primary information-processing capability, what would the sequence, roles, handoffs, and documentation look like?

1. Audit your current knowledge architecture for retrievability and cumulativity. How much prior campaign learning is actually usable by the next team or by an AI system?

1. Map your AI use cases by criticality and judgment intensity. Develop tiered governance playbooks that explicitly address irreproducible failure modes and human-AI delegation.

1. Redefine success metrics. Supplement traditional operational KPIs with leading indicators of knowledge accumulation and organizational learning velocity.

1. Engage cross-functionally early — CMC, Regulatory Affairs, Quality, IT/OT, and Data Science — because this is a structural program, not a technology project.

From Adoption Trap to Structural Transformation

Tancev concludes that the pharmaceutical industry faces a structural adoption trap: an operating model built for centralized human information processing that AI renders increasingly obsolete. The four works he outlines — operating model redesign, knowledge architecture, measurement infrastructure, and governance architecture — are not optional enhancements. They are the minimum conditions for AI to deliver transformative rather than merely incremental value in CMC.

History shows that general purpose technologies eventually deliver enormous productivity gains — but only to the organizations willing to rebuild around them. For regulated CMC teams, the choice is clear: continue applying powerful new tools to 20th-century workflows, or undertake the deliberate, evidence-based structural redesign that turns AI into a genuine source of competitive and regulatory advantage.

The regulated life sciences industry has always excelled at rigorous, risk-based transformation when the stakes are patient safety and product quality. AI in CMC is simply the latest — and perhaps most consequential — arena in which that discipline must now be applied at the level of the operating model itself.

The paper by Georgi Tancev provides both the diagnostic clarity and the constructive roadmap. CMC leaders who internalize its lessons will be far better positioned to navigate the coming decade of AI-enabled pharmaceutical development and manufacturing.

Further Reading

• Tancev, G. (2026). Artificial Intelligence as a General Purpose Technology: The Case of the Pharmaceutical Industry. IEEE Transactions on Engineering Management.
• FDA & EMA (2026). Guiding Principles of Good AI Practice in Drug Development.
• ISPE GAMP 5 Second Edition and AI-related appendices.
• FDA draft guidance: Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products.

Facing questions about incorporating AI into CMC operations, or broader drug development challenges? Contact our team at info@windshire.com or +1 844-686-5750 for expert guidance.

Below are some questions we often run into.

Frequently Asked Questions

Here are 5 FAQs distilled from the piece, written for skimmers:

1. Why isn’t adding AI tools to existing CMC workflows delivering big productivity gains? Because the tools are being bolted onto operating models still designed for centralized, human-mediated information processing. Without redesigning the workflow itself, AI only produces incremental gains — the same “productivity paradox” seen with electrification and computing.

2. What is the “unit drive question,” and why does it matter for CMC? It’s the exercise of asking: if every process could draw on its own dedicated power source (here, AI), what would the workflow look like built from scratch? Applied to CMC, it means redesigning roles, handoffs, and documentation around AI as a primary information processor — not just plugging AI into today’s process.

3. What are the four things CMC organizations actually need to change? Operating model redesign, knowledge architecture (making past learning retrievable and reusable), measurement infrastructure (new KPIs beyond traditional ones), and governance architecture tailored to AI-specific risks.

4. Does this require loosening GxP or data-integrity standards to move faster with AI? No — the opposite. Structural redesign is what makes regulatory alignment (GxP, ALCOA+, etc.) sustainable. Treating AI governance as a bolt-on compliance step is what creates friction; rebuilding the operating model around it makes compliance a natural outcome.

5. What’s the first practical step a CMC leader should take? Run the “unit drive question” exercise on your top 5–10 CMC workflows, and audit your current knowledge architecture to see how much institutional learning is actually retrievable and usable by the next team or an AI system.