AI is accelerating discovery. Pharma’s next challenge is orchestrating the CMC network that turns candidates into manufacturable products.
When we talk about the lab of the future, we often picture an increasingly autonomous experimental environment: connected instruments, automated workflows, and data flowing directly into the next decision. That work is real, and it has delivered. It also addresses a narrower problem than the one many drug developers now face: moving an entire portfolio of molecules through development and into manufacturing, across a network of partners they do not own, without losing time, data, and context at every handoff.
That is what I mean by CMC-in-the-Loop. Chemistry, Manufacturing, and Controls (CMC) encompasses the process knowledge, analytical methods, manufacturing controls, specifications, and quality evidence required to turn a candidate into a reproducible product and maintain its quality throughout the product lifecycle. CMC-in-the-Loop coordinates that work across R&D, Manufacturing Science and Technology (MSAT), manufacturing, internal sites, and external partners through a shared, governed data model. The digital thread follows the molecule from research through clinical manufacturing and technology transfer.
CMC is the choke point, and AI is tightening it.
CMC is where timelines and costs concentrate. It carries the iterative exchange between process development and clinical manufacturing that turns a candidate into something a company can produce at scale and support with defensible evidence.
AI-driven discovery is pushing more candidates into that stretch, and pushing them in faster. The front of the pipeline has accelerated, while the development and manufacturing path those candidates enter has not scaled at the same rate. A longstanding constraint now plays a greater role in determining how quickly a portfolio reaches patients. If you run CMC operations, you have felt it: candidates arrive faster than the organization can move them through the gates.
From experimental loops to lifecycle orchestration
Lab-in-the-Loop connects computational models, automated experiments, and experimental results in a continuous feedback cycle. Within a defined scientific problem, it can significantly accelerate learning.
CMC-in-the-Loop changes the unit being coordinated. A molecule must cross development functions, quality gates, manufacturing sites, technology transfers, and external partners. The challenge is preserving its process knowledge, execution context, and decision history through each transition.
That changes how scale is measured. The question expands from how quickly an experiment can run to how efficiently a new molecule, process, or manufacturing partner can enter the network without requiring the underlying structures and integrations to be rebuilt. The automated bench retains its value, but it becomes part of a governed system spanning the broader scientific outcome.
CMC no longer lives inside four walls
A modern development program runs across specialized CROs, CDMOs, clinical manufacturing sites, analytical laboratories, and fill-finish partners. In large programs, the external network can number in the dozens.
Those partners are often connected by documents. Batch records get translated by hand from one site’s format to another, certificates of analysis move as PDFs, and study status lives in email and spreadsheets. Each exchange creates another opportunity for data to be re-entered and context to be lost. Across a network of that size, document handoffs can become a rate-limiting step.
Why the current operating model does not scale.
Pharma does not lack capable systems. The problem appears in the spaces between them, and in the operating model used to coordinate work across the organization.
Point solutions create integration debt
The natural response to a gap is to buy a tool for it: a notebook for one team, a specific tool for another, a separate instrument integration, a separate analysis package. Each addresses a real local need, which is why the best-of-breed approach has been the default for so long.
The cost shows up between the tools. Every system added on its own terms introduces another translation burden. Data must be reshaped each time it crosses a boundary, and each crossing is another place where it can break. The way I put it to teams is that a body does not get healthier by adding organs. It needs a nervous system to coordinate the organs it already has. Most pharma IT estates have strong ELN, LIMS, MES, PLM, and QMS systems, each capable in its own lane, with no layer governing the sample-to-report or molecule-to-submission lifecycle across them.
Gartner describes the same pattern. In its 2025 Market Guide for Laboratory Information Management Systems, the firm observes that point solutions can expand laboratory capability while adding technical debt, and identifies digital unified platforms as an emerging architectural direction. The coordinating layer increasingly deserves to be treated as strategic infrastructure rather than another integration project surrounding the applications.
AI is exposing an old organizational gap
The disconnection I am describing predates the current AI wave. The silos, manual handoffs, and missing governance layer were already there. Organizations had enough room in their timelines to bridge the gaps with people and paperwork.
