Afresh Store Ordering: Frequently asked RFP questions

1. How does Afresh generate an order recommendation?

Afresh generates an order recommendation by running several models in parallel, then resolving them through a decision engine that picks the order that reflects real business tradeoffs, rather than a single arithmetic calculation.

Afresh’s demand forecast is a full probability distribution, not a single number

Afresh’s forecasting model uses deep learning built on billions of data points, taking into account more than 40 variables per forecast, including seasonality, promotions, holidays, pricing, and cross-item relationships like cannibalization and halo effects. Rather than predicting a single expected sales number, the model predicts an entire probability distribution of possible demand outcomes for each item, store, and day, capturing how much uncertainty actually exists around that day’s likely sales. The forecast runs at the item/store/day level and can regenerate intraday for high-volume items. Because the model runs the same way across every retailer without manual configuration, no separate parameter tuning is required to launch a new store, department, or region; everything is learned directly from the data.

Afresh calculates inventory position without a manually maintained perpetual inventory

In parallel with the demand forecast, Afresh maintains a modeled estimate of each item’s inventory position by continuously reconciling physical counts, shipments, sales, known scanouts (if available), and self-tuned unrecorded shrink. Because Afresh doesn’t depend on scan-out or known-loss data, which tend to be incomplete for fresh categories, the inventory model is built to tolerate and correct for those gaps rather than assume they don’t exist.

Afresh assigns every item its own shelf-life estimate

Every item gets a shelf-life estimate, capturing how perishable that item is and how that perishability can shift over the course of a year as vendors, lead times, or product quality change. This estimate feeds directly into the inventory estimate and order recommendation, since a shorter shelf life item needs a different order quantity than one with a long shelf life, even when the demand forecast looks identical.

Afresh’s decision engine simulates thousands of scenarios to select a single order quantity

The outputs of the demand, inventory, and perishability models are combined inside a simulation that runs thousands of possible scenarios per item, weighing the cost of running out of stock against the cost of over-ordering and the resulting shrink. Case pack sizes, vendor minimums, and delivery schedules are applied as constraints, and rounding logic decides whether to round a quantity up or down based on which choice is actually less costly given current inventory and shelf life, always landing on a full case. The result is a single recommended quantity per item, per store, per day, reflecting an actual profit-maximizing tradeoff rather than a static formula.

Afresh submits high-confidence orders automatically and routes low-confidence orders to associates

Only the recommendations Afresh isn’t confident in get surfaced to a store associate; the rest are submitted automatically. Where a recommendation is surfaced, associates see the context behind it (recent sales, orders, and any active promotions) so they can decide whether to accept or override it. Afresh’s insights and reporting for corporate teams show override rates as well as how many overrides were actually beneficial to better understand store adherence and performance patterns.

2. How does Afresh calculate and manage inventory position?

Afresh calculates inventory position probabilistically rather than relying on a manually maintained perpetual inventory, which tends to drift out of accuracy quickly in fresh categories due to unrecorded shrink, mis-scans, and inconsistent counting discipline.

Afresh estimates inventory using shipment, sales, and shrink data

The estimate updates as new data comes in rather than being reconciled on a fixed schedule, and accounts for factors like late trucks, rejected shipments, and item-specific perishability that a simple point-in-time count would miss.

Afresh does not rely on scan-out or known-loss data to determine inventory position

Many ordering systems assume shrink and scan-out data is complete and accurate, then use it to calculate inventory position. In practice, this data is inherently incomplete in fresh departments (a bruised avocado or a spoiled batch of berries rarely gets formally logged), so systems built around it tend to compound small data gaps into meaningfully wrong order recommendations over time.

Afresh uses scan-out and known-loss data when accurate, but doesn’t treat it as absolute truth; it’s one signal that helps estimate inventory position. Afresh’s model is built to expect and correct for the data gaps rather than assume they don’t exist.

Afresh determines which items need a physical count based on the likelihood it would change an order recommendation

Rather than scheduling blanket physical counts on a fixed calendar or targeting items simply because they’re on reorder, Afresh targets counts based on the likelihood of a change to each item’s inventory estimate to materially change a recommendation. Nearby items on the same display are grouped together in the count list to reduce the physical effort involved.

Store associates can log transfers and adjustments, but aren’t required to for Afresh’s estimate to stay accurate

When store teams do log transfers, adjustments, or scan-outs, Afresh factors that data in to strengthen its estimate. But it isn’t required: targeted, confidence-driven count requests replace the daily maintenance burden a manually tracked perpetual inventory would otherwise require.

3. What makes Afresh’s approach to fresh categories different?

Afresh was built specifically to handle fresh, rather than adapting a system originally designed for general retail or center-store categories.

Afresh was built from the ground up for fresh, not adapted to it

Other vendors in this space took software designed for general retail or supply chain and adapted it for grocery. Afresh took the opposite path, building from the ground up around the realities of perishable, store-level fresh operations.

Afresh’s item mapping resolves disconnected grocery item codes across order, POS, and supplier systems into a single view of that item, giving it a complete movement history to use in estimating demand, inventory position, and order recommendations. Afresh’s recipe mapping models relationships where one item is prepared into another, like ingredients combining into a finished bakery item, rather than treating the finished item as disconnected from what went into it.

