Detect Franchisee Off-Network Sourcing with ML

The procurement team's efforts are pivotal in ensuring consistent quality across restaurants and helping locations achieve high operating margins. However, some franchisees undermine the effort to capture what appears to be a small, immediate saving on individual purchases. The damage from off-contract procurement is multifaceted:

• Revenue Loss: Purchases that bypass the approved supply chain can reduce procurement-related revenue and undermine negotiated commercial agreements.
• Reputational Damage: Lower-quality or inconsistent ingredients can affect the customer experience and, ultimately, the brand.
• Contractual Volume Risk: Reduced purchases through approved distributors can make it harder for the network to meet contractual volume commitments and maintain negotiated discounts.
• Compliance Risk: Off-system purchases can make it difficult to accurately track items such as alcoholic beverages, packaging, plastics, and other regulated or reportable products.
• Skewed Operational Metrics: Network-level KPIs, such as operating margin, waste ratios, and inventory turnover, are skewed, affecting reporting and strategic business focus.

Supply chain teams have built a series of metrics, including same-store comparables, the ratio of procurement cost to total sales, purchasing patterns across related items, and others, to flag locations that may be sourcing outside the approved network. Unfortunately, over the years, franchisees have learned these proprietary signals, allowing some to procure from unauthorized distributors without tripping the wires.

Nevertheless, in the last decade, companies have transformed their data strategy. Franchisors now hold years of granular sales and procurement data, including item-level sales, detailed product purchases, quantities, prices, and location-level operating metrics. This depth of data creates an opportunity to employ a broader array of analytical and Machine Learning (ML) techniques.

Deviation Signals

Sophisticated procurement teams have developed a series of signals and corresponding thresholds over the years. However, all rule-based signals have an inherent flaw: they are learnable, and once known, someone can find ways to stay under the radar. A key metric is the procurement ratio of related ingredients, such as pizza dough and sauce, but franchisees learned to off-channel source both in proportion; similarly, with the procurement ratio of items with steep discounts provided by partner distributors and those with low discounts, franchisees began to either follow proportionate off-channel sourcing or choose to forgo a few high-reward items.

Some advanced metrics combine sales and procurement data, such as the ratio of packaging materials, plastic cutlery, and take-out containers purchased to the number of delivery and take-out sales channel orders, but franchisees began registering sales under different channels. Comparing procurement and sales volume for items sold directly to consumers, such as beverages, is susceptible to systematic suppression of both sales and procurement, which doubles revenue loss by losing both royalties and procurement discounts. Similarly, same-store comparables can be systematically suppressed in sales or growth over years.

A range of innovative signals has emerged, but the core challenge remains: for any rule you write, someone can stay under it.

Mapping

Years of sales and procurement data allow a shift from using proxy signals to directly estimating deviations between procured volume and the requirements needed to meet reported sales for individual items. A model can be developed by mapping products sold to ingredients required using recipes. The variance between estimated ingredient requirements and actual procurement can flag franchisees potentially involved in off-channel sourcing. However, generating the model manually has several challenges, such as:

• Interdependent Relationship: Multiple ingredients are used to prepare a food item, and those same ingredients are used to prepare various products, creating complexity for manual mapping.
• Wastage and Pilferage: Raw ingredient waste varies by location, depending on SOP adherence and demand uncertainty.
• Ingredient Shelf Life: Each ingredient has a different shelf life, so procurement frequency varies, making it challenging to build these rules.
• Regular Updates: As new products are introduced, existing ones are replaced, or the distribution partner changes, mapping must be updated frequently.
• Modifications: Customers often request modifications to their food orders, most of which have limited overall impact, but a few can skew raw-ingredient requirements, and these are often geography-specific.

Predictive Methods

A supervised ML model can instead learn this relationship autonomously from historical sales and procurement data. Supervised ML architectures can capture complex relationships even amid uncertainty and noise in the data. The ability to handle the noise is essential, as the procurement data does not always perfectly align with sales for every location because of different levels of wastage, a few locations underbuying as a result of unauthorized sourcing, a handful of locations overbuying as they may be diverting deeply discounted ingredients to their other restaurants or reselling locally at a margin, and glitches in the record-keeping systems of distributors. Some ML architectures are better suited given the complexity of the relationship, the need for continuous output, data availability, and other factors. Random Forest and XGBoost are likely candidates for the initial modeling, while a neural network can be adapted for higher accuracy when ample historical data is available. During training, the model learns iteratively, predicting required procurement based on sales data and adjusting based on the difference between predicted and actual procurement. The model learns how custom modifications affect ingredient requirements.

Distribution Network Evolution

The supply chain evolves as the franchisee network expands into new territory and builds new distributor relationships; thus, the range of items supplied through networked distributors changes with territory and time. Therefore, raw ingredient procurement data can vary over time and across regions. Moreover, specific products may be launched or modified in certain regions to match local taste preferences. Thus, key geographic and temporal information is essential for training an accurate, robust model. However, it is critical to ensure this information does not identify individual restaurants, as that could undermine the entire effort. The historical data the model is trained on will show lower procurement volumes for locations procuring from off-channel distributors, and specific restaurant identifiers will teach the model to predict lower procurement requirements for those stores for similar sales. Preventing this kind of information leakage requires analyzing each input feature's influence and having business and domain experts evaluate model performance before deployment.

