When Scotland beat Haiti 1-0 at Boston Stadium on June 13, the city's bars weren't ready. The Sam Adams Boston Taproom ran through nearly 90 kegs as the Tartan Army (50,000 Scottish fans marking their team's first World Cup appearance in 28 years) drank it nearly dry. Emergency deliveries were called in. Hennessy's Bar tripled its St. Patrick's Day sales and sold out of beer altogether. The Phoenix Landing ran out of 20 of its 24 draft taps. The story crossed into mainstream cultural commentary within days.
The beer shortage wasn't a fluke. It was the kind of demand forecasting failure that plays out at enterprise scale whenever operators plan from historical data and normal seasonal benchmarks without accounting for the specific variables that make an event genuinely unprecedented. It also made for a perfect AI experiment.
We gave five leading AI models the same scenario: a mid-sized Boston sports pub, 20 taps, 200-person capacity, normal Saturday sales of 15 kegs, Scotland vs Haiti kicking off at 6pm, 50,000 Scottish fans in the city. The question: how many kegs should you stock?
Then we gave them a follow-up with new information: it's 8pm, Scotland just won, the bar is still packed, and you've already sold through 28 kegs. How many more do you need?
The results reveal something important about how AI models approach demand planning problems — and where even the best reasoning breaks down without the right data inputs.
We ran the same prompt verbatim across ChatGPT, Gemini, Grok, Perplexity, and Claude, all on free tiers. Identical inputs, no customization. The free tier constraint matters: these are the tools a bar manager or operations lead would realistically reach for, not enterprise software. It's the same methodology we used when we asked five LLMs to build a winning March Madness bracket — same models, same principle, different operational problem.
The scenario was designed to test demand planning reasoning rather than general knowledge. It included specific variables (the 28-year absence, the 50,000 traveling fans, the 6pm kickoff) that should meaningfully affect any serious forecast. Models that treated this like a generic "busy Saturday" would produce a very different answer than models that factored in supporter culture, dwell time, and the emotional stakes of a once-in-a-generation fixture.
The spread across models was wider than expected.
ChatGPT produced the most conservative forecast at 28–30 kegs, recommending 30 as the target. Its reasoning was methodical: establish baseline Saturday volume, apply a 50–100% event uplift, add a safety buffer. The approach was structurally sound but the uplift multiplier was cautious. ChatGPT modeled the capacity constraint carefully (a 200-person bar can only serve so many people) and didn't fully account for the cultural dimension of traveling Tartan Army behavior.
Perplexity landed at 35 kegs as its clean middle estimate, with a range of 30–40. It used a 2.5x multiplier on normal Saturday volume, acknowledged the 6pm kickoff as a driver of extended dwell time, and kept its reasoning compact. Perplexity has web search enabled by default, which means it was potentially drawing on live reporting about the Scotland fan situation — but its forecast sat lower than models reasoning purely from training data, which was counterintuitive.
Gemini recommended 38–40 kegs, with detailed operational notes. It introduced what it called a "Scottish consumption factor," estimating 1.5–2 pints per hour per person for traveling Tartan Army fans versus the 1 pint per hour average for sports fans generally. It also flagged the tap profile problem: with 20 taps, blowing through your most popular lagers early leaves you with 200 fans forced to drink whatever's left. The operational specificity was strong.
Grok arrived at 35–45 kegs with 40 as its target, and was the only model to directly cite real-world reports from the actual event in its reasoning — specifically the Sam Adams emergency deliveries and the bars running dry across Boston. That's a meaningful advantage when historical data meets live event data. It modeled the extended drinking window explicitly and flagged the Tartan Army's documented behavior across previous tournaments.
Claude came in highest at 40 kegs minimum, targeting 45 if storage allowed. It worked through the multipliers systematically: once-in-a-generation fixture, 50,000 traveling fans actively looking for pubs, near-ideal 6pm kickoff with a 6–8 hour drinking window. It noted the asymmetry of the risk (unused kegs go back, running dry costs you revenue and customer loyalty) and flagged Scottish ales and Tennent's as priority tap lines given the crowd composition.
The actual answer, based on what happened at Sam Adams alone: closer to 90 kegs over the opening weekend, with multiple emergency deliveries. Every model underestimated. But the gap between ChatGPT's 30-keg recommendation and Claude's 45-keg recommendation is the difference between running out before the final whistle and surviving the night.
The more revealing test was the follow-up. At 8pm, Scotland have just won. The bar is still packed. You've already sold 28 kegs. How many more do you need?
