Review of Afresh, Grocery Inventory and Replenishment Software Vendor

By Léon Levinas-Ménard

Last updated: November 4, 2026

Afresh provides grocery software for store replenishment, inventory counting, prepared-food production, recipe management, and distribution-center buying. Store associates use its applications to check inventory and execute recommended orders and production tasks; distribution-center buyers use it to replenish warehouses and manage purchasing exceptions. Developed initially for fresh departments, the platform expanded its advertised coverage to packaged groceries, frozen goods, general merchandise, and health and beauty in March 2026. 1 2 3 4

Overall score: 6.0/10.

Dimension	Score
Supply chain depth	6.4/10
Decision and optimization substance	6.8/10
Product and architecture integrity	5.0/10
Technical transparency	5.6/10
Vendor seriousness	6.0/10

Afresh has a substantive technical foundation for perishable inventory: a published ordering objective, probabilistic demand and inventory models, and explanations of how shelf life changes decisions. Its strongest evidence concerns fresh store operations. The broader promise of optimizing grocery economics remains less established, particularly where merchandising rules, warehouse purchasing, and store labor interact. The software can improve an order while leaving the rule that makes that order unprofitable untouched. 5 6 10

Afresh publishes enough engineering detail to assess several mechanisms, and its customer references include sustained use and operational improvements. Its public evidence is thinner on total ownership cost, security containment, and the economics of the newest distribution-center and full-store capabilities.

Afresh versus Lokad

Afresh packages grocery decisions into applications for store teams and buyers. Lokad’s approach centers on a maintained decision system that evaluates feasible actions economically under uncertainty and sends the selected actions to operational systems. Its broader design perspective is developed in Introduction to Supply Chain. The difference concerns the ownership and scope of the decision logic, rather than whether either vendor uses probabilistic forecasts.

Afresh’s published ShrinkAware policy assigns costs to spoilage, missed sales, and inventory holding. Its inventory model also treats the stock position as uncertain. These overlap with Lokad’s inventory optimization service, where Supply Chain Scientists build and maintain the customer’s numerical decision logic. 5 6

The harder distinction appears when a commercial assumption is wrong. A display minimum may require more inventory than can sell before expiry. A warehouse may improve its own clearance figures by sending aging goods to stores. The relevant decision system must expose and correct those incentives, not merely calculate orders consistent with them. Afresh’s own account of these problems provides a concrete test of how far its optimization extends. 10

For a grocer, Afresh offers considerable ready-made knowledge of counters, recipes, yields, stock counts, and intraday preparation. Lokad’s proposition places more emphasis on adapting and maintaining the economic model across interacting decisions. The practical comparison is which arrangement produces better executable decisions after counting the continuing work of associates, buyers, merchandisers, engineers, and managers.

History: technical continuity, followed by a broader AI promise

Afresh was founded in 2017. Its July 2020 funding announcement already described an AI-powered fresh-food platform and promoted sales, waste, and margin improvements. Those early claims concerned a bounded grocery application; the subsequent engineering record gives them more substance than a retrospective change of vocabulary. 12

In 2021, an engineering article explained why an accurate point forecast is insufficient for ordering perishables. It worked through a small dynamic-programming example, then showed how tracking product age makes the state space explode. The article explicitly distinguished this tutorial from the production method. Its simulated cost and waste improvements were illustrative results, not customer outcomes. 7

The sequence continued with probabilistic ingredient ordering in 2022, the ShrinkAware policy in 2023, and inventory inference and promotion forecasting in 2024. Each addressed a recognizable grocery problem and disclosed part of the mechanism. The progression supports a coherent technical roadmap: uncertainty, perishability, and imperfect observations remain central across the articles. 8 5 6 9

Customer-side reporting also supports continued deployment. Albertsons’ 2025 sustainability report states that it used Afresh in produce, meat, and seafood during 2024. That confirms operational use independently of Afresh’s sales literature, although the passage does not isolate a financial return attributable to the software. 13

The evidence changes character with the November 2025 Fresh Buying launch. Its announcement described AI agents coordinating forecasting, vendor selection, truck building, and issue resolution; Wakefern was identified as an early customer due to deploy. The March 2026 expansion then widened the proposition across the store. Grocery Outlet’s June announcement was a selection for full-store ordering. These announcements establish product direction and commercial commitments; the published engineering articles primarily describe earlier fresh-store components. 14 4 20

The historical concern is therefore the widening distance between demonstrated mechanisms and advertised coverage. Afresh has earned credit for several years of specific engineering work. Its newer agent and enterprise-wide claims need equivalent disclosure of deployed decision scope, failures, and net operating results.

