AI agents are increasingly presented as the next leap in industrial transformation — systems that can move beyond analysing data to making decisions and taking action autonomously.
In theory, they promise a future where manufacturing becomes faster, smarter and more adaptive.
But in food manufacturing, there is a harder truth we need to confront first: AI cannot fix processes we do not yet fully understand.
Manufacturing is not a stable system
This is especially evident in food systems, where manufacturing is far from stable.
Unlike highly controlled digital environments, food production deals with biological raw materials that are inherently variable. Moisture content shifts with storage conditions. Protein functionality changes depending on source and prior processing history.
Small differences in formulation or temperature can lead to significant changes in final product quality.
Take extrusion, for example — a process commonly used to produce puffed snacks and plant-based protein products.
A successful outcome depends on balancing moisture, temperature profile, screw configuration and ingredient behaviour with precision. When conditions align, the product expands and forms as intended. When they do not, the result may collapse, become dense, or fail to form the desired structure.
These are not rare anomalies.
They are part of the everyday reality of manufacturing.
The promise of AI — and what it assumes
In my own work at SIT, including pilot-scale trials at FoodPlant, I am often asked whether AI can be used to predict outcomes or recommend processing conditions in extrusion.
It is an understandable question. If enough production data is collected, it seems reasonable to expect that AI should be able to identify patterns and optimise performance.
In principle, this promise is compelling.
It suggests a shift from trial-and-error towards more predictive, data-driven manufacturing.
But this vision rests on a critical assumption: that the data available fully captures how the system behaves.
In food manufacturing, that assumption rarely holds.
Where the gap lies
AI systems can only learn from what is measured.
Yet some of the most influential variables in food processing — such as how materials behave under heat, pressure or shear — are not always directly observed or consistently recorded. Many process interactions remain tacit, built through experience rather than explicit data capture.
Even when data exists, relationships between variables are often non-linear and context-dependent.
The same processing condition can produce different outcomes depending on formulation, material history, or environmental conditions.
What AI receives, therefore, is often only a partial and unstable representation of reality.
When AI performs poorly in such settings, the conclusion is often that the technology is not mature.
In many cases, the issue lies elsewhere.
We are asking algorithms to optimise systems that remain insufficiently characterised.
More data is not the same as a better understanding
There is growing emphasis on shared datasets, digital toolboxes and industrial AI platforms.
These are important developments — but more data alone does not resolve the underlying challenge.
If variables are defined differently, measured inconsistently across facilities, or recorded without a common structure, combining datasets does not improve understanding.
Also Read: Costing comparison of top 7 popular ERP software for food manufacturing in Singapore
It amplifies inconsistency.
A meaningful dataset — much like a well-designed dashboard — reflects a clear understanding of what variables matter and how they relate to outcomes.
Without that structure, aggregating more data does not lead to better insight.
It simply scales the same limitations.
Why this matters now
These questions extend beyond manufacturing efficiency.
For Singapore, they are becoming increasingly relevant as food resilience rises on the national agenda. Recent geopolitical tensions and disruptions to global supply chains have once again highlighted how vulnerable food systems can be under external shocks.
Singapore imports more than 90 per cent of its food.
In such a context, resilience cannot be defined only by where supply comes from.
It must also include our ability to convert available inputs into stable, nutritious and scalable food products locally.
That capability is a resilience multiplier.
What needs to be built
AI can play an important role in this future.
It can accelerate learning, improve consistency and help detect patterns that are not immediately obvious.
But AI is not the foundation.
Before autonomous systems can make reliable decisions, manufacturing systems must first become more observable, more structured and better understood.
This means:
- Better characterisation of material behaviour
- Clearer definition of operating windows
- More consistent ways of capturing process–material interactions
Also Read: Anomaly Bio powers the future of ingredient manufacturing with US$2.6M in pre-seed funding
Only then can AI move beyond pattern recognition into dependable decision support.
Beyond automation: Building real capability
The future of intelligent manufacturing will not be built by algorithms alone.
It will be built on a deeper mastery of process.
Until we close that gap, AI will not transform manufacturing. It will simply make visible how much of it we still do not fully understand.
—
Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.
The post AI won’t fix manufacturing, until we fix our understanding appeared first on e27.