AI for UK Craft Brewing: Consistency Without Losing Safety

AI can help a brewer spot a fermentation that is moving away from a known good profile, compare laboratory results with process history or forecast how much of a seasonal beer might sell. It cannot guarantee flavour, make an unsafe vessel safe or turn incomplete production records into compliant evidence.

Beer is biological, sensory and industrial. Raw materials vary, yeast behaves differently across generations, sensors drift and customers do not experience a statistical average. The useful system helps the brewing team see evidence earlier and make a controlled decision. It does not silently write to plant equipment or declare that every batch will taste identical.

This guide is current to 31 July 2026. Alcohol duty is UK-wide, but food labelling, food-safety enforcement and some public-health requirements involve nation-specific authorities and rules. Licensing, planning, water, waste and workplace obligations also depend on the site and activity. Confirm the current position for the brewery and product; this is not legal, tax or food-safety advice.

Choose a bounded brewing decision

Start with one decision that already has an owner:

Avoid “autonomous brewmaster” as a requirement. Write the intended input, output, permitted users, latency, operating range, prohibited action and fallback. A recommendation can be useful without receiving control authority.

Do not measure success as model accuracy alone. A fermentation alert is valuable only if it arrives early enough, produces an appropriate investigation and does not overwhelm staff with false alarms. A demand forecast is useful only if it reduces waste or stock-outs without disguising uncertainty.

Establish trustworthy batch data first

Many apparent model failures begin as process-data failures. A timestamp may represent sample collection in one system and laboratory entry in another. A tank can be renamed, a probe moved, a manual correction omitted or a product blended across lots.

Create a traceable batch record covering:

• malt, hops, yeast, water treatment and processing-aid lots;
• recipe and approved version;
• vessel, transfers, clean-in-place cycle and equipment status;
• timestamps, units, sensor identity and calibration;
• manual and automated additions;
• gravity, pH, temperature, dissolved oxygen and relevant laboratory results;
• yeast generation and pitch details;
• deviations, holds, rework, blending and losses;
• packaging materials, code, strength determination and release; and
• the person and evidence behind each consequential decision.

HMRC’s Alcoholic Products Technical Guide on records and accounts sets out production and stock information expected for alcohol-duty control, including materials, fermentations, additions, packaging, quantities, strengths, losses and unusual incidents. It also explains record-retention expectations. An AI dashboard is not a substitute for the underlying duty records and audit trail.

The government’s commercial brewing rules overview brings together food-business registration, alcohol-duty and product information. HMRC’s approvals guidance covers the Alcoholic Products Producer Approval regime. Keep the approved premises, process and records aligned with what the plant actually does.

Monitor fermentation without promising identical flavour

Fermentation models can compare a live batch with historical profiles for the same beer, yeast, vessel and operating conditions. Useful outputs include expected-range bands, rate-of-change warnings and a ranked list of plausible data-quality or process checks.

They do not prove the cause. A gravity curve can depart because of yeast health, temperature, oxygen, recipe, sampling, stratification or a faulty instrument. Require a brewer to review the raw trace and obtain the defined sample before changing the process.

Build separate limits:

• Safety and equipment limits enforced by validated plant controls and interlocks.
• Product specification limits that determine hold, investigation or release.
• Model alert bands that prompt attention but do not override the first two.

The model must never widen a safety or specification limit to reduce alerts. Changes to cooling, pressure, agitation or dosing should stay within an approved recipe and equipment envelope, with confirmation and a recorded reason. Preserve local operation when the model service is unavailable.

Track precision, missed material deviations, alert lead time and interventions. Slice results by beer, vessel, yeast generation, season and sensor. A model that performs well on a flagship pale ale may be unreliable for a high-gravity stout or mixed-fermentation product.

Use generative ideas as hypotheses, not recipes

A model can search an approved ingredient catalogue, suggest combinations or explain similarities across prior trials. It cannot taste the beer, confirm ingredient compatibility, predict every fermentation outcome or know that a supplier specification has changed unless those facts are in the governed input.

Constrain generation by:

• available and approved raw materials;
• allergen and processing-aid status;
• target product category and declared strength range;
• equipment, yeast, packaging and shelf-life constraints;
• prohibited claims and brand rules;
• cost and supply assumptions with dates; and
• a complete source trail for borrowed recipe information.

Pilot-brew candidates before scale-up. Use a defined tasting protocol, record the panel and conditions, and include stability and microbiological checks appropriate to the product. A high predicted preference score must not bypass technical release.

Protect confidential recipes and supplier data. Verify whether vendor prompts or files are retained, used for training or accessible to support staff. Generated names, label copy and artwork also need intellectual-property, advertising and packaging review.

Keep food safety and allergens outside model discretion

Every brewery needs a food-safety management system suited to its process. The Food Standards Agency’s MyHACCP service provides a structured route for developing a HACCP-based plan for small food manufacturers. Scotland’s regulator and local enforcement arrangements differ, so use the relevant national guidance and competent advice.