Decision mechanisms: meaningful probabilities, incomplete economic boundaries

A real cost objective for perishables

ShrinkAware combines spoilage, lost-sales, and holding costs. It differentiates floor stock from backroom stock and makes the shortage penalty begin before shelves are completely empty. Its authors also explain that estimation effort should follow decision sensitivity: a small shelf-life error matters much more for some products than others. Afresh reported a further 5–10% shrink reduction against its previous policy after launch in 2023. The article does not publish the deployment sample or a reproducible evaluation. 5

This is a stronger starting point than optimizing forecast accuracy and hoping the resulting inventory is profitable. However, assigning a cost to a low display does not establish the right value for that cost. The penalty can encode an economically justified loss of demand, or it can encode a merchandising preference that systematically overfills shelves. The quality of the decision depends on how that distinction is measured and maintained.

Inventory uncertainty and the value of counting

Afresh’s 2024 InvHMM article models the true inventory as hidden and physical counts as noisy observations. Its particle filter carries alternative inventory trajectories through sales, shipments, shrink, and expiration, then updates their relative plausibility when counts arrive. Afresh reports 10–40% lower mean squared error than perpetual inventory on a subset of reliable measurements. 6

The useful contribution is the connection to action. If a count could change the next order materially, it has value; if all plausible stock positions imply no order, recounting has little immediate benefit. Sampling alone is a basic numerical technique. Choosing which uncertainty is worth resolving can remove actual store work. The public account supplies a mechanism for that choice, while leaving the joint calibration of multiple items and supply events largely unexplored. 6

Forecast validation with an operational horizon

Afresh’s promotion-forecasting article describes an encoder-decoder network using sales, prices, store and item identifiers, display sizes, and promotion types. It evaluates three-day demand, linking that horizon to the ordering interval and delivery delay. An ablation removes promotion inputs; the resulting deterioration is compared through quantile loss and calibration. The reported 77.5% coverage for an 80th-percentile forecast at one point near promotion start is a calibration result, not a claim of 77.5% forecast accuracy. 9

This is considerably more informative than a near-perfect accuracy headline. The chosen aggregation serves a stated replenishment horizon, and uncertainty remains visible. The next economic question is whether the better forecast changes the orders that can still be executed in time, and how much margin those changed orders preserve. A forecast improvement after the first promotional day can be valuable for later replenishment while arriving too late for the opening allocation.

Recipes connect items; dependence remains consequential

The 2022 recipe tutorial translates demand for prepared products into demand for shared ingredients. It samples recipe-item distributions, applies conversion factors, and derives an ingredient-order quantile. The example explicitly assumes independent demands and immediate receipt and preparation before opening. It leaves substitution, varying yields, and ingredient shortages as extensions. 8

A common promotion can raise demand for several recipes simultaneously; a competing dish can divert it. Independent draws will misrepresent the tail of the combined ingredient requirement when those relationships matter. Adding samples cannot repair a missing dependence. The tutorial makes this limitation explicit.

Afresh’s October 2026 recipe documentation now describes cross-department ingredient attribution and store-specific ingredient identities and yields. It also recognizes that overrides can reveal missing context in the model. This is a practical advance in maintaining the representation of the business. The documentation does not show how the current production system models joint demand, yield, and supply uncertainty. 11

Distribution-center buying broadens the decision perimeter

The October 2026 DC documentation specifies probabilistic demand, vendor reliability in inventory buffers, freight-versus-spoilage tradeoffs, capital limits for forward buys, and demand transferred to substitutes. Routine purchase orders can be submitted automatically; forward buys require buyer approval. 3

The unresolved question is how these considerations combine. Reliability statistics and configurable lead times can improve a buffer without representing the distribution of future delivery dates and quantities. A locally profitable forward buy can consume capacity or cash better used elsewhere. The public description does not expose the joint stochastic model or the method for selecting coupled actions. Such a model must also remain correct as supplier practices, commercial terms, and store behavior change. That maintenance burden is central to holimization, not an implementation detail that disappears once a solver returns an answer.