This is the question that separates real-time forecast revision from static demand planning. A pre-match forecast is built on prior event data and assumptions. A real-time update requires integrating new information — actual sales pace, match outcome, crowd behavior — and revising the model on the fly.
ChatGPT estimated 8–12 additional kegs, projecting a final total around 36–40. It modeled the post-match drop-off carefully, estimating peak-game pace at 4.7 kegs per hour and applying a 25–40% drop for post-match drinking. The capacity math was rigorous. What it didn't model was the overflow scenario: when other bars run dry, the crowd doesn't go home. It goes to whichever bar still has beer on tap.
Perplexity gave the most compact response: 10 more kegs as a clean middle estimate, range of 8–12. It noted that beer sales typically slow after the initial emotional peak and that the late-night tail should be meaningful but not match the pre-game surge. The reasoning was defensible but didn't model the Scotland win factor deeply. A 1-0 victory in the first World Cup match in 28 years is not a normal post-match scenario.
Gemini recommended 10–12 additional kegs, bringing the total to 38–40. It added useful operational detail about the tap profile problem (28 kegs already gone likely means your most popular lagers are blown) and flagged the stadium crowd dispersal dynamic: fans leaving the stadium after the match will try to flood into nearby bars, keeping your capacity at maximum even as some existing patrons leave. Its 8pm action plan was the most operationally specific of any model.
Grok went highest in round two at 12–18 additional kegs, targeting 15 more. It directly cited Boston bar reports showing emergency deliveries needed city-wide, modeled the extended hours (Boston bars were licensed to serve until 3am during the tournament), and flagged the potential for restocks going well into the night. It recommended monitoring sales hourly and calling distributors immediately for possible late delivery.
Claude recommended 18–20 additional kegs — the highest estimate and the only one that explicitly modeled the overflow scenario. It framed the Scotland win as a "New Year's Eve trajectory rather than a typical Saturday night tail-off" and identified the risk that other bars running dry would send additional crowds your way. It also noted that the injury prediction signal for a bar running out isn't the current sales pace — it's the combination of sales pace, remaining competitor inventory across the city, and the emotional magnitude of the event. Total projected kegs for the day: 46–48.
The models were not bad. Several produced genuinely useful pre-match forecasts with well-structured reasoning. But the experiment exposes three limitations that matter for anyone thinking seriously about AI-driven demand forecasting in an enterprise context.
First: all five models were reasoning from static inputs. None had access to the real-time signals that determine what actually happens to demand during a live event: distributor stock levels across the city, foot traffic data from fan tracking apps, social media sentiment about which bars were running low, or comparable event data from previous Tartan Army tournaments. Without those inputs, even sophisticated reasoning produces forecasts that are structurally sound but operationally incomplete. Traditional forecasting methods face the same constraint — the difference is that production AI systems can ingest those signals continuously and revise in real time using predictive analytics that updates against actual outcomes.
Second: the models handled real-time revision very differently. Round two — the dynamic update — is where the quality gap opened up. ChatGPT and Perplexity modeled a standard post-match drop-off. Claude and Grok modeled the overflow and celebration dynamics explicitly. The difference wasn't intelligence; it was whether the model was reasoning about this specific event or applying a general template for post-match trade. Effective AI-driven demand forecasting requires models trained on domain-specific data — seasonal patterns, event type, supporter behavior, venue size — not general-purpose reasoning applied to operational scenarios ad hoc.
Third: the asymmetry of errors was understood by some models and missed by others. Running out of beer on Scotland's first World Cup night in 28 years is not the same cost as having five unused kegs at close. Claude and Grok both modeled this explicitly: excess stock has low carrying costs in this context, running dry has high reputational and revenue costs. ChatGPT and Perplexity treated the optimization as roughly symmetric. In real demand planning, getting the cost asymmetry right is often more important than getting the central estimate right. That applies whether you're managing keg inventory for a Boston pub or optimizing inventory levels for a retail chain across hundreds of locations.
The beer shortage is the proof. Every bar that ran dry that night had made a planning error — but the more expensive error was underestimating by 50% rather than over-ordering by 15%. The models that recommended 40–45 kegs would have kept the taps running. The models that recommended 28–30 would have been calling for emergency restocks at halftime.
The experiment is illustrative precisely because a general-purpose AI model and a production AI demand forecasting system are solving different problems. The models in this experiment were given a scenario and asked to reason. A production system is continuously ingesting data — market trends, consumer behavior patterns, IoT sensor data from supply chains, and external signals including economic indicators — and running predictive analytics against actual outcomes as they happen.
