How Torfs increased sales 50% with AI-driven stock reallocation

Belgian footwear retailer Torfs has reported a 50 per cent increase in sales for inventory that was reallocated between stores using artificial intelligence, according to the company. The result comes from an AI-driven allocation system that analyses store-level demand signals and moves stock multiple times per week, transforming shoes that were stagnating in one location into fast sellers at another. For an industry grappling with excess inventory and rising markdown pressure, the case illustrates how granular, algorithm-driven distribution can unlock revenue from existing stock without additional purchasing.

The uplift matters because footwear assortments are notoriously complex. A single style may be carried in eight to twelve sizes, and consumer preferences vary sharply by region, store format and demographic. Traditional allocation relies on historical averages and manual replenishment, often leaving popular sizes out of stock in high-demand stores while they gather dust elsewhere. AI systems promise to solve this mismatch by processing real-time sales velocity, local customer behaviour and transfer costs to recommend moves that maximise sell-through.

Torfs, a family-owned chain with more than 80 stores across Belgium and Luxembourg, began piloting AI allocation in 2023 and scaled the system across its network in 2024. The retailer now reallocates inventory several times per week, using algorithms that weigh store performance, stock levels and logistics costs. The company has not disclosed the total revenue impact or return on investment, but the 50 per cent sales gain on reallocated items suggests meaningful potential for retailers willing to invest in data infrastructure and change management.

What exactly did Torfs do with AI allocation?

Torfs implemented an AI-powered stock allocation system that analyses sales data at the store and SKU level to identify inventory that is underperforming in one location but likely to sell faster elsewhere. The system generates reallocation recommendations multiple times per week, prompting the retailer to move shoes between stores based on predicted demand rather than waiting for stock to reach markdown thresholds. The company began testing the technology in 2023 and expanded it across its full network in 2024, according to industry reports.

The algorithm considers several variables, including recent sales velocity for each style and size at every store, local customer preferences, current stock levels, and the cost of transferring inventory between locations. When the model identifies a mismatch, such as a popular size sitting unsold in a low-traffic store while a high-traffic location is out of stock, it recommends a transfer. Torfs staff review the recommendations and execute the moves, typically using the retailer's existing logistics network.

The reported 50 per cent sales increase applies specifically to the items that were reallocated, not to total company sales. In other words, shoes that were moved from one store to another sold 50 per cent more units after the transfer than they would have if left in the original location, based on the company's internal analysis. Torfs has not published a controlled study or disclosed the baseline sales rate, so the figure represents the retailer's own attribution rather than an independent measurement.

How often does the system reallocate inventory?

Torfs reallocates stock multiple times per week, a sharp departure from the traditional monthly or seasonal replenishment cycles used by many footwear retailers. The high frequency allows the system to respond quickly to emerging trends, local events and shifts in customer behaviour. For example, if a particular sneaker style begins selling rapidly in urban stores but slowly in suburban locations, the algorithm can recommend transfers within days rather than waiting weeks for a manual review.

Frequent reallocation requires robust logistics and clean data. Torfs must track inventory in real time across all stores, process sales transactions immediately, and operate a distribution system capable of handling small, frequent shipments. The retailer's relatively concentrated geography in Belgium and Luxembourg makes this operationally feasible, as transfer times between stores are short and transport costs relatively low. Retailers operating across larger territories or in markets with fragmented logistics infrastructure may find higher reallocation frequencies more challenging.

What impact does AI allocation have on sell-through rates?

Higher sell-through rates are the core promise of AI allocation systems. By moving inventory to stores where demand is strongest, retailers can sell more units at full price and reduce the volume of stock that eventually requires markdown. Torfs has not disclosed its overall sell-through improvement or markdown reduction, but the 50 per cent sales uplift for reallocated items implies that a meaningful share of inventory was previously misallocated. If even 10 to 20 per cent of stock can be redirected to higher-performing locations, the cumulative effect on margin can be substantial.