Configuration maintenance, productivity, and accountability

Who owns the assumptions that destroy the return?

An October 2026 article by Afresh’s VP of Account Management describes disappointing shrink despite apparently good store-ordering metrics. One example traces the loss to a distribution center overbuying and sending short-dated products downstream. Another gives a display requiring two cases of blood oranges where weekly sales are one tenth of a case. It argues that an ordering system filling the display is doing what it was asked to do. 10

That last argument exposes a weak economic boundary. The display represents roughly twenty weeks of demand before considering freshness. A system sold to reduce waste should make that contradiction actionable and assign responsibility for changing the display. Successful compliance with a loss-making instruction is not successful optimization.

The author says that, in his consulting experience, reported system failures almost always traced to misaligned people and processes. The display example reveals the limit of that distinction: the software accepts a merchandising input that predictably generates waste. Economic responsibility needs to cover detecting and correcting such inputs, including who must act when the application cannot change them. 10

Afresh’s recipe-attribution article supplies a better stance: repeated overrides may indicate missing model context. Its current forward-deployed engineering vacancy also assigns engineers responsibility for discovering the customer’s actual problem, building the solution, and measuring outcomes such as hours saved and waste reduced. These are constructive commitments to maintaining the model. The vacancy describes an operating approach, rather than evidence of contractual risk sharing or audited project returns. 11 19

The work removed, and the work retained

Afresh’s store-ordering page advertises 40–50% faster ordering and describes guided counts and order reviews. It also offers recommendations through a backend API for an existing retailer interface. The relevant productivity benefit is fewer calculations and better-targeted observations; the core store workflow still includes associate judgment and physical verification. 1

Production Planning links ingredients and finished goods and schedules preparation within the day. Associates review inventory estimates before production. Configurable minimums remain binding, and the product records execution, overrides, and adherence. 2 The distinction between necessary physical work and avoidable planning work is essential: cutting fruit requires labor, whereas repeatedly repairing quantities that a model could calculate is an ongoing software cost.

Fresh Thyme’s case study describes replacing paper-based production planning across a 70-store business. It reports adherence rising from roughly 40% to 90%, 70% year-on-year growth in rotisserie-chicken sales, and 8.8% growth in deli sales. It also describes corporate visibility used to coach stores. These figures show an operational change; they do not give the additional labor, waste, margin, or implementation cost needed to calculate its net return. 15

Heinen’s provides a different and narrower example: replacing price-based period-ending counts with item-level scanning. Its CFO says the count now takes four hours, but the published account supplies no numerical duration for the old process. The improvement in item identification is tangible; a percentage labor saving cannot be calculated from that account. 16

These gains should be judged against the full recurring workload: maintaining recipes and yields, reviewing merchandising minimums, correcting feeds, investigating exceptions, counting, and supervising adherence. A faster store interface is useful. A durable productivity advantage requires the total amount of decision work per store or purchasing desk to fall without equivalent work appearing in a central support team.

Bill of software materials and security

Current recruitment and partnership disclosures identify a reasonably coherent platform direction. The Ordering backend vacancy explicitly lists Python, JavaScript, TypeScript, React Native, Databricks, and GraphQL as Afresh’s stack; Azure appears in its experience requirements. The forward-deployed engineering vacancy states that its team runs Databricks and Azure with PySpark and dbt. These are current, undated vacancies retrieved for this review. 18 19

Component or boundary	Evidence	Consequence
Ordering applications and services	Named languages, React Native, and GraphQL in the Ordering vacancy	Mobile and service code must evolve together as workflows change.
Data engineering	Databricks, Azure, PySpark, and dbt in forward-deployed engineering	Processing, orchestration, data transformations, and their operating costs remain part of delivery.
Customer data exchange	Databricks Open Sharing, formerly Delta Sharing	Can reduce duplicate integration work for retailers already using the platform.
Operational integration	Recommendation API and purchasing-system connections	Execution still depends on maintained mappings and retailer systems.