The Scotland scenario had variables that a well-trained forecasting model would have flagged automatically: a first World Cup appearance in a generation produces a different fan travel pattern than a routine qualification. A 6pm weekend kickoff in a city with 50,000 traveling supporters produces a different demand profile than the same fixture at noon. The emotional stakes of a win versus a draw change post-match dwell time in ways that appear predictably in models built on comparable events — from previous Tartan Army tournaments to major sporting events in similar cities — with seasonality patterns capturing how celebration behavior differs from routine weekend trade.
Industries that face genuinely volatile demand have already built this infrastructure. Healthcare uses AI-driven demand forecasting to manage resource allocation across patient volumes that swing dramatically with outbreak patterns. Automotive manufacturers use it to align production schedules and procurement with demand signals that arrive months before orders are placed. Both rely on domain-specific model training, not general-purpose reasoning applied to operational scenarios. The architecture is the same whether you're managing inventory for a hospital pharmacy or a Boston sports bar on the most significant night in Scottish football history.
What the AI models got right was the structure of the reasoning. What they lacked was the data infrastructure to run that reasoning against real-world signals in real time. That gap between good reasoning and production-ready demand forecasting and inventory management is exactly what enterprise AI systems are built to close. Inventory decisions made at 8pm shouldn't rely on a bar manager's instinct or a general-purpose chatbot's best guess. They should be informed by a system that has seen this pattern before, knows what generative AI capabilities add to scenario modeling, and uses data analytics to tell you what happens next with operational specificity.
Demand forecasting at the operational level requires automation of the data ingestion layer — continuous feeds from point-of-sale systems, distributor networks, and external signals — combined with models that revise in real time as new information arrives. The beer shortage wasn't inevitable. It was an inventory problem in reverse: the same forecasting failure that leaves warehouses full of unsold product, just pointed the other direction. Closing that gap means treating the forecast not as a plan but as a living model that improves with every data point. The 2026 AI demand forecasting playbook for supply chain teams covers what that looks like in practice.
Invisible builds AI demand forecasting systems that help enterprise operators close the gap between good reasoning and production-ready decisions. See what that looks like at invisibletech.ai/solutions/forecasting, or get started.
AI demand forecasting uses machine learning to predict future demand by analyzing historical sales data, market trends, consumer behavior, and external signals including weather patterns and economic indicators. Unlike traditional forecasting methods that rely on static models, AI-driven systems update continuously as new data arrives, improving forecast accuracy over time and reducing both supply shortfalls and excess inventory across inventory management cycles.
AI demand forecasting consistently outperforms traditional methods on complex, high-variability scenarios. Traditional approaches apply fixed rules to past data and struggle with unusual events, new product launches, or rapid shifts in consumer behavior. Neural networks identify non-linear patterns across multiple data sources simultaneously, and models built on domain-specific training data improve predictive accuracy particularly where demand is driven by factors that don't appear in a business's own sales records.
The foundation is historical sales data, replenishment records, pricing history, and procurement data — but data quality matters more than volume. Poor quality compounds quickly in machine learning algorithms. Advanced analytics built on comparable event patterns is especially valuable where demand spikes are predictable but irregular, such as healthcare resource allocation or automotive manufacturing where overproduction ties up capital for months.
Demand sensing uses real-time signals — point-of-sale data, distributor stock levels, social media activity, foot traffic data — to update near-term forecasts continuously. Traditional demand forecasting produces a plan and holds it until the next review cycle. Demand sensing uses automation to treat the forecast as a living model, which is critical for operational efficiency in fast-moving situations where conditions change faster than planning cycles can track.
The most common errors come from over-reliance on past data without accounting for structural changes, poor data quality in training records, and failure to model the asymmetric cost of running dry versus carrying excess inventory. Using generative AI for scenario planning helps, but the data infrastructure has to support real-time revision. Forecasting failures at major events typically result from applying normal baselines to genuinely unprecedented demand.
AI demand forecasting reduces stockouts and overstocking by generating more accurate demand signals than traditional methods produce. Better forecast accuracy means procurement decisions and resource allocation choices are made against realistic projections rather than averages. The downstream effects include fewer emergency restocks, better replenishment timing, the ability to optimize inventory across locations, and tighter alignment between production schedules and actual demand — whether you're managing pharmaceutical supply chains, parts inventory, or food and beverage for a major sporting event.
.avif)