Sell-through improvements depend on the accuracy of the demand forecast and the speed of execution. An algorithm that correctly predicts which store will sell a given size and style can generate significant gains, but errors can be costly. Moving inventory to the wrong location incurs transfer costs, ties up stock during transit, and may leave the original store understocked if demand suddenly picks up. Torfs mitigates this risk by reallocating frequently, so mistakes can be corrected in subsequent cycles, and by focusing on styles and sizes with clear demand signals rather than slow-moving outliers.

Industry data on AI allocation outcomes remain sparse. Most retailers treat their algorithms and results as proprietary, and few publish controlled trials. Torfs' disclosure of a 50 per cent uplift is unusually transparent, but the figure lacks context on baseline performance, the share of inventory reallocated, or the net impact after accounting for transfer costs and mis-allocations. Decision makers evaluating AI allocation should therefore request pilot data, establish clear success metrics and compare AI recommendations against existing manual processes before committing to full-scale deployment.

Which algorithm types are used for stock reallocation?

AI allocation systems typically rely on machine learning models that forecast demand at a granular level, often using regression, time series analysis or neural networks. The models ingest historical sales data, inventory positions, pricing, promotional activity, weather, local events and other variables to predict how many units of each SKU will sell at each store over a defined period. The system then compares predicted demand to current stock levels and recommends transfers to balance supply and demand across the network.

Some platforms use reinforcement learning, in which the algorithm iteratively tests different allocation strategies and learns from the outcomes. This approach can adapt to changing conditions without manual retraining, but it requires substantial transaction volume to generate reliable feedback. Other systems combine predictive models with optimisation algorithms that calculate the most cost-effective set of transfers, taking into account shipping costs, handling time and store capacity constraints.

Torfs has not disclosed the specific algorithm or technology vendor behind its system. The retailer may have developed the solution in-house, partnered with a specialised supply chain software provider, or worked with a consultancy to build a custom model. Regardless of the technical details, the success of any allocation system depends more on data quality, organisational discipline and change management than on the choice of algorithm. Retailers with fragmented inventory data, inconsistent SKU coding or siloed IT systems will struggle to achieve meaningful results, even with sophisticated models.

What return on investment can retailers expect?

Return on investment for AI allocation depends on several factors, including the cost of the technology, the current level of inventory inefficiency, the retailer's margin structure and the operational cost of moving stock. Torfs has not published ROI figures, but a 50 per cent sales uplift on reallocated inventory suggests the system is delivering value, assuming transfer costs and technology expenses are a small fraction of the incremental revenue. For retailers with high margins and low transfer costs, the payback period may be less than one year. For those operating on thin margins or in geographies with expensive logistics, the business case is more marginal.

Costs include software licensing or development, data infrastructure, integration with existing ERP and warehouse management systems, staff training and ongoing operational expenses such as additional shipments and handling. Benefits include higher full-price sales, lower markdowns, reduced need for clearance inventory and potentially lower overall stock holdings if the retailer can serve the same demand with better-distributed inventory. Some retailers also report improved customer satisfaction, as popular items are more likely to be in stock at the right location.

A 2025 report by McKinsey on retail supply chain technology estimated that AI-driven allocation and replenishment could improve gross margin by 100 to 200 basis points for fashion retailers with complex assortments and multi-channel distribution. The consultancy noted that benefits are highest for categories with short selling seasons, high SKU proliferation and significant regional variation in demand, all of which apply to footwear. However, McKinsey also cautioned that realising these gains requires disciplined execution, clean data and a willingness to override manual decision-making, which can meet organisational resistance.

What are the risks and limitations of AI allocation?

AI allocation systems carry several risks. Forecast errors can lead to costly mis-allocations, moving inventory to stores where it will not sell and leaving high-demand locations understocked. This risk is especially acute for new styles with limited sales history, seasonal items with short selling windows and SKUs affected by sudden external events such as weather or viral social media trends. Algorithms trained on historical data may also perpetuate past biases, such as under-serving certain store formats or regions, unless explicitly designed to test new allocation strategies.

