Where the algorithm stops and the producer starts: drawing the craft line at Asterley Bros

The debate in craft drinks production is no longer whether to use automation. It has moved to something harder: how to use it without dissolving the judgment that gives the product its character. At Asterley Bros we have spent the last six months automating every operational layer we could reach, and the exercise has taught us that the line between what a system should own and what a person must own is not obvious in advance. You find it by paying attention to where the algorithm starts making confident recommendations you would not follow.

Six months of going deep on AI integrations

We have been building our own software tools and integrating AI across the entire operation: operational meeting summaries, client reports, recipe tracking, demand plans, new product development workflows, supply chain monitoring, sourcing, pricing, design briefing, marketing scheduling. Every facet, in the words we use internally. Not dabbling. Not piloting a single use case. Going to the root of how the business runs and asking what should still require a human decision and what should not.

Our assessment after six months is that this is an absolute superpower for an SME. Not because it makes us smarter. Because it releases the team from the parts of the job that are purely mechanical, and lets us spend the reclaimed hours on work that is actually interesting. A Infinity Sky 2026 review of AI adoption across breweries, wineries and distilleries found that operational automation consistently ranked above product development AI as the higher-value use case for small producers. We feel that.

The tasks that should just be automated

There is a category of work that is data-heavy, time-consuming, structurally repetitive, and adds no value by having a person do it manually. For us that is: supplier price-change monitoring, demand forecasting against batch-production schedules, meeting transcription and action-point extraction, client reporting, and the initial drafting of marketing copy against a brief. These are not creative tasks. They are processing tasks. An algorithm does them faster, more consistently, and without getting bored.

The argument for releasing people from the mundane is not abstract efficiency. For a team of six operating a South London workshop on real stakes and thin margins, every hour spent formatting a client report is an hour not spent on a customer conversation, a recipe iteration, or a growth initiative. The opportunity cost of administrative overhead at our scale is brutal. Automation is, in that framing, a craft decision.

Where the algorithm and the producer disagree

The interesting territory starts when the system makes a recommendation we do not follow. This happens most often in two places: demand forecasting and product development.

On demand forecasting, the models we use are good at extrapolating from sales history. They are not good at incorporating qualitative signals that are not in the data yet. Every month we run direct customer events. The feedback from those sessions, the specific questions people ask about botanicals, the reactions to new expressions, the confusion about category positioning in the British market, none of that shows up in a spreadsheet before we act on it. When the forecast and the customer signal diverge, we go with the customer signal. The algorithm has not been in the room.

On product development, the system can track recipe versions, flag ingredient substitutions, model theoretical flavour interactions based on prior batches. What it cannot do is tell us whether a new expression is worth making. That is an aesthetic judgment. Tastry in California has built impressive flavour-prediction modelling that can anticipate how consumers will score a product before it reaches them. Bespoken Spirits uses wood-science AI to accelerate spirit maturation. Diageo's FlavorPrint maps flavour language at scale. All of these are genuine advances. None of them answers the prior question: what should the thing taste like?

What Circumstance Distillery's Ginette tells us

Circumstance Distillery in Bristol worked with the Ginette project to explore AI-assisted gin recipe development. The experiment is worth looking at carefully because it is honest about what the collaboration produced. The system could suggest ingredient combinations with high predicted coherence. It could not tell the distillers whether the result was interesting. That judgment sat, as it always has, with the people doing the making. Craft Spirits Magazine 2026 covered several comparable projects and reached the same structural conclusion: AI narrows the search space for development decisions; it does not make the development decisions.

I have come to think the tension between automation and craft is slightly false as usually framed. The real question is not whether machines make better decisions than people. It is which decisions are clearly improved by removing judgment from them. Formatting a demand report: yes. Deciding what your vermouth should smell like: no.

The table: what we automate and what we do not

The monthly market as a corrective to the algorithm

One of the clearest illustrations of the human side of this divide is the monthly customer event. Having a few hours of direct conversation with people tasting the products, asking questions, expressing confusion or enthusiasm, is irreplaceable for calibrating production and development decisions. It is education in both directions: we explain the products, the category, the botanical choices; customers tell us what lands and what does not.

No automation touches that. The proximity to the decision-making line that comes from being physically present, pouring and talking, is not something you can replicate in a dashboard. It is also the most direct source of the kind of qualitative signal that, when it contradicts the forecast model, should win. We have built our AI integrations with that hierarchy in mind. The system informs. The people decide. When those two things are in conflict, we have found that the people are usually right.

What this means for other small producers

The practical conclusion for any small batch producer thinking about AI is this: start with the overhead. Not with product development AI, not with flavour prediction, not with the applications that large companies like Diageo are exploring at scale. Start with the operational grind that is eating your team's hours and producing no insight. Automate that first. The returns are immediate, the risks are low, and the recovered time goes directly into the work that actually matters.

For deeper context on where automation fits into drinks business operations and where it does not, we write more of these honest dispatches over at Absolution Labs. The work we are doing at Asterley Bros is the test case for everything we write about from both sides of this AI and drinks manufacturing divide: the maker and the technologist, working in the same South London workshop on the same batch, asking the same question about where the line should sit.