The October 2026 Databricks announcement describes access to customer data in an existing governed environment and sharing standardized data back. That is a useful integration arrangement. The partnership name contributes no independent evidence that the ordering model is better, cheaper to operate, or easier to maintain. Those results require measurements of duplicated storage, processing, integration effort, and ongoing engineering time. 17

Afresh’s DC documentation identifies SSO, SCIM, access controls, encryption, and end-of-agreement data destruction. Its public description stops short of explaining tenant isolation, the reach of compromised service credentials, restrictions on support access, or recovery after destructive changes. 3 These are the mechanisms needed to assess security by design. Compliance status adds no answer to those engineering questions.

The engineering roles also leave the permissions of production code and privileged operations unspecified. Their access to Afresh’s hosted and customer-connected environments needs an enforceable boundary, including limits on credentials, dependencies, and network access.

Supply chain score

The overall score is the equally weighted mean of the following 25 sub-scores. These assess the public evidence available on November 4, 2026.

Supply chain depth

Criterion	Score
Economic framing	7/10
Decision end-state	8/10
Conceptual sharpness on supply chain	8/10
Freedom from obsolete doctrinal centerpieces	5/10
Robustness against KPI theater	4/10

Average: 6.4/10.

Economic framing — 7/10. The published perishable-ordering objective connects quantities to costs. The weakness lies in commercially imposed settings that can overwhelm those economics without being corrected by the decision system. 5 10

Decision end-state — 8/10. Store orders, preparation quantities, and purchase orders are concrete outputs. Integration with purchasing systems gives the DC product an execution path beyond advice and reports. 1 2 3

Conceptual sharpness on supply chain — 8/10. Afresh explains perishability, imperfect inventory, ingredient transformations, and the gap between a good forecast and a good order. These mechanisms engage with the physical business at a useful level of detail. 6 7 11

Freedom from obsolete doctrinal centerpieces — 5/10. Probability distributions improve on static planning buffers, but service preferences and merchandising minimums still shape the result. The economic validity of those settings cannot be delegated indefinitely to policy maintenance. 2 5

Robustness against KPI theater — 4/10. The technical articles distinguish calibration, error, and decision cost. Commercial case studies emphasize sales and adherence without publishing a complete return calculation, allowing successful intermediate metrics to substitute for profitability. 9 15

Decision and optimization substance

Criterion	Score
Probabilistic modeling depth	7/10
Distinctive optimization or ML substance	7/10
Real-world constraint handling	8/10
Decision automation and labor productivity	6/10
Resilience under real operational complexity	6/10

Average: 6.8/10.

Probabilistic modeling depth — 7/10. Demand distributions and uncertain stock estimation have disclosed mechanisms and validation. The evidence is less complete for jointly uncertain deliveries, yields, substitution, and shared resource use. 6 8 9

Distinctive optimization or ML substance — 7/10. ShrinkAware and the inventory model address specific shortcomings of generic replenishment. Their contribution comes from the formulation and operational use, rather than the novelty of sampling or neural networks. 5 6

Real-world constraint handling — 8/10. Recipe attribution, freshness, preparation steps, and purchasing constraints receive explicit treatment. The remaining concern is the interaction of those constraints across the network, not an absence of domain detail. 2 3 11

Decision automation and labor productivity — 6/10. Recommended quantities and automated routine purchasing remove work. Guided store reviews, exception handling, and adherence supervision remain substantial, and customer cases do not measure the complete continuing workload. 1 3 15

Resilience under real operational complexity — 6/10. Afresh connects recurring overrides to missing recipe context and describes correcting ingredient mappings and yields. Its forward-deployed engineers turn operational problems into pilots and reusable platform changes. These provide concrete routes to revising the decision model. Historical replay, economic acceptance tests, and controlled release of those revisions remain insufficiently documented. (11, 19)

Product and architecture integrity

Criterion	Score
Architectural coherence	6/10
System-boundary clarity	6/10
Security seriousness	3/10
Software parsimony and ownership cost	4/10
Compatibility with programmatic and agent-assisted operations	6/10

Average: 5.0/10.