Operational complexity is another challenge. Frequent reallocations require reliable logistics, real-time inventory visibility and coordination between store staff, distribution centres and the IT team. In practice, many retailers struggle with data quality issues, such as inaccurate stock counts, delayed transaction updates or inconsistent SKU coding across channels. If the data feeding the algorithm is unreliable, the recommendations will be flawed, and trust in the system will erode. Retailers must invest in process discipline and technology infrastructure before deploying AI allocation at scale.

There is also a risk of over-optimisation. An algorithm that maximises short-term sales may recommend reallocating all high-demand stock to a few flagship stores, leaving smaller locations perpetually understocked and eroding the brand's presence in certain markets. This can harm long-term customer relationships and limit the retailer's ability to serve diverse segments. To avoid this, allocation systems should include guardrails such as minimum service levels for all stores, constraints on the share of inventory that can be moved in a single cycle and regular reviews of the strategic implications of the algorithm's recommendations.

Finally, the Torfs case represents a single retailer's internal assessment rather than a peer-reviewed study or independent audit. The 50 per cent sales uplift is a company claim, and the methodology for calculating the figure has not been disclosed. It is unclear whether the comparison accounts for seasonality, promotional activity, or other confounding factors, and whether the result is sustainable over multiple seasons. Decision makers should treat vendor case studies and self-reported metrics with appropriate scepticism and insist on pilot programmes with rigorous measurement before committing to large-scale investment.

How does footwear differ from apparel for AI allocation?

Footwear presents distinct challenges and opportunities for AI allocation compared to apparel. Shoes are sold in discrete sizes with limited flexibility, unlike clothing where a customer might accept a slightly different fit. A store that stocks only size 38 and 42 of a popular sneaker cannot serve customers who need size 40, creating a binary in-stock or out-of-stock outcome. This makes size-level allocation critical and creates a larger potential benefit from moving inventory to match local demand patterns.

Footwear assortments are also more fragmented. A single style may be offered in eight to fourteen sizes, and each size represents a separate SKU with independent demand. Multiplying styles by colours and sizes can generate thousands of SKUs, even for a focused collection. This proliferation makes manual allocation impractical and increases the value of algorithmic decision-making. However, it also raises the data burden, as the system must forecast demand for each SKU at each store, requiring substantial transaction history and sophisticated models.

On the other hand, footwear has less pronounced fashion risk than apparel for certain categories. Classic sneaker styles, work boots and formal shoes can sell for multiple seasons, giving algorithms more historical data to train on and reducing the stakes of a mis-allocation. Trendy or seasonal styles, by contrast, behave more like fashion apparel, with short life cycles and limited sales history. Retailers operating across both stable and volatile categories may need different allocation strategies for each, with more conservative rules for new or risky items.

Frequently asked questions

What should fashion companies do now?

1. Audit your inventory data quality before investing in AI allocation. Ensure SKU codes, stock counts and sales transactions are accurate and updated in real time across all stores and channels.
2. Run a controlled pilot programme in a subset of stores or for a limited product category. Measure sales, sell-through and margin impact against a control group using your existing allocation process.
3. Map your logistics capabilities to determine how frequently you can reallocate inventory. Assess transfer costs, lead times and handling capacity to set realistic reallocation intervals.
4. Define clear success metrics beyond sales uplift, including full-price sell-through, markdown rate, gross margin, inventory turn and customer satisfaction. Track these metrics before and after AI deployment.
5. Establish governance and guardrails for the algorithm, such as minimum stock levels per store, limits on the share of inventory that can be moved in one cycle, and manual override protocols for strategic decisions.
6. Invest in change management to ensure store staff, buyers and allocation teams understand the system, trust the recommendations and provide feedback. Resistance to algorithmic decision-making is a common barrier to ROI.
7. Review results every season and refine the model based on forecast accuracy, operational challenges and business outcomes. AI allocation is not a one-time implementation but an ongoing process of learning and improvement.