Beyond AI Efficiency: The Road to Error-Free GMP Manufacturing
AI is already proving its worth in boosting efficiency in bio-manufacturing, but its real potential is delivering right first time manufacturing.
The self-driving paradox
Every day, millions of people climb into two-tonne metal boxes and hurl themselves down narrow strips of tarmac at 70 miles an hour. Families strapped in, inches from potential disaster, with a painted white line as the only mitigation.
When spelt out, this sounds absurd. And if you imagine a world where the car was only invented today, it almost certainly wouldn't be legal: ‘ You want untrained humans, potentially drunk, tired or distracted to control heavy machinery at high speeds, inches apart, relying on only paint and etiquette to avoid death?!’ It would never fly.
And as we all know, the statistics back this up. Road accidents kill over a million people a year, devastating families and communities. This is more people than the most deadly diseases, such as Malaria and breast cancer. But we still accept driving as normal.
Having worked in GMP manufacturing, I couldn't help notice a parallel. In GMP manufacturing, life-saving therapies are on the line and humans - however well trained - are often the weakest link. And believe me, I've seen it all, from mixing up critical labels on a container to accidentally pouring millions of pounds' worth of product down the drain. Even small transcription errors can be immensely costly. No matter how well-controlled your environment may be, humans mess things up.
This reality is what should make self-driving cars and AI in GMP so compelling. Not because humans are lazy, but because human error is the weakest link in the system.
However, the truth is that although self-driving cars are statistically safer, one high-profile Tesla crash still makes global headlines and reinforces mistrust in AI. The bar we set for machines is astronomically higher than the one we set for ourselves.
The same psychology plays out in GMP manufacturing. We accept deviations, missing signatures, and transcription errors as part of the process, even though they risk product quality, safety and patient lives. But suggest that AI might one day help release a batch or catch those errors in real time, and suddenly the scepticism sets in. In both cars and GMP, it’s not the technology that’s holding us back; it’s trust. But trust must be earned through evidence-based validation and explainability.
The route forward
Of course, letting an LLM rip, making mission-critical decisions in a GMP facility would be a terrible idea at this point. In the same way that AI has been incorporated in self-driving cars iteratively, the same principles can be applied in GMP.
Levels of autonomy in self-driving applied to GMP:
Level 0: No automation
Cars: Full human control.
GMP: Paper records, spreadsheets, manual checks - every action is human-driven.
Level 1: Driver assistance
Cars: Cruise control helps, but you’re still in charge.
GMP: AI highlights missing signatures or anomalies in records, but decisions are fully human.
Level 2: Partial automation - We are here now
Cars: Tesla Autopilot steers and brakes, but you must keep your hands on the wheel.
GMP: AI predicts deviations, suggests CAPA actions, flags risks - but humans must review and approve.
Level 3: Conditional automation
Cars: The car can drive itself in certain conditions, but a human must be ready to intervene.
GMP: AI drafts batch release or deviation reports and suggests corrective actions - humans oversee and can override.
Level 4: High automation
Cars: Fully self-driving within specific geofenced zones, no human needed.
GMP: AI runs most QC and documentation workflows, while autonomously controlling unit operations based on AI predictions: humans check in periodically.
Level 5: Full autonomy
Cars: No steering wheel, no human intervention.
GMP: Fully autonomous GMP facility. AI controls, tests, documents, and releases everything.
Roadblocks
While the technology itself is racing ahead, and with our systems already approaching Level 3 autonomy, the real roadblocks are regulation and liability.
If AI were to miss, or even cause, a deviation, who carries the blame?
Is it the operator who trusted the system?
The AI developer who wrote the code?
Or the QP who signed off the batch?
When a human makes a mistake, GMP has clear CAPA procedures. But when an AI makes a mistake or falsely flags an issue that halts production, what happens then?
Regulators are beginning to think about this. Recent guidance from the FDA, EMA, and MHRA emphasises explainability. But major technological challenges remain:
How do we retrain models if new datasets aren’t available or are too small?
How do we trace which parameters or weights drove a given decision?
How do we know if poor-quality data biased the model from the start?
And remember: GMP leaves no margin for error. Patients’ lives depend on every batch being safe, pure, and effective. To reach Level 4 autonomy, two things need to converge:
Validation: AI must consistently and repeatedly prove it can outperform human decision-making.
Explainability: Either errors must become so rare they are practically eliminated, or every error must be fully explainable, traceable, and correctable, with a clear process for preventing recurrence.
Until these two pieces are reconciled, regulators will be right to hold AI to a higher standard than humans. And that’s exactly why we demand near-perfection from machines before we trust them with autonomy.
Conclusion
It’s now clear that AI isn’t some distant future technology; it’s already here, delivering measurable efficiency gains. At Native Labs, we’ve reached Level 2 autonomy: catching transcription errors in real time, generating instant reports, forecasting resources and schedules, and unifying data with extremely lightweight implementation. Our partners are already seeing reduced deviations, faster documentation cycles, and less wasted operator time.
But efficiency is just the first step. The real opportunity is building toward a future where we achieve right-first-time, every time manufacturing. We are now pushing level 3 automation, leveraging our breakthrough AI to predict deviations before they happen and suggest corrective action.
Getting to level 4 and beyond will require working hand-in-hand with:
Regulators, to iteratively validate and build trust in AI-assisted decisions.
Manufacturers, to scale data capture and feedback loops so models can continually improve.
Quality teams, remaining firmly in the loop, shifting their focus from manual data entry to higher-value oversight and decision-making.
This collaborative journey is how we unlock full autonomy safely and, ultimately, a world where AI doesn’t just make GMP more efficient, but fundamentally de-risks manufacturing and accelerates therapies to patients.
If you’re ready to reduce deviations, speed up release, and prepare your facility for the next generation of manufacturing, we’d love to talk.
Interesting, yes right now it feels like if you got AI to processes data and come up with some interesting findings or conclusions you'd have to personally check it was correct which almost suggests doing it yourself instead. I suppose the benefit is that it can generate a large variety of findings and conclusions and you only have to verify the ones that matter.