AI may help draft a hazard-analysis workshop pack or identify missing monitoring records. The food-safety team must determine hazards, critical controls, validation, monitoring, corrective action and verification. A generated hazard table is not a validated HACCP plan.

Allergen control is especially vulnerable to confident automation. The FSA’s allergen guidance for food businesses covers legal responsibilities for the 14 regulated allergens. Cereals containing gluten are common in beer; other recipes or processing choices can introduce additional risks. The FSA’s PPDS labelling guidance explains product-specific information, including allergen declaration considerations for alcoholic drinks.

Link ingredient specification, recipe, production record and approved label version. Block release when they conflict. Never let a model infer allergen absence from a marketing name or historical recipe. Supplier change, substitution, rework and shared equipment need governed review.

For wider traceability patterns, see AI in UK food-safety inspection and traceability.

Preserve pressure, gas and machinery protections

Brewing involves hot liquids, steam, chemicals, pressure, confined areas, moving equipment and carbon dioxide. Optimisation must not weaken engineering controls, safe systems of work, permit arrangements or competent inspection.

The Health and Safety Executive’s carbon-dioxide guidance explains that CO₂ is colourless and odourless and can accumulate, creating asphyxiation risk. A fermentation forecast is not a gas detector. Keep appropriate ventilation, fixed or portable detection, alarms, exposure controls and emergency procedures independent of the model.

Do not let AI raise vessel pressure, bypass a relief device, restart equipment during maintenance or recommend entry into a vessel. Safety-instrumented functions and emergency stops must operate locally without cloud availability. Engineers should approve any new control path through formal change management.

Measure near misses and manual overrides, not only energy saved. If operators routinely ignore an alert or work around a recommendation, investigate interface, workload and process design rather than training the model on unsafe acceptance.

Protect operational technology

Connecting a cloud analytics service to historians, sensors, programmable controllers or packaging equipment creates a route into operational technology. Follow the NCSC’s secure connectivity principles for operational technology, including explicit business need, controlled architecture, least privilege, monitoring and the ability to revoke access.

Prefer read-only collection for early pilots. Separate corporate, guest, vendor and plant networks; inventory assets and data flows; use managed identities rather than shared accounts; log configuration and model changes; and test recovery without the analytics service. Remote vendor support needs time-bounded approval and traceable sessions.

Threat-model malicious recipe files, poisoned sensor history, account takeover, ransomware, altered set-points and a compromised software update. Back up recipes and plant configurations offline and rehearse safe manual production or shutdown.

Forecast demand without manufacturing certainty

Demand models can combine orders, historical depletion, events, weather and promotion plans. Their uncertainty matters more for small breweries because one listing, festival or distributor change can dominate the result.

Produce ranges and scenarios, not a single “correct” volume. Separate committed orders from modelled demand and display the age and owner of each commercial assumption. Include tank time, packaging capacity, ingredient lead time, minimum runs, duty, shelf life and channel return risk.

Do not use individual-level profiling to target vulnerable drinkers. The ASA’s alcohol advertising guidance concerning young people explains restrictions around under-18 appeal and targeting. The Portman Group Codes add industry rules for naming, packaging, promotion and sponsorship. Generated campaigns need the same approval as human-written ones.

For adjacent inventory and purchasing practice, see AI in UK food supply chains and restaurants-automation-uk).

A measurable 90-day pilot

Days 1–30: choose one beer, vessel family and read-only use case. Reconcile batch, sensor, lab, duty and quality records; define alert, investigation and release responsibilities; document non-AI baseline, OT boundary and safe fallback.

Days 31–60: back-test on known good batches, material deviations and sensor faults. Then run in shadow mode through enough process variation. Include calibration drift, missing samples, time-zone errors, vessel changes, yeast generations, recipe revisions, network loss and malicious input.

Days 61–90: expose recommendations to trained staff without automatic plant writes. Review alerts each shift, potential safety or food-quality incidents immediately and performance weekly by product and vessel.

Release only when:

• every output traces to the batch, source, units, sensor and model version;
• safety interlocks, HACCP controls and release specifications remain independent;
• alert lead time improves without exceeding the agreed false-alarm burden;
• all material historical deviations are detected at the required rate;
• sensor faults are distinguished from likely process deviations or clearly left unresolved;
• duty, strength, loss and packaging records reconcile with approved records;
• allergen and label conflicts block release;
• no cloud or model failure can create unsafe pressure, temperature, gas or machinery behaviour;
• OT access is least-privilege, monitored and revocable; and
• brewers can reject a recommendation and record the evidence without penalty.

Pause after a missed safety or allergen event, unauthorised equipment write, duty-record discrepancy, unexplained model change, harmful marketing output, data breach or loss of safe local control. Revalidate after recipe, yeast, vessel, sensor, supplier, model, packaging or network changes.

The practical verdict

AI can make brewing evidence easier to see. It cannot replace the craft judgment that connects raw material, fermentation, sensory result and safe release.

Use it to call attention, compare and plan. Keep plant protection, food safety, duty evidence and final batch authority with accountable people and independent controls. Consistency comes from a controlled process—not a promise that biology will obey a prediction.