Artificial intelligence (AI) is receiving enormous attention, but for many small wineries, the subject can still feel remote. We hear about global companies using AI to forecast demand, optimize prices and personalize promotions. It is less obvious how a small winery with limited staff, limited data and a relatively small number of annual transactions can use the same ideas.
I recently decided to find out.
My goal was not to allow a computer to set wine prices automatically. I wanted to see whether AI could help me better understand three practical questions at Northern Cross Vineyard:
- What should each wine be priced at?
- Which promotions are likely to increase profit rather than simply increase volume?
- Which customers or visitor groups should receive each offer?
The experiment showed me that AI can be useful to a small winery—but only when it is treated as a decision-support tool, not as an all-knowing answer machine.
Starting with data I already had
I began by exporting three years of sales records from Square. The file covered 2022 through 2024 and included individual transactions, products sold, quantities, prices, discounts, tasting purchases and customer identifiers.
I uploaded the CSV file into ChatGPT and asked it to analyze the data from a pricing and promotional perspective.
This required no special software development. I did not build a complex machine-learning model or hire a data scientist. I simply exported the information I already had, uploaded it and described the business questions I wanted to answer.
The first lesson was that even ordinary point-of-sale data can reveal useful patterns when it is organized and examined systematically.
The dataset included several years of item-level sales, transaction, and customer information. This provided enough detail to identify patterns in product mix, purchasing behavior, discounts, and repeat visits, while also revealing gaps in how consistently customers were identified across transactions.
Those figures immediately raised several questions. Were most customers actually visiting only once, or were repeat visitors not consistently being identified in Square? Were we doing enough to encourage a second visit? Were our promotions rewarding incremental purchases, or simply discounting bottles customers would have purchased anyway?
AI did not eliminate the need for judgment. It helped identify where judgment should be focused.
What the analysis suggested about pricing
One of the clearest findings was that most of our wines had remained at nearly the same list prices for three years, with core reds generally priced at $26 and several whites at $24. That stability followed a significant earlier adjustment. We had originally priced our white wines at $17 and our reds at $19, but after reviewing pricing data collected through the New York Wine Classic, I recognized that our wines were underpriced relative to the market. That information gave me the confidence to raise prices to their current levels.
Because those prices had changed very little, the system could not reliably calculate price elasticity—the relationship between a change in price and the resulting change in demand. A model cannot determine how customers respond to different prices when there are few different prices in the historical data.
That was an important limitation.
My next step is to add all of the sales data and ask for an elasticity analysis by comparing the $19/$17 bottle sales to the $26/$24 bottle sales.
Based on sales volumes, realized prices and the lack of significant discounting, the analysis suggested testing modest increases rather than making dramatic changes. It proposed moving the principal red wines from $26 to approximately $28, increasing the whites from $24 to approximately $25 and creating greater separation between regular Marquette and Marquette Reserve.
The Reserve had often been priced at the same level as the standard wine. From a pricing perspective, that weakened the meaning of “Reserve.” The analysis suggested testing a price of approximately $29 for the Reserve to better communicate differentiation.
These were not presented as mathematically perfect prices. They were proposed as controlled experiments.
That distinction is critical. A small winery does not necessarily need an AI system that declares the “correct” price. It may be more valuable to have a system that recommends a reasonable test and explains what should be measured.
For example, after a price increase, the winery should track:
- Bottles purchased per visitor
- Tasting-to-purchase conversion
- Average transaction value
- Contribution profit per transaction
- Customer objections
- Changes in the product mix
A modest decline in bottle volume may be acceptable if revenue and contribution profit increase.
Looking beyond blanket discounts
The promotional analysis produced an equally useful insight.
The typical bottle transaction in the data included approximately 2.2 bottles. Looking at sales percentages, 53% were one-bottle transactions 35% were two-bottle transactions, but only 12% of transactions were contain three bottles.
That created a logical promotional objective: encourage customers who were already purchasing one or two bottles to purchase a third.
A conventional approach might be to offer 10 percent off. However, a blanket percentage discount reduces the price of every qualifying bottle—including bottles the customer may already have intended to purchase.
The AI analysis recommended a different test:
Apply the tasting fee as a credit when a visitor purchases three bottles.
This type of offer could be more profitable because it rewards an incremental behavior. Instead of discounting the first and second bottles, it encourages the customer to add another bottle to the transaction.