Architectural coherence — 6/10. Ordering and production share ingredient and forecast information, and the disclosed engineering stack has a recognizable platform direction. Public material gives limited visibility into how the newest buying and agent capabilities share that architecture. 2 18 19

System-boundary clarity — 6/10. The operational interfaces and associate workflows are understandable. Economic responsibility is less clear when an order follows a configured rule that creates predictable waste. 1 3 10

Security seriousness — 3/10. Authentication and access-control capabilities are listed. Containment, privileged access, destructive-failure recovery, and the consequences of a compromised integration are not explained at a comparable level. 3

Software parsimony and ownership cost — 4/10. The data platform, application services, and customer-specific engineering create recurring operating costs. Neither hosting consumption nor continuing configuration effort is quantified. 17 18 19

Compatibility with programmatic and agent-assisted operations — 6/10. APIs and automatic order submission enable machine execution. The agent proposition is credible only to the extent that it eliminates decision work while preserving control of economic assumptions; its announced breadth exceeds the published implementation detail. 1 3 14

Technical transparency

Criterion	Score
Public technical documentation	7/10
Inspectability without vendor mediation	6/10
Portability and lock-in visibility	4/10
Implementation-method transparency	6/10
Evidence density behind technical claims	5/10

Average: 5.6/10.

Public technical documentation — 7/10. The engineering articles describe objectives, assumptions, models, and evaluation measures in unusual detail for enterprise planning software. The corresponding disclosure for the newer product perimeter is more functional than mathematical. 5 6 9 3

Inspectability without vendor mediation — 6/10. Readers can examine meaningful mechanisms and tutorial code without a demonstration. They cannot reproduce the reported production improvements or inspect the complete deployed decision logic. 7 5

Portability and lock-in visibility — 4/10. Data exchange and a recommendation API reduce interface dependence. They do not explain how a retailer would export the learned behavior, reconstruct policy choices, or replace the model without rebuilding its decision process. 1 17

Implementation-method transparency — 6/10. Product documentation and engineering recruitment reveal data work, rollout, and continuing customer engagement. This helps identify the work involved, but does not establish its predictable cost or who absorbs it when assumptions change. 2 19

Evidence density behind technical claims — 5/10. Component-level evaluations and named deployments support core fresh-store claims. Wider economic results and the newest platform extensions rely more heavily on vendor summaries and launch announcements. 6 9 13 14

Vendor seriousness

Criterion	Score
Technical seriousness of public communication	6/10
Technology roadmap integrity	7/10
Conceptual sharpness	7/10
Economic accountability	4/10
Defensibility in an agentic-software world	6/10

Average: 6.0/10.

Technical seriousness of public communication — 6/10. Engineers distinguish illustrative models from production and explain limitations that matter. Broad AI and margin claims receive less exact treatment in commercial material. 7 8 12 14

Technology roadmap integrity — 7/10. The dated sequence shows sustained work on the same fresh-food problems, followed by related product expansion. The recent agent branding increases the burden to show equivalent continuity in the new capabilities. 5 6 7 14

Conceptual sharpness — 7/10. Afresh understands that inventory observations, shelf life, and transformations are part of the decision problem. Its treatment of display-driven losses is less convincing because it accepts an uneconomic instruction as the boundary of software responsibility. 6 10 11

Economic accountability — 4/10. Customer examples make the benefits plausible, and forward-deployed engineering emphasizes measured outcomes. The public record still omits full project economics and leaves room to attribute unfavorable results to surrounding operations. 10 15 19

Defensibility in an agentic-software world — 6/10. Perishable-inventory knowledge and decision-relevant data modeling can remain valuable as interfaces become easier to build. The durable advantage must be reliable, maintained economic decisions, rather than an assistant presenting another list of tasks for a buyer. 6 11 14