Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

When the model is the machine

Hardware-embedded AI can help pharma manufacturers navigate Annex 22.

[Photo: Getty Images]

Profile photo of Nagesh Nama

CEO · xLM Continuous Intelligence

There is a tension on the pharmaceutical manufacturing floor that does not get discussed enough in AI circles. On one side, operations and quality leaders are under pressure to deploy machine learning that can meaningfully reduce defects, detect anomalies, and accelerate real-time decision-making. On the other side, every model running near a critical operation carries a compliance burden that most developers never designed enterprise AI architectures to meet.

The question I keep coming back to is not whether AI belongs on the line—it clearly does—but how do we deploy it in a way that gives regulators, auditors, and quality teams something they can stand behind.

A NEW REGULATORY FRAMEWORK

The regulatory backdrop shifted when the European Commission released the draft of EU GMP Annex 22, the first formal guidance specifically governing AI and machine learning in pharmaceutical manufacturing. There are numerous implications, but the short version is this: generative AI—including large language models—is explicitly not permitted for critical good manufacturing practices (GMP) operations.

The new guidance calls for “deterministic” models. Deterministic simply means the model produces the same output every single time it receives the same input. That predictability is foundational to pharmaceutical quality—and it is precisely what Annex 22 is designed to protect.  The deeper challenge is not which model type to use. It is how to govern that model once it is running in production. Under GMP logic, a validated model must remain frozen, unless a formal, documented change control process authorizes the update. Unauthorized retraining, silent updates pushed over a network, or version drift between environments are not just technical problems. They are audit failures. If an inspector asks, “Is the model running today the model you validated six months ago?” and you cannot answer with documented certainty, you have a problem.

A NEW AI APPROACH IN MANUFACTURING

This is where an approach gaining serious attention in regulated manufacturing becomes relevant: embedding validated AI models directly into purpose-built hardware chips, rather than running them as software on general-purpose servers or in the cloud.

This distinction is like the pressed vinyl record versus a rewritable hard drive. Once a record is pressed, you cannot remix it. Hardware-embedded models work on a similar principle. The validated model is baked into silicon during chip manufacturing, and it cannot be modified in the field. If the model must change for any reason, the hardware board must be physically swapped out through a documented process.

Annex 22 expects manufacturers to maintain a defensible audit trail covering the full lifecycle of a model—training data provenance, validation records, deployment history, and any changes made post-deployment. When the model is locked into hardware, that lifecycle becomes dramatically simpler to document. Version control is no longer a software configuration management problem; it is a hardware serial number. The attack surface for unauthorized modification shrinks to near zero, because there is nothing to patch remotely.

Reproducibility is another core Annex 22 expectation that is guaranteed by the physics of the chip, not by the discipline of a DevOps pipeline. As I explored in a recent Fast Company piece, the foundational work of AI compliance in pharma is often less about the model itself and more about the infrastructure of trust around it.

General-purpose processors are designed to be flexible—they can run many different tasks, but that flexibility comes at a cost. A chip purpose-built to execute one fixed model at the edge incurs none of that overhead. In some benchmark reports this increases throughput by 1,000 times for certain inference tasks. For high-speed tablet inspection, fill-finish monitoring, or inline spectroscopy, that kind of speed difference is not academic. It is the difference between detecting an anomaly in time to act and logging it after the batch has moved on. 

When a model needs to change the operational overhead of a physical board swap is higher than pushing a software update. Organizations must manage spare board inventories, validate replacement hardware, and build supply chain resilience around a specific chip. There is also vendor dependency to consider. Revalidation timelines for a new hardware version may be longer than with a software model update. These concerns are manageable, but only with thoughtful governance from the start.

A PRACTICAL STEP FORWARD

A practical path forward looks something like this. Identify one well-bounded, deterministic model supporting—not autonomously controlling—a single critical or near-critical operation at the edge. Validate it against predefined acceptance criteria. Instrument monitoring to detect performance drift over time. Document every element of the deployment, and treat any hardware swap as a formal change control event with versioned records and a full validation delta package. Start small, build confidence, and let the audit trail speak for itself in the next inspection.

The broader point is this: hardware-secured, deterministic AI at the edge is not a workaround for Annex 22. Used correctly, it is one of the more structurally honest ways to meet its intent. Regulations like Annex 22 are early signals of a much larger global conversation about how AI earns trust in high-stakes environments. The organizations that build their AI infrastructure around auditability and immutability now will be better positioned when that conversation expands—as it will—beyond European GMP into other frameworks, other regions, and other critical industries. Compliance built into architecture tends to age better than compliance bolted on afterward. 

Nagesh Nama is the CEO of xLM Continuous Intelligence.