The data showed that tasting-room visitors frequently purchased wine. Approximately 87 percent of transactions that included a tasting also included at least one bottle. That suggests the tasting experience already converts well. The opportunity may therefore be to increase the number of bottles purchased rather than simply increase the percentage of tasters who buy something.
The system also identified frequently purchased combinations, including Frontenac with La Crescent, Battenkill Red with La Crosse and Battenkill Red with Frontenac.
Those patterns could support fixed-price mixed bundles. A three-bottle selection might carry an effective discount of only 4 or 5 percent rather than 10 percent. The customer receives a visible benefit, but the winery protects more of its realized price.
Again, the objective is not to maximize the number of bottles leaving the tasting room. The objective is to maximize profitable, sustainable sales.
Not every customer should receive the same offer
The customer data also suggested that promotions should be more selective.
More than 300 identified customers appeared only once in the three-year file. Some may have returned without being properly identified, but the number still indicates an opportunity to improve customer tracking and follow-up.
The analysis proposed several basic customer groups.
First-time tasting visitors buying one or two bottles could receive the three-bottle tasting-credit offer at the point of sale.
High-value first-time customers—those buying three or more bottles or spending at least $75—might receive a return-visit invitation instead of an immediate discount. A complimentary future tasting, access to a limited release or an invitation to a vineyard event could encourage a second visit without reducing the margin on the first purchase.
Customers who had already visited twice could be candidates for a preferred-customer program or wine club. Their behavior demonstrates more commitment than that of a first-time visitor, so the objective should be to formalize the relationship.
The most loyal customers may not need the deepest discounts at all. They may value access, recognition, vineyard events, reserve tastings and early releases more than a coupon.
This is one of the most important pricing lessons AI can help reinforce: a promotion should be connected to the behavior the business is trying to change.
The information that was missing
The experiment also exposed several weaknesses in the data.
The Square export did not include reliable variable-cost data for each bottle. Without the cost of the wine, bottle, closure, label, packaging, credit-card processing, and other incremental expenses, the system could not calculate true contribution profit. The next step will be to add those cost inputs and use AI to evaluate profitability by wine, price point, and promotion.
It also did not contain enough variation in historical prices to estimate dependable price elasticity.
Customer identification was inconsistent, and the file did not clearly distinguish tourists, local residents, event visitors, wine-club members or referrals from other wineries.
These limitations do not make the exercise unsuccessful. They show what data should be collected next.
The next version of the analysis should include:
- Estimated variable cost by wine
- Inventory by vintage
- Visitor ZIP code
- New versus returning visitor
- Promotion or offer received
- Source of the visit
- Event attendance
- Wine-club status
- Weather and seasonal conditions
- Whether a customer visited other wineries in the region
For the broader Upper Hudson wine region, a wine-trail passport could become an especially valuable data source. With appropriate permission and privacy protections, the program could help measure winery visitation patterns, route preferences, repeat visits and the effectiveness of regional promotions.
What I learned about AI
AI was useful in several ways:
My biggest conclusion is that a winery does not need enormous amounts of data to begin using AI. It does, however, need clearly defined business questions and a willingness to examine the quality of its information.
- It organized a large transaction file quickly.
- It identified product and customer patterns that would have taken much longer to calculate manually.
- It proposed testable pricing and promotional ideas.
- It explained the assumptions behind its recommendations.
- It identified missing data that limited the analysis.
What it did not do was replace business judgment.
AI did not know the reputation of each wine, the inventory available, the vintage quality, the customer experience or the long-term positioning of the Northern Cross brand unless I supplied that context.
The technology is most valuable when the winery owner combines three things: the data, the model’s analysis and practical knowledge of the business.
For wineries that have not yet experimented with AI, the best first step may be surprisingly simple: export a year or two of sales data, remove unnecessary personal information and ask a focused question.
Do not begin by asking, “What can AI do for my winery?”
Begin with a specific decision:
- Which products are commonly purchased together?
- Which promotions are reducing price without increasing purchase quantity?
- Which first-time customers should receive a return offer?
- Which wines may support a controlled price test?
- Which tasting-room behaviors are associated with larger purchases?
The answers may not be perfect, but they can lead to better questions, better experiments and ultimately better decisions.
For a small winery, that may be the most practical definition of artificial intelligence: not a machine that takes control of pricing, but a tool that helps the owner see the business more clearly.
