Waste and forecasting inaccuracy are expensive, persistent problems for Irish FMCG businesses. AI can reduce both — but only with the right data foundation and realistic expectations.
Waste and forecasting inaccuracy are two of the most persistent margin problems in Irish FMCG. They're connected: poor forecasting drives over-production and over-ordering, which drives waste. Improving one typically improves the other. AI has a genuine role in this — but the path to that improvement is less straightforward than vendor demonstrations suggest.
The waste ROI case
For Irish FMCG businesses producing or distributing perishable goods, waste is rarely below 3% of revenue. For businesses with shorter shelf lives, seasonal demand volatility, or promotional complexity, it's frequently higher. At 4% waste on a €15 million turnover, that's €600,000 leaving the business annually in product that's produced, stored, transported, and discarded.
Even a 20% reduction in waste — which is a conservative target for businesses moving from manual to AI-assisted forecasting — represents €120,000 in annual saving. The ROI case for investing in forecasting improvement is almost always compelling. The question is what it takes to get there.
What AI actually improves in forecasting
Manual forecasting in FMCG — whether done in spreadsheets, ERP tools, or someone's head — tends to underperform on specific, predictable dimensions. It handles seasonality reasonably well when it's regular and historical. It handles promotional uplifts inconsistently, often by applying a blanket percentage uplift that doesn't reflect actual promotional performance variation by SKU, by retailer, or by time of year.
It handles weather correlation poorly, despite the fact that weather effects on FMCG demand in Ireland — particularly in food service-adjacent categories — are well documented and quantifiable.
It handles new product introductions badly, because there's no historical data to draw on and analogue selection is often done informally.
AI forecasting tools improve performance on all four of these dimensions, because they can process multiple variables simultaneously and identify patterns in historical data that a human analyst would not spot at scale. The improvement is typically an increase in forecast accuracy of 15% to 30%, which translates directly into lower safety stock requirements, lower write-offs, and improved service levels.
What data is needed
The AI forecasting improvement above is contingent on data quality. The minimum viable data set for AI-assisted demand forecasting in FMCG includes at least 24 months of clean historical sales data at SKU level, with promotional flags and channel separation. Weather data can be overlaid from third-party sources. Promotional calendars need to be structured consistently.
The common failure mode is that the historical data exists but isn't clean. Sales data across multiple channels is reconciled manually and inconsistently. Promotional periods aren't flagged in the system. SKU codes have changed over the period, breaking historical continuity. Write-off data isn't consistently captured at SKU level.
Fixing data quality is not glamorous, but it is the prerequisite. Two to four weeks of data preparation work before an AI forecasting implementation will deliver better results than a faster deployment on poor data.
Inventory AI
Beyond demand forecasting, AI can improve inventory management at the operational level: identifying slow-moving stock before it becomes a write-off problem, flagging SKUs where the reorder point is misaligned with actual usage patterns, and alerting buyers to emerging demand shifts in time to adjust ordering.
These are incremental improvements rather than transformations, but they operate continuously. A system that catches a slow-moving inventory problem two weeks earlier, across hundreds of SKUs, saves more cumulatively than any single intervention.
Common implementation mistakes
The most common mistake is deploying AI forecasting without fixing the data first. The second is underestimating the change management requirement: demand planners and buyers need to trust the AI output enough to act on it. Trust is built through transparency — showing how the model arrived at a recommendation — and through a transition period where AI recommendations are reviewed against human judgment before being acted on.
The third mistake is measuring AI forecasting performance against a benchmark that hasn't been clearly defined. Before deployment, establish what the current forecast accuracy actually is. It's often surprisingly hard to calculate, because the definition of "accuracy" varies across teams and the measurement isn't systematic. Without a baseline, you can't demonstrate improvement — which undermines the business case and the internal credibility of the project.
A practical starting point
For Irish FMCG businesses starting from scratch, the most tractable starting point is a focused pilot on a defined product category with clean historical data. Run the AI forecast alongside the current method for one full seasonal cycle. Compare accuracy. Calculate the waste reduction and service level improvement. Use the results to build the case for broader deployment.
This takes longer than a full deployment. It is also significantly more likely to succeed, because it builds the organisational confidence and data discipline that makes full deployment viable.