Afresh’s grocery-specific approach and technology show up in the metrics that matter most to a fresh-led grocer: shrink, in-stock rate, freshness, and labor effort.

Most competing systems use a point-based approach to inventory position that breaks down in fresh

Afresh’s AI/ML approach was developed by its CTO and co-founder, a PhD in AI/ML and professor whose research (published at top AI/ML conferences including ICML and NeurIPS) focuses on understanding predictive uncertainty in machine learning models. Nearly every other system on the market uses a point-based, arithmetic approach to understanding inventory position, then feeds that number into an ordering algorithm. In fresh, inventory data is inherently unreliable (a bruised avocado or spoiled berries rarely gets logged), so a system built on the assumption that inventory data is accurate ends up compounding bad data into bad orders.

Afresh applies its probabilistic approach across the full platform, not just to demand forecasting

This includes validating and scoring data quality before it’s used in any calculation, probabilistically modeling the perishability of each item, using modeled estimates of inventory position instead of maintaining a perpetual inventory, and generating order recommendations that account for all of it together. Robust demand forecasting alone isn’t sufficient to calculate accurate order quantities in fresh; it has to be paired with a modeled estimate of inventory and shelf life, and fed through a store-item specific ordering policy.

4. What level of automation does Afresh support: fully automated, or review-based?

Afresh’s automation level is intentionally configurable, since retailers differ in how much risk they’re willing to delegate to an automated system versus keep in human hands.

Afresh evaluates its own confidence in every recommendation before deciding whether to automate it

For every item, every day, Afresh’s decision engine generates an order recommendation and separately evaluates how confident it is in that recommendation. High-confidence orders are submitted without any associate involvement. Lower-confidence orders, typically involving a newer item, an unusual demand pattern, an upcoming promotion, or a flagged data quality issue, are routed to a store associate for review first.

5. How does Afresh handle exceptions and workflows?

Rather than bolting on a separate alerting system, Afresh treats exception handling as something built into every stage of the ordering process.

Afresh continuously checks incoming data for mismatches, pricing errors, and missing records

Afresh’s data-scoring layer flags ship-to-sale mismatches, pricing errors (including distinguishing a legitimate price change from a decimal or sign-flip error), and missing sales or shipment records before they can distort a recommendation.

Afresh automatically corrects common fresh data gaps

Where gaps exist, Afresh applies targeted fixes automatically: imputing likely demand during stockout periods, inferring sold-by-weight behavior from sales patterns, and using AI-based item-mapping to catch case-pack changes, code recycling, and broken item relationships in a retailer’s master data, without requiring a manual cleanup project before go-live.

Afresh surfaces only the specific items and quantities that need review, not entire orders

During the ordering process itself, only orders Afresh isn’t confident in are surfaced to an associate, and within those, the specific items or quantities that actually need review are highlighted rather than asking someone to re-check an entire order line by line.

6. How does Afresh account for seasonality, promotions, holidays, and local events?

Afresh’s forecasting model learns most demand-shaping factors directly from historical data, without requiring a retailer to manually tag or configure them.

Afresh learns seasonal and weekly demand patterns automatically from historical data

Weekly, monthly, and annual seasonal patterns are learned at the item, store, and day level, capturing short-term variation and longer seasonal cycles without manual setup.

Afresh learns promotional lift by item type, price point, and ad position

Forward-looking promotional data, including pricing, ad placement, and display type, is ingested and used to learn lift patterns, including cannibalization and halo effects where a promotion on one item measurably shifts demand for related items. A supplementary weighting mechanism is available to add extra emphasis to key promotional displays where additional lift needs to be captured.

Afresh incorporates holiday effects automatically, by category

Holiday effects don’t require a retailer to manually flag which days count as holidays.

Afresh relies on human review, not automated detection, for local one-off events

For events that aren’t captured in structured data, a nearby school event or local festival, Afresh relies on its human-in-the-loop workflow, surfacing orders for review when the model detects unusual demand it doesn’t have an explanation for, rather than guessing at unstructured local signals. Any recommendation can be edited by an associate, not just the ones Afresh proactively surfaces for review.

7. How does Afresh minimize waste while maintaining availability?

Traditional ordering systems often manage this tradeoff with static safety stock, a fixed buffer added to every order to guard against running out. This approach can lead to over-ordering and waste, since a fixed buffer doesn’t adjust based on how confident the system actually is on a given day. Afresh takes a different approach.

Static safety stock treats every item and day the same, regardless of forecast confidence

In fresh categories, where both demand and shelf life vary constantly, a static buffer tends to either waste product on predictable days or still fall short on unpredictable ones.

Afresh simulates thousands of outcomes per item to select an order quantity that balances stockout risk against shrink

Afresh’s decision engine incorporates the demand forecast, inventory estimate, and perishability model together, and selects the order quantity that balances the cost of a stockout against the cost of shrink from over-ordering. When the model is highly confident, it recommends a tighter order; when uncertainty is higher, it orders up to protect availability.

