Inventory approval services can now record decision rules

From tools that count inventory to tools that support decisions

A new kind of inventory tool can help people make purchasing and allocation decisions, rather than simply count stock. Atomic, a Boston-based company in the United States, raised $12.5 million, approximately KRW 16.9 billion at a reference exchange rate, on September 29, 2026. Its work is not limited to inventory quantities. It covers deciding what to buy and produce, when to do so, and which warehouse should receive it.

Atomic compares options using demand, inventory, supplier lead times, and conditions at individual warehouses. It handles repeat work through predefined rules, while sending unusual situations to a person for review, revision, and approval.

The important change is not that AI replaces people. It is that criteria previously applied only in a worker’s head—such as “order when stock falls to this level” or “send scarce goods to this customer first”—can be recorded and used again.

Smaller manufacturers in Korea need an even narrower starting point. In a 2024 survey by the Ministry of SMEs and Startups (중소벤처기업부), the government ministry responsible for small and medium-sized enterprises, 18.6% of SMEs with factories had adopted smart-factory systems. Among adopters, 75.5% remained at the basic stage.

An analysis of the same survey by the Korea SMEs and Startups Institute (중소벤처기업연구원) found that 75.7% of manufacturers collecting production data entered it manually, while 19.7% collected it in real time. Automating decisions on top of late or incorrect inputs only repeats bad orders faster. The condition for success is therefore current data and an approval process, not elaborate forecasting.

How a 12-person food manufacturer’s day could change

Consider one user. Kim is responsible for both raw-material purchasing and inventory at a frozen side-dish manufacturer with 12 employees.

Each morning, Kim transfers warehouse receiving messages and production-team usage into an inventory file. Kim then pastes in the sales team’s order sheet and looks up supplier lead times and minimum order quantities in old files or from memory.

When spinach runs short, Kim decides which product to make less of. When packaging containers are delayed, Kim decides which orders to ship first. Decisions happen over phone calls and messaging apps, so the next person can see the outcome but not easily understand the reason. When Kim takes a day off, the owner reviews the same materials again.

A small decision service does not immediately replace the inventory file. It takes the file Kim already uses and first flags only conditions such as “stock is below expected usage,” “stock will run out before delivery,” or “meeting the minimum order quantity will leave too much inventory.”

It then asks Kim for a short reason for changes. Reasons such as “event orders this week,” “supplier holiday,” or “switching suppliers because of a quality issue” can accumulate. When the same reasons recur, the service can apply those rules to the next order proposal and show only items that differ from normal.

The screen compares the option to keep the original order, reduce the order quantity, or split the order between two suppliers. It shows when stock could run out under each option and how much extra stock would remain. But Kim or the owner must approve the order before it is sent.

Some work still belongs to people. The person responsible must judge circumstances that are difficult to reduce to numbers, such as a quality incident, a sudden large order, or a supplier relationship. This service is likely to fail if it tries to automate such exceptions, or if it presents recommendations as correct answers while stock movement records are delayed.

Operations that already separate rules from exceptions

Fresh-food ordering at US grocery stores

US company Afresh forecasts demand by store and item, recommends fresh-food order quantities, and lets store employees accept or change the recommendation. In a three-month trial at CUB, which operates about 100 stores in the US Midwest, Afresh reported a 92% recommendation acceptance rate and an 8.8% improvement in store operating efficiency.

Afresh sells through enterprise software contracts with grocery chains, and its pricing is not public. The key point is not fully automated ordering for perishable products. It is that store employees retain the authority to revise recommendations.

Exception management at a Swiss pharmaceutical distributor

Swiss pharmaceutical distributor Galexis adopted RELEX demand forecasting and replenishment services. Rather than having people review every item, it receives alerts for items with sudden demand changes or possible stockouts. The system proposes urgent deliveries and optimizes safety stock based on exceptions defined by users.

According to results published by RELEX four months after implementation, the automated processing rate rose from 44% to 88%, while the rate of immediately supplying needed products remained above 98%. Contract pricing is not public. The structure automates repeat items while leaving risky exceptions to people.

A beer company that started in Canada and expanded across countries

Global beer company AB InBev began using o9’s planning service in Canada in 2020, then expanded it to Brazil, China, the United States, Europe, Africa, and Asia. It compared afterward whether manual changes made by staff improved forecasts. Helpful changes were retained as rules, while interventions that did not help were reduced.

o9 reported an approximately 10% improvement in forecast accuracy and an approximately 25% inventory reduction. In some markets, 70% to 90% of recurring planning was handled without human intervention. AB InBev evaluated planners’ revisions afterward to measure the effect of manual intervention.

Four things you could build now

1. A raw-material order approval inbox

This service takes an inventory file and purchase requests, then separates items to order today, items to delay, and items that need confirmation. It is for purchasing staff at small food manufacturers that record stock movements manually and buy raw materials regularly.

A full factory system may be burdensome, but adding an approval step on top of an existing file can start small. The first screen should show current stock by item, expected depletion date, recommended order quantity, the reason for the recommendation, and an approval button.

2. A shortage allocation board

When incoming supply is lower than orders, this service creates delivery options based on due dates, contract priority, and whether substitutes are available, then asks a responsible person to approve them. It is for wholesalers that supply imported parts to multiple manufacturers and divide scarce goods by phone whenever stock runs out.

As Atomic and the international examples show, a practical starting point for supply-chain automation is closer to handling shortage exceptions than automating every decision. The first screen should place unfulfilled orders, promised due dates, recommended allocation quantities, the reason for prioritization, and changed orders together.

3. A perishable-product ordering assistant

This service uses sales and remaining stock to propose the next day’s order quantity, then reflects the store manager’s reason for changes in future recommendations. It is for multiple neighborhood stores that order short-shelf-life products such as prepared side dishes or sandwiches themselves.

As the Afresh example shows, operations where frontline employees review recommendations already exist, so the scope can be reduced for small stores. The first screen should show tomorrow’s recommended quantity, the quantity likely to remain unsold, stockout risk, and buttons to “order as recommended” or “revise.”

4. A returns-processing rule inbox

This service takes return reasons and product condition, proposes the next action—resell, repair, return to supplier, or discard—and leaves an approval record. It is for logistics staff at household-goods wholesalers that decide how to handle returns through photos and messaging conversations.

Returns involve recurring decisions about whether an item should go back into inventory, but the criteria and responsible person are often unclear. That makes returns a good function to separate out and start with. The first screen should include return photos, purchase date, condition-check items, recommended handling method, expected inventory change, and the approver.

What to check today

Ask one person responsible for inventory or purchasing to open their 10 most recent urgent orders and explain why each order was made for 30 minutes. If the same two or three conditions explain six or more of the 10 orders, and the approver for the remaining exceptions is clear, that decision is worth turning into a rule and approval screen first.

Why this matters where you are

You can check whether urgent inventory decisions in your own operation are explained by a small set of repeat conditions, and whether exceptions have a clear approver. Data collection methods, supplier relationships, and existing systems may differ by market. But an existing spreadsheet can be a starting point for testing an approval workflow before replacing the whole inventory system.