In the intake lab of a malthouse, barley is still judged much the way it was fifty years ago: a trained eye, a cut test, a germination count on damp filter paper. On 8 October, the journal Beverages published the most complete map so far of what could change. Three Polish researchers — two based in Legnica, one at the Wrocław University of Science and Technology — reviewed 188 publications on artificial intelligence in beer brewing, covering everything from grain intake to the bottling line. For small breweries and malthouses trying to separate what works from what is conference-stage talk, it is the most useful document published this year.
Where does AI actually get used in brewing?
Mostly in two places. Quality assurance accounts for 54 of the studies the review examined, and fermentation for 45. Everything else — raw material assessment, bottling, production planning — is thinly covered. The research mirrors what matters commercially: knowing your beer is right, and keeping the slowest, most capital-intensive step of the process under control. The authors' central conclusion is that AI in brewing is not one technology but a set of phase-specific tools, each worth judging on its own.
The method counts tell their own story. Among the primary studies, neural networks appear 67 times, classical machine-learning methods 50 times and computer vision 43 times. Large language models and multi-agent systems appear in exactly 7. That last figure is worth keeping in mind the next time a vendor suggests that chat-style agents are already running breweries. A further 37 studies deal with Industry 4.0 architectures and digital twins — plumbing for data, rather than learning methods themselves.
The research is genuinely international, with a European core. Contributions came from 39 countries, led by Poland with 19, Australia and Brazil with 17 each, and Germany with 16. For a European brewer, this matters: much of the published work was done on the same barley varieties, hop regions and brewery scales you operate with.
What does this look like on a fermentation tank?
Neural networks are used to estimate what is happening inside the tank without opening it: predicting how a fermentation will progress, flagging deviations early and supporting temperature control decisions. The commercial appeal is concrete. Fermentation is where a brewery's capital sits idle for days or weeks, so anticipating a sluggish fermentation even one day earlier translates directly into tank capacity and fewer dumped batches.
Sensors do the seeing, models do the interpreting. The review describes electronic noses and spectroscopy combined with machine-learning models for authentication, defect detection, sensory prediction and quality classification. None of this requires a data centre; it requires instruments you may partly own already, plus models trained on your own production history.
One warning deserves bold type. The authors stress that sensor-based models must be revalidated whenever instruments, sample handling or the sample population change. A model trained on last season's malt does not automatically understand this season's. Whoever sells you a sensor-plus-model package should be able to explain their recalibration plan before you sign.
And in the malthouse?
Raw materials are the less-studied end of the pipeline, but the examples are the most relevant for anyone who handles grain. The review describes convolutional neural networks classifying malting barley kernels and detecting defects, X-ray imaging combined with neural networks finding granary weevil inside kernels, image-based hop variety classification, an electronic nose that distinguishes hop varieties, and machine-learning models fed by field sensors to predict hop-cone ripening.
For a malthouse, the practical translation is intake grading. A camera-and-model setup that scores incoming barley lots consistently — kernel size, broken grains, visible defects — gives you an objective record for every delivery, fewer disputes with suppliers and a chance to catch an infestation before it reaches the silo. These are bounded, repetitive visual tasks: precisely the category where computer vision has become affordable for businesses of malthouse size, not just multinationals.
Can AI write a beer recipe?
It has been tried, with varying degrees of seriousness. One team in the review trained a transformer pipeline on public recipe databases and actually brewed the resulting proof-of-concept IPA. Another used deep networks to predict alcohol content, bitterness and colour from a recipe before anything touches a kettle. The honest reading: recipe prediction is a useful sanity check that can save a pilot brew, while recipe generation remains a curiosity. Nobody credible is claiming you can replace a brewmaster's palate with a model trained on internet homebrew recipes.
When this isn't the right move
The review is refreshingly blunt about limits, and they map well onto small-business reality. Validated datasets are scarce, and data-driven models do not generalise beyond the data they were trained on. The authors note that classical statistical methods sometimes match neural networks — and classical methods are cheaper to run and far easier to explain to a quality auditor. They also observe that much of the published work has never been integrated into routine brewery operations.
If your production data lives on paper, start there. A brewery producing a few thousand hectolitres a year with inconsistent logging will get more from a disciplined digital brewsheet than from any model. The sequence that works is unglamorous: capture your measurements consistently for a few seasons, then let a model find the patterns. Buying the model first is how pilot projects die.
FAQ: AI in brewing and malting
Is this only relevant for large industrial brewers? No. The heaviest research areas — fermentation monitoring and quality assurance — scale down well, because the sensors and compute involved have become cheap. What small breweries lack is usually structured historical data, which costs discipline rather than money to build.
What data should a brewery collect before trying any of this? Consistent, timestamped records per batch: gravities, temperatures, pH, times, raw-material lots and tasting outcomes. A model is only a way of finding patterns in records you already keep. Two seasons of clean brewsheets beat ten years of scattered notes.
Does the EU AI Act affect a brewery using these tools? Lightly. Quality-control and fermentation models fall outside the high-risk categories, so the main duty is the AI-literacy requirement that applies to any business using AI professionally. Your bigger regulatory anchors remain food safety and traceability rules, which good data practice serves anyway.
Where is the fastest payback? In most cases, wherever you currently dump or rework product: fermentation deviations caught late, packaging faults found after the fact, or barley lots accepted on a quick visual check. Count what a single avoided bad batch is worth and work backwards from that number.
The broader lesson of the review applies well beyond beer: AI arrives in an agrifood business one bounded task at a time — a camera on an intake line, a model on a tank — not as a single system that runs the plant. Figuring out which task comes first in your brewery, malthouse or cooperative is the real work. It is also the work Cresly does with farms, cooperatives and food producers, and one we know from the inside: our founder runs an organic malthouse in Dinant (mad.be), where questions like barley intake grading are daily reality rather than theory. If you want to think it through for your own production, write to [email protected].