What doesn’t travel

1The gap is in execution, not in documentation

Regulated manufacturing runs on documents, and for good reason. SOPs, master batch records, validated methods and transfer packages define what should happen, and the quality system makes sure it is checked. But every experienced operator, analyst or MSAT engineer knows that executing a critical process also depends on things those documents only partly describe: the conditions under which a step behaves differently, the exceptions they have learned to recognise, the judgement calls they make many times a shift, and the small adjustments that keep a process inside its window.

That knowledge is not missing because someone forgot to write it down. Much of it is hard to put into words at all, and most of it only shows up while the work is being done. It lives in people, it moves through apprenticeship, and it leaves when they do.

Where the gap becomes expensive

Technology transfer. McKinsey has estimated that transfers of sterile dosage forms take from 18 to more than 30 months, and that an average transfer requires six to nine full-time equivalents drawn from about a dozen disciplines, across hundreds of individual activities.³ The same analysis argues that best-in-class transfers can be completed in eight to eleven months. In our view, a large part of that difference comes down to how well process knowledge moves between people and sites.

Deviations and batch failures. A widely cited BioPlan survey found that biomanufacturing facilities experienced, on average, one batch failure every 40.6 weeks.⁴ The data dates from 2008, but the figure is still quoted today. Recent analyses link around half or more of GMP deviations to human factors or weaknesses in operational design, and note that training gaps explain only a minority of them; the larger share comes from cognitive overload, unclear instructions and procedural complexity.⁵

Onboarding and qualification. Operators and analysts still largely qualify through “read and understood” training followed by supervised practice, an approach whose limits are well documented: without practice, knowledge decays quickly.⁵ Every new hire, shift change and site expansion draws again on the same few experts.

These are not separate problems. They are the same gap seen at different moments in the life of a process: the distance between how a process is documented and how it is actually executed.

2Why this is solvable now

Tacit knowledge has been recognised as a problem for decades. What has changed is that several conditions now hold at the same time.

Capture has become practical

Until recently, turning expert practice into something reusable meant asking a busy expert to sit down and write it. Multimodal AI models can now interpret an execution from video, speech, documents and instrument data together. That does not take the expert out of the loop, but it changes their role: from author of documentation to reviewer of what was observed.

Guidance can reach the point of work

Tablets, smart glasses and mixed-reality devices are moving into labs and production areas, and mixed-reality formats address many of the ergonomic problems that held back earlier virtual-reality pilots. GMP-grade hardware is not yet standard, and in regulated settings the software that produces compliant records matters more than the device.⁵

Regulation is engaging with AI

In 2025 the European Commission opened a consultation on a revised Chapter 4 and Annex 11 of the EU GMP guidelines and on a new Annex 22 dedicated to artificial intelligence.⁶ Regulators currently favour models that are “frozen” once validated, with mechanisms such as predetermined change control plans as the route towards models that can evolve safely.⁷ Data-integrity expectations under ALCOA+ remain the baseline for any record a system produces.⁸ The direction is clear: AI will enter GMP environments with human oversight and explainability as conditions, not afterthoughts.

The digital stack is maturing without its human layer

Manufacturing execution systems, LIMS, process models and digital twins describe the process itself better every year. Industry thinking on autonomous, “lights-out” manufacturing now depends explicitly on synchronising people, robots and software agents.⁷ And the value of process data is increasingly understood to depend on the context captured with it, at the source.⁹ The context of human execution is the part of that picture that nobody is capturing yet.

3The state of the art: a map, not a ranking

Several families of tools already address parts of this problem, and each is good at what it was built for. Looking at what each one represents, and what it leaves outside, shows where the gap sits.

Two observations follow. First, the gap is not a missing feature in any one of these systems; it is a missing layer between them. Second, as AI makes content cheaper to produce, the lasting value is likely to sit in the layer that hosts, versions, validates and governs operational knowledge, rather than in the content itself, a point HINA Bioventures makes about immersive training.⁵

4What becomes possible

If the unit of work changes from the document to the verified execution, operational knowledge can be represented as a model: the actions in a procedure, the states and conditions that matter, the exceptions experts recognise, the decisions they take, and the evidence each execution leaves behind. Experts review and approve that model instead of writing it from scratch.

The same model then serves purposes that today are handled separately:

• Technology transfer. The receiving site gets the conditions, exceptions and judgement calls that do not travel with the transfer package, alongside the documents.
• Onboarding and qualification. Competency is built and measured on real executions rather than on attendance and signatures.
• Troubleshooting and deviations. What someone on another shift already knows becomes available when the problem appears, and investigations start from what actually happened.
• Complex procedure execution. The steps with the most variability or risk are guided in context, with evidence of every execution.

Further out, the same representation can feed the rest of the digital stack: deviation analysis with a human in the loop, which is already regarded as relatively low risk;⁷ immersive training generated from real executions rather than written by hand; process models enriched with human context; and, eventually, the coordination of people, robots and software agents that autonomous manufacturing will require. It is the movement from tacit to explicit knowledge that Nonaka and Takeuchi described three decades ago,¹⁰ with one difference: much of the conversion can now be done by machines and checked by experts.

5Open questions

An honest account of this space has to include what is not yet settled. These are the questions we are working through with the people who would use such a system.

• Validation. What can deliver value outside the GMP record, and what is the path towards validated use as the rules for AI in GMP take shape? Human oversight and frozen models are the obvious first step, but not the last.
• Ownership of operational knowledge. When knowledge about a sponsor’s process is captured at a CDMO, who owns it, and who may reuse it? The same question is being asked about process data more broadly,⁹ and the answer will shape how these systems are adopted.
• Recording people at work. Capturing execution means observing operators and analysts. In Europe this raises data-protection and workplace-representation questions that have to be designed in from the start, together with the trust of the people being observed.
• Explainability. Every recommendation or check should be traceable to the evidence behind it. It is a regulatory expectation and a condition for experts to accept the system.
• Measuring value. Transfer duration, time to qualify an operator and deviation rates are natural measures, but they need baselines that many sites do not collect yet.

6What we are building

Opmetis is developing an operational-intelligence system for regulated manufacturing and laboratory environments. It uses multimodal AI and structured operational knowledge to capture tacit expertise from people, documents and workflows, turn it into reusable procedural knowledge, and provide contextual guidance and traceability during execution.

We are looking for design partners among CDMOs, pharmaceutical and biotech manufacturers and laboratories, to co-define and validate use cases in technology transfer, onboarding, troubleshooting and complex procedure execution. At this stage we work through discovery projects, proofs of concept and pilots. If the gap described here sounds familiar, we would like to hear from you.