Data Frequency

Another challenge is data frequency: while sales occur throughout the restaurant's operating hours, procurement typically happens in batches weekly, biweekly, monthly, or even quarterly for a few long-shelf-life items. Moreover, delivery days and dates are inconsistent, especially across locations; for example, monthly delivery can occur on the 4th at one location and the 7th at another, and next month it can be the 2nd and 8th, respectively. Also, not all procured items are consumed in the same cycle, and restaurants can order small-batch refills if sales exceed projections.

The models can handle differences in delivery dates across time intervals by aggregating sales to the same procurement intervals and using the days between deliveries as an input parameter. The small stopgap batches are summed with other procurements in the same time interval. Two models are trained: one for raw ingredients, aligned with the standard procurement cycle, and another for long-shelf-life items such as packaging containers, spice mixes, and sauces, using a longer cycle, typically quarterly or half-yearly. Variation in waste across locations and differences in the amount of ingredients carried from the previous cycle versus what remains at the end of the current cycle add noise to the data, which various ML architectures can handle effectively.

Augmented Baseline

The models generally start from a randomly seeded baseline and learn from the data. However, model performance can be improved and data requirements reduced by leveraging recipe-based mappings. The model can be seeded with the manually developed mapping discussed in the previous section, rather than random values, allowing it to learn quickly from less data.

Limitations

ML models are not precise; the difference between predicted procurement requirements and actual procurement will never be precisely zero. Thresholds for individual ingredients or across ingredients to flag locations are decided based on the model's accuracy and business insights. Franchisees can exploit these thresholds by procuring small volumes from unauthorized distributors. Moreover, because the model is trained on historical data, it includes data from stores that have procured from unauthorized distributors, biasing the model toward lower requirements. However, the leeway for off-channel procurement is small, and as locations involved in off-network procurement are identified and amend their purchasing policy, data quality improves, increasing measured model accuracy and reducing bias.

Systematic Sales and Procurement Suppression

A restaurant may systematically suppress both sales and procurement, reducing both royalty and procurement revenue. Threshold methods and the ML-based approach discussed earlier cannot identify systematic suppression.

An alternative approach is to predict sales using restaurant information, local demographic details, traffic, competition details, sales through third-party apps, and macroeconomic factors. This approach can be implemented in two ways: predict sales and then feed that information into the above model, or build a model that predicts procurement requirements directly. Training a model to predict procurement requirements directly can also be used when sales data is unavailable or not captured at the required item-level granularity.

The circumvention of sales data to predict procurement requirements allows identification of franchisees involved in systematic suppression, but it has two limitations: high data requirements and greater uncertainty. The model requires large volumes of data from broader sources, including census, geographic, and macroeconomic data, with sufficient examples for each category. Indirect procurement measures allow inaccuracies to creep in at each stage, reducing overall model accuracy.

Data Requirements

Data is essential for building and using ML models, and the data required for this purpose differs from business analysis and reporting needs. ML models require numerical data; therefore, categorical information, such as product names, must be converted to numerical values. Several standard techniques are available, but consistent mapping requires standardizing names across all data, regardless of data source, ingestion time, or territory.

Item names can vary by distributor and over time, as the same product may be abbreviated differently, misspelled, or include the manufacturer name or product code. Similarly, product name variations may appear in the sales data because different individuals programmed the PoS, misspelled entries, or customer-specific modifications typed differently across locations.

The quantity for specific items also needs to be consistent across locations and time. Distributors may use different units, metric vs. imperial or volume vs. weight for the same product, total vs. unit-level breakdown.

These variations need to be cleaned before ML modeling starts, and during deployment, a system needs to be developed to catch new name variations and fix them. Cleaning can be performed through manual mapping or using LLMs to build agents that map automatically, with manual review before updating the data.

Furthermore, ML requires ample data to learn underlying relationships with sufficient accuracy and robustness. The exact data requirement depends on the relationship's complexity and the data's accuracy, and it goes beyond total volume. Data should come from many locations to prevent a few from skewing predictions, and span years to capture relationships in longer shelf-life items and seasonality.

Flag not Detection

Treat discrepancies between predicted ingredient requirements and procurement data as flags for investigation rather than conclusive findings. These could be inaccurate for many reasons, such as model uncertainty, a distributor being out of stock, inaccurate distributor reporting, or a restaurant in an area underserved by the distribution network. However, ML models can highlight specific items per location that need investigation and potential volume mismatches, helping prioritize locations and enable quick, focused investigations.

Conclusion

Detecting off-network procurement is not simply about finding unusual purchasing patterns. It requires estimating what each restaurant should reasonably procure given its reported sales, product mix, operating conditions, and procurement cycles. Heuristic rules remain useful for identifying known risk patterns, but their effectiveness is limited as franchisees alter their off-channel procurement to stay within thresholds.

The unauthorized procurement pattern varies: some locations source most items from off-network distributors, others strategically source selected items, and some do so year-round while others restrict it to deep retail discount seasons. Moreover, menu changes, expansion into new territories, and expansion of the distributor network make advanced analytical methods challenging because they require continuous updates.

Machine Learning models can estimate discrepancies between procurement needed to meet sales and sourcing from authorized distributors at the individual-item level. This limits franchisees' ability to stay under the radar by manipulating procurement of select items. Nevertheless, ML does not eliminate off-network procurement but makes it much harder to hide, requiring procurement suppression that stays consistent with sales, menu mix, and seasonality, against an updating model.