AI takes that slack away. As candidates multiply and partner networks grow, an organization that routes every change through a central IT bottleneck cannot keep pace. A federated operating model allows the domains that own scientific outcomes to move faster within shared enterprise guardrails, supported by common data, standards, and governance.
That operating model cannot close the gap on its own. The underlying technology architecture must also change, because an AI system working across fragmented data inherits an incomplete view of the scientific process. The work has to start with the foundation.
What a coordinating layer requires.
Two design principles make this architecture scalable: a shared language across the network and governance embedded directly into execution.
A common language reduces the translation burden
When workflows and physical hierarchies are expressed using ISA-88 and ISA-95, an early-phase development recipe and a commercial manufacturing recipe rest on the same structural foundation. That common structure reduces the translation required at each partner boundary, so teams can focus on genuine local exceptions instead of rebuilding a bespoke integration for every site.
Some mapping will always be needed because partners have different equipment, terminology, and operating practices. In a network of dozens of organizations, however, reducing that work to the exceptions changes the economics of scale. A recipe can move between sites as governed digital content instead of being reconstructed from a paper manual on arrival.
One governance framework for people and agents
Coordination is only useful when it is governed. In a regulated environment, that means audit trails, version control across data and models, electronic change control with approvals, tracked lineage from a batch down to its samples, and evidence showing how each result was produced.
The part that often gets glossed over is how governance extends to AI agents. People and agents should operate within the same governed environment, with attributable identities, defined permissions, and actions recorded in a common audit framework. The agent works inside additional boundaries: a constrained set of permitted actions, human approval at checkpoints that carry risk, and risk-based validation appropriate to its intended use.
Treating an agent as interchangeable with a trained scientist would be a mistake. The goal is to apply one governance framework while giving each participant controls suited to its role. That structure gives Quality and CMC teams the evidence they need to evaluate, control, and defend the use of AI in regulated execution.
This governed operating model is the foundation behind L7|ESP®, the agentic operating system for precision science we have built at L7 Informatics. Its architecture connects governed data models and ontologies with workflow orchestration, scientific applications, and operational intelligence. Within that environment, L7|SYNAPSE™ can work with the process context, permissions, lineage, and audit history surrounding the scientific work. The AI layer comes last by design because its usefulness depends on the foundation beneath it.
Regulatory information is becoming more structured.
The pull is regulatory as well as operational. FDA’s Knowledge-Aided Assessment and Structured Application (KASA) program is modernizing how the agency captures, manages, and assesses pharmaceutical quality knowledge. In parallel, FDA’s Pharmaceutical Quality/Chemistry, Manufacturing and Controls (PQ/CMC) data standards initiative is defining standardized structures for CMC information submitted in eCTD Modules 2 and 3.
Together, these efforts point toward a more structured and computational approach to pharmaceutical quality information. When Critical Quality Attributes (CQAs) and Critical Process Parameters (CPPs) are captured as queryable data during execution, submission teams begin with records that already exist and retain their connections to the relevant recipe, batch, sample, method, and approval history.
This does not make submission preparation automatic. It makes the process less dependent on reconstructing structured information from documents after the work has already happened. The practical result is more assembly and less archaeology.
The question worth asking.
For a long time, the reflex has been to ask which application or instrument to automate next. That question remains useful, but it cannot address the larger coordination problem. The question that scales is how to orchestrate an entire scientific outcome across internal laboratories and external partners under one governed model.
When recipes, methods, specifications, and quality data travel with the molecule as governed digital content, technology transfer requires far less reconstruction at the receiving site. Partner onboarding becomes an exercise in controlled reuse instead of rebuilding the same integration and translation work for every organization. The operational value accumulates at each handoff, long before the program reaches commercial manufacturing.
CMC-in-the-Loop gives organizations a way to add molecules and partners without rebuilding the digital infrastructure around each one. As pipelines accelerate and partner networks expand, that is the scale problem pharma now has to solve.
—
Sources
– Gartner, Market Guide for Laboratory Information Management Systems, 16 April 2025, ID G00757034.
– U.S. Food and Drug Administration, [Pharmaceutical Quality/Chemistry, Manufacturing and Controls].
– U.S. Food and Drug Administration, [PQ/CMC Data Standards Scope and Development].
– International Society of Automation, [ISA-88 Series of Standards] and [ISA-95 Standard].