The pattern is consistent enough that we have started calling it the adoption ladder. Small drinks producers pick up AI tools for the easy problems first: writing tasting notes, generating social captions, drafting trade emails. These are the surface-layer applications. Useful, certainly. But they are not where the money is.
The decisions that actually move money in a small drinks business are still manual. Which account to visit next week. Whether to restock a SKU that is moving slowly. How to price the next batch against a duty cycle that has risen 17 percent in three years. These are decision-layer problems. And most small producers are not within six months of automating them.
The legibility gap is the real blocker
The ONS's 2026 survey of AI in UK businesses found that 39 percent of firms cited difficulty identifying AI use cases as a primary adoption barrier. That is not a skills gap. It is a legibility gap. Founders know AI exists. They know it does things. They cannot see clearly where it fits in their specific operation.
Whito's research on UK AI adoption patterns among small businesses adds texture to this picture: when you ask small producers what is holding them back, the answers cluster around exactly this problem, not around access to technology or cost of entry. The data is worth reading in full, but the headline finding is simple. The barrier is not the tool. It is the gap between what the tool does and what the business actually needs.
Surface-layer tools have high legibility. You can read a social caption and know within thirty seconds whether it sounds like you. If it does not, you fix it. The feedback loop is immediate and the downside of a bad output is small: one slightly off caption.
Decision-layer tools have low legibility. If your AI account-prioritisation system tells you to visit four venues in south Manchester this week, you cannot verify that recommendation without actually going, spending two days, and measuring the result over the following quarter. The feedback loop is months long. The downside of a bad output is a wasted sales trip, a missed stockist, a quarter of pipeline drift.
That asymmetry is rational. It is not irrational fear of technology. It is a reasonable response to the different risk profiles of different AI outputs.
Why the adoption order matters
The problem is that surface-layer AI rarely builds the foundation for decision-layer AI. A tasting notes generator does not produce the structured data you need to run demand forecasting. A social caption tool does not give you the account relationship history you need for intelligent sales prioritisation.
If you build in the wrong order, you end up with a stack of tools that sit in separate silos, each doing a narrow job well, none of them talking to the operational data that would make the decision-layer work possible.
Most producers should start at the surface layer. That is the right call. The mistake is treating surface-layer adoption as the destination rather than the beginning of a data foundation build.
| AI layer | Example tools | Feedback loop | Downside of bad output | Data foundation required |
|---|---|---|---|---|
| Surface layer | Tasting notes, social captions, trade email drafting | Immediate (seconds to minutes) | Low: one off caption or email | None required |
| Decision layer | Account prioritisation, SKU restocking logic, pricing decisions against duty cycles | Months (requires running the plan to verify) | High: wasted sales trips, working capital tied up, missed stockists | Structured operational data: account history, reorder cadence, SKU granularity, seasonal patterns |
The tools are not built for small drinks businesses
Most AI tools available to small businesses are designed for a different shape of organisation. They assume a marketing team, a sales team, a finance team. Each function is separate. Each has its own tools. The data stays in its silo.
A six-person spirits company does not have three teams. It has six people, and on Tuesday one of them is writing social captions while on Thursday the same person is deciding which accounts to visit. The AI tool that helps with the captions is solving one problem. The AI tool that might help with account decisions is solving a different problem for a different shape of organisation. Neither tool knows about the other. Neither tool is interested in the combined data set.
The result is a class of small producers who are underserved not because the tools are bad, but because the tools are built for a world that does not match theirs.
What starting in the right place looks like
Starting in the right place does not mean ignoring marketing AI. It means building in parallel: while the caption generator is saving four hours a week, you are also cleaning your sales data, standardising your account records, and building the data layer that a future decision system can actually read.
At Asterley Bros, we spent three months on data architecture before we built anything. Standardising how we logged every trade account interaction. Building a clean structure for reorder history. Making sure our SKU data was granular enough to be meaningful. None of that work produced anything visible. But it is what made the next layer possible.
The insight layer came second: what does the data actually show, before we ask it to do anything? Not predictions, not recommendations, just structured visibility. Which accounts had not reordered in 90 days. Which SKUs were underperforming against the seasonal pattern. Which regions were showing clustering that the calendar did not explain.
Only after we had visibility we trusted did we start building anything that touched decisions.
The compliance and safety layers are not optional for finance-adjacent decisions
Rob's standing principle on this is worth articulating plainly: AI tools need to be layered to be trustworthy, and the layers are not all the same. Insight, data, compliance, and safety layers must all be integrated before you ask an AI to inform anything finance-adjacent.
For marketing, a bad output costs you the ten minutes it takes to notice and rewrite it. For finance-adjacent decisions, a bad output can mean restocking a SKU three months too early, tying up working capital, and missing the cash cycle.
The compliance and safety layers are not complexity for their own sake. They are the engineering equivalent of asking a second person to check the numbers before you sign the invoice. In a six-person company, you cannot always have that second person available for every decision. The layer is the substitute.
The adoption ladder is not wrong, it is just incomplete
There is nothing wrong with starting on the surface. Most producers should start there. The problem is treating surface-layer adoption as the destination rather than the beginning.
The producers who will get the most from AI in the next three years are not the ones who adopted fastest. They are the ones who used the surface-layer phase to build the data foundation that makes decision-layer work possible, and who built their decision systems with enough layered trust architecture that they are actually willing to act on the output.
That is the gap we are trying to close. Not selling AI to drinks brands. Building the operational infrastructure that makes AI for drinks businesses actually trustworthy.
Why do most AI tools for small drinks brands focus on marketing rather than operations?
The market follows legibility. Marketing AI is easy to evaluate: you can read the output and make a judgement in seconds. Operational AI requires more trust in outputs you cannot immediately verify. Tool builders go where the barrier to sale is lowest, which means the market has oversupplied surface-layer tools and undersupplied decision-layer infrastructure. The gap is commercial, not technical.
What does a "compliance layer" in an AI system actually mean for a small producer?
It means building explicit rules that the system checks its outputs against before surfacing them to you. For a restocking recommendation, a compliance layer might check whether the recommendation falls within your cash-flow parameters for the month, whether it conflicts with existing purchase commitments, and whether it is consistent with the previous month's pattern in a way that makes sense. If any of those checks fail, the output gets flagged rather than acted on. The compliance layer is not about legal compliance specifically. It is about operational coherence.
How long does it take to build a trustworthy decision-layer AI system for a small drinks brand?
In our experience, the data architecture phase takes two to three months if you are starting from a clean but unstructured state. Building and testing the insight layer takes another four to six weeks. Getting to a place where you trust the outputs enough to act on them consistently takes another two or three months of running the system alongside manual processes and comparing the results. Call it six months end to end for a scope-limited first system. The mistake is expecting it to be faster, and cutting the trust-building phase short as a result.
---
Robert Berry is co-founder of Asterley Bros, a London-based premium aperitivo company, and Absolution Labs, an AI automation consultancy for drinks businesses. He makes vermouth by day and builds AI systems in the margins.
More on how we think about AI architecture for small producers at absolutionlabs.com.