8. What departments and item types does Afresh Store Ordering support?

Afresh Store Ordering supports produce, meat and seafood, bakery, deli and prepared foods, and center store.

Afresh handles varying units of measure, bulk displays, and items sold and ordered under multiple IDs in produce

Produce ordering has to account for varying units of measure, such as bulk weight items, common scanning errors, bulk and open displays without individual UPCs, and items that are sold and ordered under multiple different IDs, which Afresh resolves through item mapping.

Afresh accounts for trim loss and complex cutting paths in meat and seafood

Afresh understands yield and cutting paths for meat and seafood, including cases where one primal cut yields multiple sellable items, to make order recommendations that account for how those items actually convert.

Afresh can optimize orders for the individual ingredients behind items produced or transformed in-store

For bakery, deli, and other prepared items, Afresh can generate order recommendations for the specific ingredients that go into a finished, in-store-produced item, rather than only ordering the finished item itself.

Afresh’s fresh-category modeling extends to non-perishable center-store items

Fresh departments already include items that behave more like center-store products, such as dressings and crackers. Afresh was originally built to handle the added complexity of fresh categories, and that same approach extends naturally to lower-complexity, non-perishable items, whether they sit within a fresh department or in center-store, which is part of why Afresh has expanded into center-store ordering as customer demand for it has grown.

9. Does Afresh support planning across the supply chain network, from store to warehouse?

Yes. Store-level demand data captured by Afresh doesn’t stay siloed at the store; it can be aggregated upstream to inform distribution center and corporate-level planning as well.

Afresh uses the same underlying forecast for store-level ordering and DC-level planning

This avoids a common failure mode in retail supply chains, where the store’s demand signal and the warehouse’s replenishment plan disagree with each other and create either bottlenecks or excess inventory.

Afresh rolls up individual store forecasts into a consolidated demand forecast for each distribution center

Order projections are aggregated across every store a distribution center supplies, giving warehouse and corporate teams an accurate, consolidated forecast of what that DC will need to fulfill, months in advance if needed.

Afresh’s fresh-category infrastructure extends to non-perishable center-store planning

Because the modeling approach was built to handle fresh categories, which are harder to forecast than shelf-stable goods, the same infrastructure extends without much additional complexity as a retailer’s use of Afresh grows.

10. What integrations does Afresh support with ERP, POS, and warehouse systems?

Afresh is designed to work alongside a retailer’s existing systems rather than replace them, which shapes how it approaches every integration.

Afresh is not a system of record

Afresh takes data in through a feed or API, in whatever format the retailer’s upstream systems already produce it in. Afresh gathers targeted input from store teams directly in its own front-end where needed, such as on-hand counts for items with lower confidence, and then outputs an order file or feed that the retailer’s own warehouse management system ingests. This means a retailer’s WMS, ERP, and other systems of record stay in place; Afresh sits alongside them rather than requiring a rip-and-replace.

Afresh has integrated with SAP and hybrid SAP/SCM environments

Afresh has worked with multiple customers running SAP or a mix of SAP and other supply chain management systems, and can scope specific nuances of a retailer’s SAP setup during technical discovery rather than requiring a standardized integration path upfront.

Afresh’s data ingestion layer accommodates a range of POS systems

Afresh’s team has experience navigating a wide variety of POS data formats and delivery mechanisms across its customer base, including specific integration experience with NCR systems. Any retailer-specific nuances in POS data are scoped and accommodated during the technical assessment phase of implementation.

Afresh generally does not integrate directly with data warehouses

Because Afresh is not a system of record, it will typically not integrate directly with a retailer’s data warehouse (for example, Google BigQuery). Instead, Afresh takes in data via feed or API and outputs its own order recommendations and reporting suite, which can be pulled into a retailer’s existing warehouse or BI tools separately. Afresh is flexible to adjust this approach to a retailer’s specific architecture and requirements where needed.

11. How does Afresh use AI?

Afresh’s core technology is built on patent-pending technology for inventory estimation, demand forecasting, and item data modeling developed specifically to handle uncertainty in fresh food, an area with far messier data and more volatile demand than most retail categories.

Afresh’s approach was developed by a team with published, peer-reviewed research on decision-making under uncertainty

This includes work published at top AI/ML conferences such as ICML and NeurIPS. That research focus, understanding and quantifying uncertainty rather than just producing a single best guess, carries through into how Afresh’s models are built today.

Afresh measures forecast quality using quantile loss instead of a single-point accuracy metric

Quantile loss is suited to evaluating an entire demand distribution rather than just a single midpoint estimate, since a midpoint number alone doesn’t capture the accuracy of decisions made under uncertainty the way Afresh’s ordering approach requires.

Afresh applies AI to inventory estimation and perishability modeling, not just demand forecasting

Afresh’s decision engine treats all of it as one connected decision under uncertainty rather than separate, disconnected steps.

AI isn’t limited to predicting demand. Afresh also uses AI it to estimate each item’s inventory position and shelf life, and its decision engine uses these three probabilistic estimates simulate scenarios, weigh tradeoffs, and make the best recommendation for the business.